Volltext PDF
Volltext PDF
TECHNISCHE BERICHTE
TECHNICAL REPORTS
ISSN 1430-211X
TUD-Fl15-05-Oktober 2015
Cognitive Computing
Collected Papers
c
Fotolia.com (Used with permission)
“Cognitive technologies are
products of the field of artificial
intelligence.They are able to
perform tasks that only humans
used to be able to. [1]”
Table of Contents
Introduction 5
Introduction
Context
“Cognitive Computing” has initiated a new era in computer science. Cognitive computers are
not rigidly programmed computers anymore, but they learn from their interactions with humans,
from the environment and from information. They are thus able to perform amazing tasks on
their own, such as driving a car in dense traffic, piloting an aircraft in difficult conditions, tak-
ing complex financial investment decisions, analysing medical-imaging data, and assist medical
doctors in diagnosis and therapy [2][3][4]. Cognitive computing is based on artificial intelligence,
image processing, pattern recognition, robotics, adaptive software, networks and other modern
computer science areas, but also includes sensors and actuators to interact with the physical
world [1].
Cognitive computers – also called “intelligent machines” – are emulating the human cognitive,
mental and intellectual capabilities. They aim to do for human mental power (the ability to
use our brain in understanding and influencing our physical and information environment) what
the steam engine and combustion motor did for muscle power. We can expect a massive impact
5
of cognitive computing on life and work. Many modern complex infrastructures, such as the
electricity distribution grid, railway networks, the road traffic structure, information analysis
(big data), the health care system, and many more will rely on intelligent decisions taken by
cognitive computers.
A drawback of cognitive computers will be a shift in employment opportunities [5]: A raising
number of tasks will be taken over by intelligent machines, thus erasing entire job categories
(such as cashiers, mail clerks, call and customer assistance centres, taxi and bus drivers, pilots,
grid operators, air traffic controllers, . . . ). A possibly dangerous risk of cognitive computing
is the threat by “super intelligent machines” to mankind [6]. As soon as they are sufficiently
intelligent, deeply networked and have access to the physical world they may endanger many
areas of human supremacy, even possibly eliminate humans. Cognitive computing technology is
based on new software architectures – the “cognitive computing architectures” [7][8]. Cognitive
architectures enable the development of systems that exhibit intelligent behaviour.
Seminar Work
This seminar worked on answers to the central question:
Which are the situation, the challenges, and the impact of cognitive
computing in the year 2025?
• Individual, guided research in the selected area and authoring of a scientific paper. Feed-
back from peer reviewers,
• A first seminar day: The participants presented their results and received feedback from
the audience,
• Improvement of the paper and the presentation, based on the peer review feedback,
• A second seminar day: The participants presented their improved results and received
feedback from the audience,
The participants learned: (a) to do focused research in a specific area (“Cognitive Comput-
ing”), (b) to author a scientific paper, (c) to improve their LaTex expertise, (d) to experience the
peer-review process, both in giving and receiving peer review advice and (e) to hold convincing
presentations, and (f) to benefit from a considerable broading of their perspective in the field of
technology, software, applications, and impact.
6
As a final outcome of the seminar, a proceedings volume including all the papers produced
by the participants was assembled (= this report) and is made available in electronic form to
anybody interested. Seminar language was English. Three seminar days were held and 3 ECTS
credits were awarded for the successful participation. Audience was limited to 8 participants.
My sincere thanks go to Georg Püschel for his highly valuable assistance during the seminar
and for assembling this Technical Report. Thanks also to the 7 participants, whose active
collaboration and innovative work is greatly appreciated.
References
[1] David Schatsky, Craig Muraskin, and Ragu Gurumurthy. Demystifying artificial intelligence
– what business leaders need to know about cognitive technologies. Downloadable from:
http://dupress.com/articles/what-is-cognitive-technology/ [last accessed: 31.12.2014], 2014.
[2] John E. Kelly III and Steve Hamm. Smart Machines: IBM’s Watson and the Era of Cognitive
Computing. Columbia University Press, N.Y., USA, 2013.
[3] Erik Brynjolfsson and Andrew McAfee. The Second Machine Age – Work, Progress, and
Prosperity in a Time of Brilliant Technologies. WW Norton & Company Inc., N.Y., USA,
2014.
[4] Eric W. Brown. Cognitive computing ushers in new era of it. Downloadable
from: http://www.forbes.com/sites/ibm/2014/02/03/cognitive-computing-ushers-in-new-
era-of-it/ [last accessed: 8.9.2014], 2014.
[5] Harry Rudin. Will the it revolution cost our children their jobs? Download-
able from: http://ercim-news.ercim.eu/en99/challenges-for-icst/will-the-it-revolution-cost-
our-children-their-jobs [last accessed: 7.10.2014], 2014.
[6] James Barrat. Our Final Invention – Artificial Intelligence and the End of the Human Area.
Thomas Dunne Books (St. Martin’s Press), New York, N.Y., USA, 2013.
[8] John E. Laird. The SOAR (State, Operator And Result) Cognitive Architecture. The MIT
Press, Cambridge MA, USA, 2012.
7
Participants
Each participant had to choose one focus and elaborate on one specific theme related to the
focus question. The following choices were made by the participants:
Name Q1 Q2 Q3
Mai, Felix X
Schröppel, Philipp X
Nett, Tobias X
Butler, Tyler X
Blau, Ben X
Richter, Robert X
Hauptseminar Papers
The following papers were authored, peer-reviewed and presented during the Hauptseminar. All
papers are reproduced in full on the following pages.
8
Applying the Subsumption Architecture to the
Genesis Story Understanding System
A Notion and Nexus of Cognition Hypotheses
Felix Mai
TU Dresden
[email protected]
1 Introduction
propose evaluate
CP/NS HC CS/AI
adapted by implemented by
Fig. 1. The feedback loop between cognitive psychology (CP) and neuroscience (NS)
on the left side and computer science (CS), i. e. AI, on the right side with the cognition
hypothesis (HC ) as mediating interface in the middle.
Elaboration
Graph
Elaboration
Graph
Fig. 4. The nexus of the story understanding workflow and the subsumption architec-
ture
4 Conclusion
References
1. H. Awad. Culturally Based Story Understanding. Master thesis, Massachusetts
Institute of Technology, June 2013.
2. R. A. Brooks. A robust layered control system for a mobile robot. Mar. 1986.
3. R. A. Brooks. Intelligence without representation. Artificial Intelligence, 47(1–
3):139–159, 1991.
4. J. D. Friedenberg and G. Silverman. Cognitive Science: An Introduction to the Study
of Mind. Sage Publications, Inc., 2006.
5. D. Kolak, W. Hirstein, P. Mandik, and J. Waskan. Cognitive Science: An Introduc-
tion to Mind and Brain. Routledge, 2006.
6. P. Langley. The cognitive systems paradigm. Advances in Cognitive Systems, 1:3–13,
July 2012.
7. A. Newell and H. A. Simon. Computer science as empirical inquiry: Symbols and
search. Commun. ACM, 19(3):113–126, Mar. 1976.
8. P. H. Winston. The strong story hypothesis and the directed perception hypothesis.
Advances in Cognitive Systems: Papers from the 2011 AAAI Fall Symposium (FS-
11-01), 2011.
9. W. V. H. Yarlott. Old Man Coyote Stories. Master thesis, Massachusetts Institute
of Technology, June 2014.
Bene ts and Drawbacks of Hardware
Architectures Developed Speci cally for
Cognitive Computing
Philipp Schr•oppel
[email protected]
TU Dresden
Abstract. The Von Neumann computer architecture has been the dom-
inant architecture for all kinds of computer systems over the last 70 years.
The upcoming idea of cognitive computing { basically meaning that com-
puters act with their environment in new, more natural ways { raises the
question if new hardware architectures, inspired by the human brain,
could be more e cient for its implementation.
By describing and comparing two existing cognitive systems, one based
on the Von Neumann architecture and the other one on a brain-inspired
architecture, I want to give a prospect on which approach will be more
promising for the future.
This paper will provide a summary of the bene ts and drawbacks of both
architectures and analyze how they are suited for dealing with the key
challenges for computer systems in the future.
1 Introduction
Most of the currently used computers are based on the Von Neumann architec-
ture: There is a memory where data is stored and a central processing unit which
can perform calculations based on the data. The list of instructions that tell the
computer how to process the data has to be provided by the programmer.
The wish for making computers act more natural, simplifying interaction
between humans and computers, is not new, but apparently technology is still
struggling with achieving this goal.
One promising approach is called cognitive computing, which diverges from
the idea of computers that strictly follow predefined instructions given by a
programmer. Basically cognitive computing means that computers interact with
the environment in new, more natural ways and draw their own conclusions.
They learn by themselves and develop new strategies to deal with problems.
They can work on data, even if the provided information is not consistent or
even contradictory [1].
Today there are several challenges where cognitive computers might be help-
ful. One example is handling big amounts of data: As more and more data is
available today, it is di cult for the programmer to write programs which solve
problems while considering all the given information. With cognitive computing
2
the programmer doesn't have to specify how to process every single piece of
information, but the computer itself can decide what is important and find a
solution.
Today there are cognitive computing systems for Von Neumann machines, but
other hardware architectures with completely new software ecosystems have been
developed for cognitive computing as well.
Some of the more successful and in uential cognitive architectures for Von
Neumann machines are SOAR, ACT-R, CLARION and EPIC [2], which have
been started to be developed decades ago. Although newer systems, like Google
DeepMind, are recently producing impressive results, I want to give an intro-
duction to SOAR as an example for the classical approaches.
On the other hand, there are not too many systems with hardware ar-
chitectures developed specifically for cognitive computing. One such system
is TrueNorth, a brain-inspired architecture developed by IBM. I will give an
overview to TrueNorth, because its development has already reached a state
where results can be seen and first applications are implemented.
Up to now, it is not clear whether new hardware architectures will become
established, or if the Von Neumann will continue to be the leading architecture.
Therefore I want to present my opinion on which architecture will be better
suited for cognitive computing in the future.
Simply put, every behavior shown by SOAR is the result of trying to get from a
current state closer to a defined goal state, following the principle of rationality:
\if an agent has knowledge that an operator application will lead to one of its
goals then the agent will select that operator" [4].
A state describes a situation the system could be in and consists of all features
and values that are important to model the situation.
As the principle of rationality suggests, SOAR will – like a human being
mostly does – perform actions in order to achieve a specific desirable situation,
3
here described by, one or more, goal states. Doing so, the system moves from
state to state, trying to get closer to the goal.
Internally, operators are used for these movements between states. An oper-
ator can be applied when predefined conditions – expressed by specific features
and values – are fulfilled and will cause a state change either by just internally
changing some features and values or by making the system perform an action
in the real world, which in turn leads to a change of perception that eventually
changes features and values as well.
Goal
state
...
operator
Initial
state ...
Goal
state
While state of the art computer chips have the problem of consuming more
and more energy and therefore producing too much heat, the brain works with
approximately 20 Watt [5]. One single damaged transistor in a chip can mean
the end of the whole system, but the human brain, providing a way higher fault
tolerance, can still work after being damaged. On top of that, the brain is able
to organize itself and find new ways to solve problems instead of just executing
predefined program code like computers [6].
A promising attempt to build a brain-inspired computer is a system called
TrueNorth which is developed by IBM in the SyNAPSE program and includes
a hardware chip and a software ecosystem.
TrueNorth consists of 4096 newly developed cores, which are called neurosy-
naptic cores by IBM. Each core simulates 256 neurons and 256 synapses per
neuron and handles processing power, memory and communication by itself, so
there is no need for memory shared by all chips, like in Von Neumann architec-
tures [7]. Communication between the cores happens through spike-events.
While designing TrueNorth, the scientists at IBM have focused on low power,
rather than compact size. Because the system uses no clock, but performs com-
putation based on events, no energy is wasted while cores are idle [7].
The following chapter will explain the basic hard- and software architecture
of this system.
While computer chips have specialized units for different tasks, the \computa-
tion" in the brain is executed by billions of more or less equal neurons, which
are strongly connected, forming so-called neuronal networks, and communicate
through electrical and chemical signals [8].
Neurons are attached to axons for sending and dendrites for receiving sig-
nals [9]. In order to transmit information, the neuron sends electrical signals
over the axon. The axon is connected to synapses where the incoming electrical
signal results in the release of neurotransmitters. Those neurotransmitters are
chemicals that in turn cause electrical activity in the dendrites of other cells,
making those cells receive the sent information.
Given this brief introduction to the main components of the human brain, we can
start regarding how they are implemented in TrueNorth's hardware architecture.
This will be done with a bottom-up approach: First we have a look at the
implementation of a single neuron, then, going one step further, how the single
neurons are linked together to form big networks.
The Neuron Model. Each of TrueNorth's 4096 cores simulates 256 neurons.
This simulation has to be based on a neuron model: A model which mathemat-
ically describes behavior of a neuron in a more or less accurate way.
5
Over the years lots of neuron models have been developed. For the imple-
mentation in a computer system it is crucial to find a model which is detailed
enough for su ciently imitating the behavior of a neuron, while being not that
detailed that it needs too much computational power [10].
One such model is the leaky integrate-and-fire model, which was developed
in 1965 and is used in the TrueNorth system today.
As the name indicates, this model consists of three parts: Integration, fire and
leak. The membrane potential of a neuron increases over time by summing up
the synaptic input (synaptic integration) until it reaches a threshold. When this
happens, the neuron fires, which means that it sends out an action potential and
resets the membrane potential. For taking leak currents into account, every time
step the membrane potential diminishes by a fixed value (leak integration) [10].
35
30
25
Membrane potential (mV)
20
15
10
0
0 0.01 0.02 0.03 0.04 0.05
Time (seconds)
Starting out from this model, the scientists at IBM have added several ex-
tensions to make their simulated neurons more versatile [10].
The first extension was adding stochasticity. It is possible to switch neurons
between a deterministic and a stochastic mode, which affects synaptic and leak
integration: While in deterministic mode, synaptic and leak integration hap-
pens normally every time step, in stochastic mode it is randomly determined if
synaptic and leak integration will happen at all, or if they will be skipped for the
current time step. Furthermore stochasticity is also introduced for the threshold,
making the threshold value vary by a random number.
Secondly, four different leak modes were supported. The leak can be positive
or negative, thus increasing, or decreasing the membrane potential, or it can
be divergent or convergent, meaning that it causes the membrane potential to
converge towards 0 V or diverge away from 0 V.
The last extension affects the resetting of the membrane potential by pro-
viding three different reset modes. The first mode, the normal one, resets the
6
membrane potential to a fixed value after the threshold is crossed. The second
mode resets the membrane potential to the value that the potential was exceed-
ing the threshold. The third mode does not reset the potential at all, relying on
synaptic integration and leak integration to diminish the membrane potential.
K x N Synapses
Synapse
K Axons
N Neurons
only external inputs and outputs [11]. On the one hand, the specification of the
network includes parameters for the neuron model, as described in 4.2, and on
the other hand it provides the anatomy of the network, meaning the connec-
tions between neurons within one core and the connections between different
neurosynaptic cores.
Therefore scientists at IBM have developed a programming paradigm for
TrueNorth which simplifies writing programs, that can be e ciently executed
by the hardware [11]. This programming paradigm introduces Corelets, which
are abstractions of a TrueNorth program, and hence encapsulate the specifi-
cation of a network of neurosynaptic cores, exposing only external inputs and
outputs. Furthermore the paradigm consists of the Corelet Language for creat-
ing Corelets, the Corelet Library, which is a repository of reusable Corelets, and
the Corelet Laboratory, a programming environment for the whole programming
cycle.
Obviously, as soon as the neuronal networks used for a program get bigger,
it is di cult for the programmer to specify the whole network. For this reason,
it is possible to compose simple Corelets in order to build more complex ones.
Axons on a Corelet that receive their input from outside of the Corelet are
called input connectors and neurons that send their output to destinations out-
side of their Corelet are called output connectors.
For composing Corelets, the programmer can connect output connectors from
one Corelet to input connectors from another one.
Of course the Von Neumann architecture has not been out there for such a long
time without a reason. When it comes to performing calculations and applying
logic, neither humans nor cognitive systems can keep up with Von Neumann
machines today. One also should not forget all the concepts, systems and tech-
nologies developed for this architecture over the last decades, which can not be
just ported to other systems.
Nevertheless the Von Neumann architecture has some drawbacks, which are
long known, but get more and more relevant now and in the near future.
For taking big amounts of data into account for computations, enormous
computing power is needed. Until today the computing power has steadily in-
creased, but this process is slowing down. Companies had to stop putting up
the clock rates of processors already years ago and for the first time since the
seventies, they are struggling to keep up with Moore's Law. The reasons for both
of this basically is, that processors consume more and more power and there-
fore produce more heat. And this heat is starting to get a problem for making
computers even faster [12].
In contrast to this, brain-inspired systems have a remarkably small power con-
W
sumption: While CPUs have a power density of around 50 100 cm 2 , TrueNorth
mW
only needs 20 cm2 . Even more: TrueNorth can run a typical recurrent network
consuming only 70 mW. This is about four orders of magnitude lower than a
simulation on a Von Neumann computer would need [7].
Probably Von Neumann machines will always be better than neuronal sys-
tems at plain calculations, but for learning new behavior and interacting with
their environment they might be left behind pretty soon. Already now, applica-
tions like speaker recognition or digit recognition have been developed and work
well for the still young TrueNorth system [13]. This gives confidence that brain-
inspired architectures can and will deliver fundamentally new ways of interaction
between humans and computers.
A major problem for neuronal systems is the development of software. As
described in 4.3, the programmer has to specify the anatomy of the neuronal
network, meaning he has to define all the connections between the neurons.
Although the behavior of single neurons is quite well understood today, it is
still not clear how neuronal networks are formed and how the neurons work
together [14]. Naturally, this makes it di cult for software developers to set up
networks that solve the problem he is dealing with.
6 Conclusion
As Von Neumann computers are used in all areas of technology all over the
world, they will definitely be present for the next decades and probably won't
disappear at all. Completely discarding the Von Neumann architecture would
mean leaving years and years of research behind and it is likely that it will take
very long to develop equivalent technologies for new hardware architectures.
9
But one should also be aware that Von Neumann computers have several
advantages, as described in the previous chapter, that justify to continue using
them, not just for the reasons of compatibility with old technologies.
However, regarding cognitive computing, apparently brain-inspired systems
are better suited. IBM describes them as \slow, sensory, pattern recognizing
machines", which makes them good at working in real time with all kinds of
– probably noisy – data from sensors, and therefore well-suited for interacting
with their environment and humans.
In the future, there won't be just one dominant system, but it will rather be
important for researchers and engineers to know benefits of both systems and
decide depending on the problem which system is more e cient.
One can also imagine combining both architectures, to get the best out of
both of them to build powerful computers that can easily learn new behavior,
adapt to their environment and interact naturally with humans.
Brain inspired computers are still a very new technology, but one can be sure
that after 70 years of having the von Neumann architecture as the dominant
architecture for all kinds of computer systems, brain-inspired architectures will
provide a promising alternative in the future, at least for specific areas like
cognitive computing.
References
[1] IBM Research. Cognitive computing. [Online] Available from: http://www.
research.ibm.com/cognitive-computing/#fbid=hD9eIGrGTwH. [Accessed 20-May-
2015].
[2] Wikipedia. ACT-R. [Online] Available from: https://en.wikipedia.org/wiki/ACT-
R#What ACT-R looks like. [Accessed 01-August-2015].
[3] Jill Fain Lehman, John Laird, and Paul Rosenbloom. Gentle introduction to
SOAR, an architecture for human cognition: 2006 update. [Online] Available
from: http://ai.eecs.umich.edu/soar/sitemaker/docs/misc/GentleIntroduction-2006.
pdf. [Accessed 01-August-2015].
[4] Allen Newell. The knowledge level. Technical report, Carnegie Mellon University,
1982. Available from: http://repository.cmu.edu/cgi/viewcontent.cgi?article=2616&
context=compsci.
[5] Paul Merolla et al. A digital neurosynaptic core using embedded crossbar mem-
ory with 45pj per spike in 45nm. IEEE Custom Integrated Circuits Conference
(CICC), Sept. 2011. Available from: http://www.modha.org/papers/012.CICC1.
pdf.
[6] Karlheinz Meier. Computer nach dem vorbild des gehirns? Ruperto Carola, 1,
2007. Available from: http://www.uni-heidelberg.de/presse/ruca/ruca07-1/vorbild.
html, [Accessed 01-August-2015].
[7] Dharmendra S. Modha. Introducing a brain-inspired computer. [Online] Available
from: http://www.research.ibm.com/articles/brain-chip.shtml, 2015. [Accessed 07-
August-2015].
[8] Eric H. Chudler. Millions and billions of cells: Types of neurons. [Online] Avail-
able from: https://faculty.washington.edu/chudler/cells.html. [Accessed 07-August-
2015].
10
[9] Eric H. Chudler. Making connections: The synapse. [Online] Available from:
https://faculty.washington.edu/chudler/synapse.html. [Accessed 08-August-2015].
[10] A. S. Cassidy, P. Merolla, J. V. Arthur, S. Esser, B. Jackson, R. Alvarez-
Icaza, P. Datta, J. Sawada, T. M. Wong, V. Feldman, A. Amir, D. Rubin,
F. Akopyan, E. McQuinn, W. Risk, and D. S. Modha. Cognitive computing
building block: A versatile and e cient digital neuron model for neurosynap-
tic cores. International Joint Conference on Neural Networks (IJCNN), 2013.
Available from: http://www.research.ibm.com/software/IBMResearch/multimedia/
IJCNN2013.neuron-model.pdf.
[11] A. Amir, P. Datta, A. S. Cassidy, J. A. Kusnitz, S. K. Esser, A. Andreopoulos,
T. M. Wong, W. Risk, M. Flicknerand R. Alvarez-Icaza, E. McQuinn, B. Shaw,
N. Pass, and D. S. Modha. A corelet language for composing networks of neurosy-
naptic cores. International Joint Conference on Neural Networks (IJCNN), 2013.
Available from: http://www.research.ibm.com/software/IBMResearch/multimedia/
IJCNN2013.corelet-language.pdf.
[12] John E. Kelly and Steve Hamm. Smart Machines: IBM’s Watson and the Era of
Cognitive Computing. Columbia University Press, New York, NY, USA, 2013.
[13] S. K. Esser, A. Andreopoulos, R. Appuswamy, P. Datta, D. Barch, A. Amir,
J. Arthur, A. S. Cassidy, P. Merolla, S. Chandra, N. Basilico, S. Carpin,
T. Zimmerman, F. Zee, M. Flickner, R. Alvarez-Icaza, J. A. Kusnitz, T. M.
Wong, W. P. Risk, E. McQuinn, and D. S. Modha. Cognitive com-
puting systems: Algorithms and applications for networks of neurosynaptic
cores. International Joint Conference on Neural Networks (IJCNN), 2013.
Available from: http://www.research.ibm.com/software/IBMResearch/multimedia/
IJCNN2013.algorithms-applications.pdf.
[14] Wikipedia. Brain. [Online] Available from: https://en.wikipedia.org/wiki/Brain.
[Accessed 07-August-2015].
Language Workbench Technology For Cognitive
Systems
Tobias Nett
1 Introduction
Within the last years, computers became capable of doing things only humans
used to be able to do. This advent of cognitive computing heavily influences
our lives, visible in self-driving cars or programs competing in quiz shows like
Jeopardy!, and invisible in many business applications [16]. Consequently, de-
signing and implementing cognitive systems becomes an increasing issue, and the
demand for good tools and environments for creating cognitive software raises.
A good starting point for research on the challenges and opportunities of
cognitive computing is the online article Cognitive Computing Ushers In New
Era Of IT by Eric W. Brown [3]. Schatsky et al. provide further insights in
the chances for business in their article Cognitive technologies: The real op-
portunities for business [16]. To readers interested in the technical background
of cognitive architectures the comparing review Cognitive architectures: Re-
search issues and challenges by Pat Langley, John E. Laird, and Seth Rogers is
recommended [13].
The remainder of this paper is structured as follows. Section 2 presents the
state-of-the-art in cognitive computing and gives a short overview of existing
cognitive architectures and frameworks. Section 3 evaluates domain-specific lan-
guages in AI and cognitive technologies. The focus lies on the structure of these
DSLs and how they are incorporated or adapted in current cognitive frameworks.
In section 4 graphical notations used in other domain-specific software are ana-
lyzed by reference to two concrete tools. Section 5 aims to combine the presented
aspects and create an enabling environment for graphical notations in cognitive
computing applications. Finally, section 6 concludes with a summary on chal-
lenges and opportunities when combining the fields of cognitive computixng and
domain software engineering.
2 Existing Work
This work is based on two important aspects of today's computer science re-
search and software engineering technologies: the advent of cognitive computing
in the last years [4, 17], fostered by business needs and growing computational
power on this field, and the urge to increase the software development process
by individual, domain-specific tooling and graphical assistance [9].
Cognitive computing might have become noted to a broader audience through
projects like IBM's Watson [11], but, as Langley et al. shows, the number of es-
tablished cognitive architectures and frameworks is way higher than the amount
of publicly known representatives [13]. Thus, we are interested to see how ex-
isting architectures face the challenges of implementing complex interactions
of cognitive technologies, e. g., computer vision, machine learning, robotics, or
project planning and scheduling.
Although cognitive systems often aim to solve problems of a specific domain
(e. g., Icarus, a cognitive architecture for physical agents [12]) we can observe
a gap between cognitive computing and domain software engineering for other
disciplines. To cite an example, Pérez Andrés et al. present an approach based
on metamodeling and view-based programming which enables languages with
graphical and textual views [15]. On the other side, domain-specific systems for
cognitive computing mostly use textual representations, such as those presented
in the next section.
Listing 1.1. Excerpt of the definition of a two player instance of Tic Tac Toe in GDL.
1 ( role xplayer )
2 ( role oplayer )
3
4 ( i n i t ( c e l l 1 1 blank ))
5 ...
6 ( i n i t ( c e l l 3 3 blank ))
7 ( i n i t ( control xplayer ))
8
9 (<= ( l e g a l ? p l a y e r ( mark ?m ? n ) )
10 ( t r u e ( c e l l ?m ? n b l a n k ) )
11 ( true ( control ? player )))
To cope the limitations of basic GDL not being able to describe games with
chance or incomplete information, the language extension GDL-II (Game De-
1
http://ii.tudelft.nl/trac/goal
scription Language for Incomplete Information) has been developed by Thielscher
[18]. Over the years, many adaptations of the language to specific environments
and use cases have appeared. Therefore, GDL is established as base description
language for general gameplaying tasks.
In common with other DSLs for AI, GDL is purely textual. The language is
the formal frame to communicate game rules (such as allowed moves, shape of
board, or winning conditions) to game players using a conceptualization of games
in terms of entities, actions, propositions, and players. The language specification
does not have any game-specific constants or keywords, but consists only of
general functions.
Listing 1.2. An example for a problem instance associated with the gripper domain.
1 ( d e f i n e ( problem g r i p p e r )
2 ( : domain g r i p p e r )
3 ( : o b j e c t s rooma roomb b a l l 1 left )
4 (: init ( room rooma )
5 ( room roomb )
6 ( ball ball1 )
7 ( gripper left )
8 ( at−r o b b y rooma )
9 ( free left )
10 ( a t b a l l 1 rooma ) )
11 ( : g o a l ( a t b a l l 1 roomb ) ) )
In its core, PDDL allows for domain descriptions, i. e., what are the elements
common to all problems of the domain, and problem descriptions, i. e., deter-
mining the specific planning-problem. As stated above, extension of the core lan-
guage is necessary for concrete planning problems. Therefore, several extensions
and derivations for various field which differ in representation and functionality
exists. Although these derivations and extensions are all based on core PDDL,
they have to be seen as independent languages for planning problems.
The pool of languages derived from PDDL indicates that a general mecha-
nism for compatible language extensions is necessary to keep the language family
clear and consistent. Optimally, domain experts are able to choose a representa-
tion of a planning problem in their desired notation, without breaking compati-
bility with other language adoptions facing a different concern.
Domain-specific languages such as GDL or PDDL are not closely integrated with
most cognitive architectures. However, there exist translators for some language-
architecture pairs. For instance, the Soar architecture offers translators for both,
GDL and PDDL. A similar translation from PDDL to the ACT-R framework
has been provided by Amant et al. [1].
The major drawback of chaining different language tools, translators, and
compilers is that the user has to learn and master the correct collaboration of
these tools. Figure 1 illustrates the resulting toolchain of different bridging and
glueing techniques. Although problem specifications often are similar, the user
has to implement architecture specific solutions which are not compatible to
other systems. Moreover, the interleaving of tools in complex toolchains is not
easily comprehensible for new developers or domain experts without program-
ming experience, and strong dependencies on specific tools restrict the reusability
of general solutions.
PDDL A1
...
GDL A2
From the point of view of software engineering, the employed model transfor-
mations and code generation to glue generalized languages of AI to specific cog-
nitive platforms is not transparent. The presence of various architecture-specific
languages emphasizes the need for a modular and extensible base language for
cognitive computing. A prime example is PDDL and all its problem-specific
extensions which are mostly incompatible to each other.
4 Language Workbench Technology for Domain-specific
Tooling
We have seen in the previous section the predominance of textual problem and
domain specifications in cognitive technologies. Listings 1.1 and 1.2 conveyed a
good insight into typical text-based representations of environment and problem
descriptions. Domain-specific language and tools in other disciplines can be taken
as a reference for advanced concepts in domain software engineering. Thus, we
look at two distinct approaches which are both based on Language Workbench
Technology aiming to simplify software development processes. Section 4.1 in-
troduces MetaR, a data analysis toolbox for biologists and bioinformaticians.
Section 4.2 presents mbeddr, an extensible C-based language used in embedded
systems engineering.
Language Workbenches are tools which enable an efficient development and usage
of (programming) languages. They simplify the generation of productive DSLs
and provide means to create and manage sets of related languages. The leading
language workbenches (such as MPS2 , Spoofax3 , or Rascal4 ) offer many generic
facilities, i. e., usage analysis, language modularization and extension, or model
visualization.
Figure 2 demonstrates the close relationship between language workbenches
and the implementation target. Language transformations and semantic adjust-
ments are solely performed within the language workbench environment. In the
end, only one concise transformation step is required to map the domain-specific
solution developed in the language workbench to the implementation target.
Extensions
2
https://www.jetbrains.com/mps/
3
http://strategoxt.org/Spoofax
4
http://www.rascal-mpl.org/
The two DSL projects presented in the following sections are both based on
language workbench technology, in particular on JetBrains’ Meta- Programming
System (MPS). The core of this language workbench instance is projectional edit-
ing (or view- based programming). Instead of using the traditional approach of
compiler techniques, i. e., parsing, transformation, and code generation, MPS
maintains programs directly as abstract syntax trees with reference overlay
graphs. This program representation allows to easily specify diverse notations
for specific model elements. To cite an example, nested conditional statements in
standard programming language can be presented in different views, e. g., using
conditional statements itself, a control flow graph, or a decision table.
Figure 3. MetaR allows to integrate graphical representations (e. g., plots) and user
interface elements (e. g., buttons) into analysis programs.[6]
MetaR demonstrates how textual problem descriptions can be mixed with
other notations, e. g., tables or plots as can be seen in fig. 3. Taking the biomarker
development as an example, it embeds the standard notation used in biology
within the source code. Thereby, it integrates well with the original domain lan-
guage and domain experts can easily express the desired functionality. Common
solution patterns are abstracted to intuitive user interface elements which hide
the implementation complexity.
User Extensions
Figure 5. The proposed language ecosystem for cognitive computing based on lan-
guage workbench technology. The language families are decoupled from the backend
architectures and implementation languages.
6 Conclusion
To sum it up, domain software engineering is an important part of cognitive
computing. The amount of architectures and domain-specific language variations
in AI and cognitive technologies shows the need for comprehensible and usable
notations of programs. However, we have seen that the cognitive computing
domain is cluttered with various approaches tailored for specific tasks, but with
a poor compatibility between language or architecture variants.
We proposed projectional language-metaprogramming for cognitive comput-
ing based on language workbench technology to face the challenge of creating a
better domain-specific ecosystem for cognitive computing. The core idea was to
introduce an extensible base language for cognitive technologies which can be
adapted to specific domains as needed. Language workbenches allow for mod-
ular language specifications and extensions and can target multiple backend
platforms.
The major open challenges are to find common base concepts which can be
used as base language concepts. The established DSLs such as GDL or PDDL
are promising candidates. Several architectures already provide translators which
can be integrated in the code generation phase of the language workbench. Ad-
ditionally, existing language extensions and derivations have to be modeled as
components of the language workbench. Matching the semantics between dif-
ferent language layers is important for the correctness of the program transfor-
mation and implementation target generation. Finally, the graphical notations
used in cognitive science and systems engineering need to be ported to the new
ecosystem.
References
[1] Robert St. Amant, Sean P. McBride, and Frank E. Ritter. An AI planning
perspective on abstraction in ACT-R modeling: Toward an HLBR language
manifesto. In: ACT-R Workshop proceedings (2006).
[2] Victoria M. Benson and Fabien Campagne. Language workbench user
interfaces for data analysis. In: PeerJ 3.e800 (2015).
[3] Eric W Brown. Cognitive Computing Ushers In New Era Of IT. 2014.
url: http://www.forbes.com/sites/ibm/2014/02/03/cognitive-
computing-ushers-in-new-era-of-it/.
[4] Erik Brynjolfsson and Andrew McAfee. The second machine age: Work,
progress, and prosperity in a time of brilliant technologies. W.W. Norton
& Company, 2014.
[5] mbeddr Core Development Team. mbeddr website. May 17, 2015. url:
http://mbeddr.com/.
[6] Weill Medical College of Cornell University. MetaR website. May 17, 2015.
url: http://campagnelab.org/software/metaR/.
[7] Mehdi Dastani. 3APL Platform - User Guide. Tech. rep. November. 2006.
[8] Mehdi Dastani et al. A Programming Language for Cognitive Agents;
Goal Directed 3APL. In: Programming Multi-Agent . . . (2004), pp. 111
130.
[9] W.B. Frakes and Kyo Kang. Software reuse research: status and future.
In: IEEE Transactions on Software Engineering 31.7 (2005), pp. 529 536.
[10] Malik Ghallab et al. PDDL - The Planning Domain De nition Language.
Tech. rep. Yale Center for Computational Vision and Control, 1998.
[11] John E Kelly III and Steve Hamm. Smart Machines: IBM's Watson and
the Era of Cognitive Computing. Columbia University Press, 2013.
[12] Pat Langley and Dongkyu Choi. A unified cognitive architecture for phys-
ical agents. In: Proceedings Of The National Conference On Arti cial In-
telligence 21.2 (2006), p. 1469.
[13] Pat Langley, John E. Laird, and Seth Rogers. Cognitive architectures: Re-
search issues and challenges. In: Cognitive Systems Research 10.2 (2009),
pp. 141 160.
[14] Nathaniel Love et al. General game playing: Game description language
speci cation. Tech. rep. Stanford: Stanford University, 2008.
[15] Francisco Pérez Andrés, Juan De Lara, and Esther Guerra. Domain spe-
cific languages with graphical and textual views. In: Lecture Notes in
Computer Science 5088 LNCS (2008), pp. 82 97.
[16] David Schatsky, Craig Muraskin, and Ragu Gurumurthy. Cognitive tech-
nologies: The real opportunities for business. In: Deloitte Review 16 (2015).
[17] David Schatsky, Craig Muraskin, and Ragu Gurumurthy. Demystifying
Arti cial Intelligence. Tech. rep. Deloittie University, 2014.
[18] Michael Thielscher. A general game description language for incomplete
information games. In: Proceedings of the Twenty-Fourth AAAI Confer-
ence on Arti cial Intelligence (AAAI-10) (2010), pp. 994 999.
[19] Markus Voelter and Sascha Lisson. Supporting Diverse Notations in MPS
Projectional Editor. In: GEMOC 2014 (2014), p. 7.
[20] Markus Voelter et al. mbeddr: an Extensible C-based Programming Lan-
guage and IDE for Embedded Systems. In: Proceedings of the 3rd Annual
conference on Systems, Programming, and Applications: Software for Hu-
manity (2012), pp. 121 140.
Networked Brain‐based Architectures for more Efficient
Learning
Tyler Butler
Technische Universität Dresden
Institut für Software‐ und Multimediatechnik
Fakultät Informatik
[email protected]
Abstract. In the future, computers will need to make autonomous decisions
based on data much in the way that human beings do. For example, one day
an autonomous robot may be presented with the task of navigating in an unfa‐
miliar environment and will need to be able to recognize obstacles and to
adapt to changes in real time. This is something humans and other mammals
are able to do very quickly and without much thought, but current‐day com‐
puters have great difficulty doing as they must be programmed to account for
each scenario they encounter and cannot make real‐time adaptations. This pa‐
per suggests taking a brain‐based approach to cognitive software develop‐
ment, using the emergent paradigm of cognitive science and taking advantage
of its environment‐based adaptions to create cognition much in the same way
it evolved in the brains of mammals. One of the shortcomings of using the
emergent paradigm is that agents need significant exposure to their environ‐
ments before they can properly adapt and achieve cognitive function. I will
present a potential solution to this problem by applying methods from Ma‐
chine to Machine communication and cloud computing to create a network of
agents running emergent‐based software which are connected to a database
where they can share experiences with other agents of the same type. This
would increase the speed at which agents are able to adapt to their environ‐
ments and give them a much higher quality of decision making than if they had
only their own experiences available for reference.
Keywords: emergent, Machine to Machine, cloud, brain‐based
1 Introduction
Many software architectures used today and in the past rely on the
programmer to account for all uncertain variables that the software may
encounter while executing in its environment. This may come of little
consequence when a software engineer is trying to create a web application
for his company to store data in a database, where all of the variables can be
somewhat easily accounted for and tested. Even in modern industrial
robotics, this approach of algorithmic programming is usually sufficient when
faced with the task of assembling a car step‐by‐step, seen in many factories
such as those of BMW in Leipzig and Munich. However, when we step out of
these convenient artificial environments where most variables can be
accounted for by if‐else statements, and into the real world where even the
best programmer cannot account for every circumstance that may occur, the
old method of programming breaks down and becomes insufficient for even
the most basic of tasks such as navigating without running into obstacles.
A potential solution to these problems is to abandon the typical
algorithmic approach to writing software, and instead write software that
mimics the biological processes of the mammalian brain. There are currently
several projects making significant advances in the field of brain‐based
software architectures. One is the MoNETA program at Boston University's
Neuromorphics Lab which is looking at the brains of animals at a high level of
abstraction and trying to create a software architecture that can replicate the
learning processes of an animal when given behavioral tests. Another project
at the forefront of artificial intelligence research is Google Deepmind's Deep
Q Network (DQN) project, which uses deep artificial neural networks and
advanced algorithms to learn through reinforcement how to do a variety of
general tasks, much in the way that an animal brain would do. These projects
will be discussed in further detail in the Existing Work section.
Fig. 1. An overview of my paper is outlined in the following figure:
My mission for this paper will be to show how robots in the future that
use software architectures based on the brain and reinforcement learning to
achieve cognition can be much more efficient and effective when networked
together. I will begin with a detailed state‐of‐the‐art to show the current
state of software architectures that are based of the brain and learn through
reinforcement. Then I will introduce the Emergent Paradigm, which is a
formal way of describing how a system can achieve true cognition through
reinforcement learning and exposure to its environment. I will introduce a
major problem of systems based on the Emergent paradigm, namely that
systems learn slowly and are prone to initial failure. I will conclude by
examining a potential solution to these problems, which is to add a
networked component to the typical reinforcement‐learning paradigm which
will allow artificial systems such as robots to share experiences among one
another and learn more quickly and effectively. This would be of great
importance, because it would help robots make much more effective
decisions autonomously and learn from their mistakes, allowing them to
perform tasks in the physical world that they were previously unable to do.
2 Existing Work
Neuromorphic Engineering, the emergent interdisciplinary subject
taking inspiration from fields such as computer science and biology to create
artificial neural systems such as vision systems and autonomous robots
whose physical and design principles are based on those of biological
nervous systems [2], has seen progress recently due to investment from
DARPA in the form of the SyNAPSE (Systems of Neuromorphic Adaptive
Plastic Scalable Electronics) with contributing major players including IBM,
HRL Laboratories, and several large universities. However, most of the work
being done is in relation to creating new hardware that more closely relates
to the biological brain rather than bio‐based software architectures [3].
A project funded by the SyNAPSE which does focus more on cognitive
software architectures is MoNETA at Boston University, in which the primary
goal is to “create an autonomous agent capable of object recognition and
localization, navigation, and planning in virtual and real environments.” [5]
MoNETA uses a cognitive software architecture based on a mammalian brain
with reinforcement learning, whereby actions leading to the best outcomes
are repeated when presented with a similar scenario. Using this style of
learning, MoNETA when placed in a virtual environment was able to emulate
the behavior of real rats when performing the Morris Water Navigation Task,
a common procedure used to measure spatial learning and memory in
behavioral neuroscience [6].
Another organization on the forefront of research into brain‐based
artificial intelligence is Google Deepmind, which recently published a paper
in Nature revealing its work on a deep neural network (called a Deep‐Q
network, or DQN) that uses reinforcement learning to play Atari video games
at a superhuman level [12]. What's special about this new development is
that it only takes in the raw pixels of the Atari game's screen as input (as a
human would), and is able to play a variety of games involving a large
diversity of tasks using a single brain‐based algorithm. In other words, the
Deep‐Q network is able to learn to do a variety of different tasks and make
continuous adaptations without being specifically programmed for each
separate task. The figure below shows the DQN compared against the best
linear‐learning algorithms when faced with the task of playing Atari games.
Fig. 2.
This method of general purpose learning far surpasses the
capabilities of automated machines of the past, which were mostly unable to
perform tasks they were not specifically programmed to do, and represents a
great leap forward in machine learning and general artificial intelligence.
There are several other ongoing projects which involve brain‐based
computer systems including IBM's TrueNorth program [4] and Darwin, a
cognitive software architecture presented in Artificial Cognitive Systems: A
Primer by David Vernon, which uses similar learning techniques based on
interactions with the environment and reinforcement to achieve cognition.
[7]
The Emergent Paradigm of Cognitive Science is one of the two paradigms of
cognitive science outlined in the book Artificial Cognitive Systems: A Primer
by David Vernon. As described in the text, “The ultimate goal of an emergent
cognitive system is to maintain its own autonomy, and cognition is the
process by which it accomplishes this. It does so through a process of
continual self‐organization whereby the agent interacts with the world
around it but only in such a way as not to threaten its autonomy. In fact, the
goal of cognition is to make sure that the agent's autonomy is not
compromised, but is continually enhanced to make its interactions
increasingly more robust.” [7] Emergent systems are dependent on
embodiment in its environment which allow them through experience to
learn actions and environments that promote or harm the system's
autonomy.
Fig. 3.
This is opposed to the cognitivist paradigm of cognitive science, which
states that mind and body are independent and cognition is achieved
through a symbol‐based recognition and learning system. Both MoNETA and
the Deep‐Q neural network described in the existing work section use the
emergent paradigm to achieve cognition. For example, in the Deep‐Q net‐
work the input pixel image of the videogame screen is its connection to the
environment, and maximizing the score of the game is seen as desirable and
actions leading to it reinforced. On the contrary, losing the game is seen as
undermining the systems autonomy and is learned to be avoided. Over time,
as shown by the DQN's ability to make superhuman predictions while playing
video games, emergent systems will achieve true cognition by being able to
anticipate events in the future and prepare for those events. This is very sim‐
ilar to the way biological brains evolved to achieve cognition.
3.1 Creating a Scaleable Emergent‐base Cognitive Architecture
Most robots today are placed in controlled environments such as
factories in which they only perform a simple repetitive task such as placing a
door on a car that is already in an exact, fixed position. If we were to take the
robots of today out of these controlled environments and place them into
the real‐world where not all variables can be accounted for, the current
software architectures that robots use would be insufficient to carry out
useful automated tasks. In the future, we could possibly look to brain‐based
software architectures that use the emergent paradigm to help robots make
well informed, quick decisions while in unfamiliar environments. While being
able to replicate the cognitive power of the human brain is most likely still a
long way off, being able to create a software architecture that mimics the
cognitive functions of a lower mammal's brain could be possible by the year
2025. In Artificial Cognitive Systems, Vernon says in regards to brain‐based
cognitive systems that “Since human intelligence evolved from the
capabilities of earlier primates, ideally a cognitive model of human
intelligence should be reducible to a model of animal intelligence. This is bio‐
evolutionary realism. Sometimes, this is taken the other way around by
focusing on simpler models of cognition as exhibited by other species – birds
and rats, for example – and then attempting to scale them up to a human‐
level cognition.” Therefore, if we are able to create a cognitive software
architecture which uses the same systems for learning and decision making
as an animal‐brain, we can reap the immediate benefits of this system in
fields such as robotics in the near‐future and let this system learn and evolve
into the long‐term future when we may be able to see truly human levels of
cognition.
4 Using the Emergent Paradigm and networked reinforcement‐
learning as a future cognitive software architecture
As seen in the figure below, the DQN took around 200 iterations of playing a
game before it maximized its potential score.
Fig. 4.
While this number of learning‐trials is sufficient for research
purposes, expensive robots may not have the luxury of multiple failures
before they properly adapt to their environment. The cognitive software
architectures used by robots and other artificial agents in 2025 could be
similar to that of MoNETA's and the DQN's in that they learn through
reinforcement and interactions from the environment, but instead of each
robot starting from scratch, a cloud database of memory and experiences
could be assembled and every robot of a similar type connected to this
database. Machine to Machine (M2M) is a broad term used to refer to
technologies that allow communication between devices of the same type. In
the past decade, M2M along with the Internet of Things has generated
unprecedented amounts of data, with new smaller electronics containing
sensors such as those on smartphones, refrigerators, and thermostats now
being connected to the internet.[14]
This abundance of data has a significant positive impact on technical
innovation. As described in Erik Brynjolfsson's and Andrew McAfee's The
Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant
Technologies ‐ “Digitation increases understanding. It does this by making
huge amounts of data readily accessible, and data is the lifeblood of science.”
If a software architecture could be developed which combines the powerful
brain‐based reinforcement algorithms used by MoNETA and the DQN with
Machine to Machine communication and the vast amounts of useful data
that comes from it, it could minimize the time it takes for artificial agents to
learn about their environments, and provide a vast array of experiences for
the agents to use to make decisions which a single agent would not have
been capable of producing alone. For example, if a cleaning robot with this
architecture in the future were to sense and encounter a specific piece of
metal in the environment which, when sucked up by the robot's vacuum
caused a part of the robot to break, it would recognize this encounter as
having a negative effect on its autonomy. It would then upload the situation
and encounter into the cloud used by other cleaning robots so that when
another cleaning robot encounters this specific type of metal, it know to
avoid it.
This would create a very effective cognitive architecture for robots of
the near future, in which new situations are consistently being added to help
robots navigate an uncertain world. This new networked architecture would
be in most cases superior to requiring all agents to be left to adapt to their
environment on their own, and would far surpass the current model of
attempting to account for every environmental variable before releasing the
agent into the world.
5 Conclusion
6 References
1. [Online]. Available: http://nl.bu.edu/research/projects/moneta/moneta‐v2‐0/new‐navi‐
gation‐and‐decision‐making‐systems/
2. [Online]. Available: http://en.wikipedia.org/wiki/Neuromorphic_engineering#cite_note‐4
3. [Online]. Available:
http://www.darpa.mil/Our_Work/DSO/Programs/Systems_of_Neuromorphic_Adaptive_
Plastic_Scalable_Electronics_%28SYNAPSE%29.aspx
4. A. Rutkin (2013, August 8). MIT Technology Review [Online]. Available: http://www.tech‐
nologyreview.com/news/517876/ibm‐scientists‐show‐blueprints‐for‐brainlike‐compu‐
ting/
5. A. Gorchetchnikov, “MoNETA: massive parallel application of biological models navi‐
gating through virtual Morris water maze and beyond”, US National Library of Medicine
National Institutes of Health, BMC Neurosci. v.12, July 2011.
6. R. Hooge (2001, August). Brain Research Reviews [Online]. Available: http://www.sci‐
encedirect.com/science/article/pii/S0165017301000674
7. D. Vernon , “Paradigms of Cognitive Science,” in Artificial Cognitive Systems: A Primer,
Cambridge, MA: MIT Press: 2014, pp. 32‐84.
8. [Online]. Available: http://www.cs.jhu.edu/~hager/Public/teaching/cs461/ObjectRecogni‐
tion.pdf
9. http://www.scientificamerican.com/article/visionary‐research/
10. J.L Krichmar and G. M. Edelman. “Brain‐based devices for the study of nervous systems
and the development of intelligent machines”, Artificial Life, 11:63‐77, 2005.
11. V. Mnih et al., “Human‐level control through deep reinforcement learning”, Nature v.
518, February 2015
12. [Online]. Available: https://en.wikipedia.org/wiki/Machine_to_machine
13. [Online]. Available: https://www.bosch‐si.com/media/bosch_software_innovations/doc‐
uments/publication/english_1/2012/2012‐07‐bigdata_industrialit_byimcramer_pub‐
lished_on_bosch‐sicom.pdf
14. E. Brynjolfsson, E. McAfee, “The Digitization of Just About Everything” in The Second Ma‐
chine Age: Progress, and Prosperity in a Time of Brilliant Technologies. New York, NY:
W.W Norton & Company: 2014, pp. 67
Developing Better Pharmaceuticals
Using the Virtual Physiological Human
Benjamin Blau
Abstract. We are at a point in history where computers are enabling what was
once considered science-fiction to become reality. The cross between comput-
ers and the medical industry is allowing for better treatment than ever before.
However, even with today’s technologies, there is still room for improvement.
Many drugs and treatment options have the potential to leave unwanted neg-
ative side effects. Additionally, the current drug development process is be-
coming more expensive and yielding fewer results. Current technologies are
laying out the future of drug development. Breakthroughs in simulation of the
body, when coupled with computer simulations of chemical reactions, will lead
to incredible discoveries in clinical medicine. Cognitive computing allows us to
further expand this technology to function on its own. This paper discusses the
future of drug research and development with the assistance of cognitive tech-
nologies.
1 Introduction
Current drug development processes are not only expensive and tedious but
they often leave us with less than ideal results. There are several ways in
which drugs come to the market. Many drugs are tested on animals but
these often don’t make it to a human testing stage [1]. The American Food
and Drug Administration requires all drugs to be tested on animals before
they can move on to clinical trials on humans. Following this, there are many
details that must be worked out before the trials can begin. Even after a drug
has been approved for human testing, there are many variables that can
leave uncertainties in a drugs capabilities. This is because of how different
every individual is. Not only does everyone differ in age, weight, etc., every
individual also has a different medical history. All of these factors make it dif-
ficult to be positive about potential negative side-effects that some drugs
may have. In addition, clinical trials can take years and do not always result
in a drugs approval.
My vision for this paper is to research new methods for the creation and de-
velopment of clinical medication and how these methods can be trans-
formed through cognitive computing. Using a cognitive computing format
similar to that of Watson Health, computerized simulations of chemical reac-
tions, when partnered with projects such as the Virtual Physiological Human
(VPH), can help us to create safer and more patient-specific drugs at a
quicker and cheaper rate than we currently are. Following the section on ex-
isting work in this field, the third section gives a more in depth explanation of
the Virtual Physiological Human, which is the backbone of this research; the
fourth section introduces recent breakthroughs in chemical reaction simula-
tion and the fifth section further discusses the potential of both of these top-
ics in regards to the future of clinical medicine.
2 Existing Work
There are two primary pieces of work that are looked at in this paper: The
Virtual Physiological Human and computerized simulations of chemical reac-
tions. The latter has been in development and use for several years now.
However, recent advances in chemistry, which were awarded the Nobel Prize
in Chemistry in 2013, has extended the use of this technology to be more
cost-effective and more capable than previous. The former is the combined
effort of computer scientists, biologists, engineers, clinicians and many other
professions. The VPH is currently a large focus of the European Comission’s
7th Framework Programme. The VPH is being developed to tackle many is-
sues in today’s medical industry including high drug research and develop-
ment (R&D) costs, declining rate of success with drug development and to
create a better method for discovering new drug treatments.
The Body as a System. In the past, doctors have looked at individual parts of
the human system to fix problems. This has worked for us to a degree. The
problem with this method is that we are not individual organs in a body, we
are an entire complex system. A treatment that may fix an immediate prob-
lem in a specific area can have lasting side-effects for another part of the sys-
tem. The VPH enables us to look at the effects on the entire system at once
allowing us to more accurately predict the side effects. Constructing the VPH
in such a way is like “putting ‘Humpty Dumpty’ back together again in a con-
certed effort to explain how each and every component in the body works as
part of the integrated whole.” [6].
A VPH will be revolutionary for many reasons. It will benefit patients, re-
searchers and doctors alike. The VPH will enable an entirely new method of
drug testing and development. Allowing doctors to test drugs on a simulation
modeled after an individual patient will save time, money and potential side-
effects for the patient. In addition, during drug development, we will poten-
tially be able to entirely skip animal testing. Animal testing only tells us so
much; when a drug moves on to clinical trials with humans it can still run into
many issues for several reasons. Every human is different in age, weight, gen-
der, etc. A VPH could be configured to a specific patient to see how a drug
will affect them.
The Cost of Research and Development. The cost of current drug research
and development is at an all-time high. It is difficult to determine the exact
price for drug R&D because companies use high R&D prices to justify more
expensive drugs. A study done by Tufts Center for the Study of Drug Develop-
ment in 2014 places the price at about $2.6 billion[7], A 2010 study by Steve
Paul et al. places the price at $1.8 billion[8], and an analysis done by Forbes
places the price even higher at $5 billion[9]. Figure 1 presents the change in
average cost for R&D per drug over the last few decades. Regardless of the
exact number, we know that the price is continuing to increase each year as
a result of higher R&D costs and higher failure rates. Using the VPH as a
method of R&D will not only save years’ worth of time on clinical trials but
also billions of dollars in man-hours, R&D and materials for the drugs.
One more piece of the puzzle is our ability to model chemical reactions.
Many of us have made physical models of atoms or chemical compounds for
our Introduction to Chemistry courses in our pre-University and University
years but this is much more complex than that. The idea here is to simulate
chemical reactions using computer technology. This technology is no longer
the future. The work of Martin Karplus, Michael Levitt, and Arieh Warshel for
the “Development of Multiscale Models for Complex Chemical Systems,”[2]
which won them the Nobel Prize in Chemistry 2013, does exactly this. The
importance of their work is that they accomplished something in computer
modeling that had not been managed before.
For many years computers have had limits on their ability to assist in chemis-
try theory. This was due to the difficulty of combining classical physics and
quantum physics in simulations. Using Newton’s classical physics, chemists
could only observe a molecule in a state of rest. In order to observe a chemi-
cal reaction, they must take into account quantum physics. The downside of
both quantum and classical physics simulations is the amount of computer
power that they consume. “The computer has to process every single elec-
tron and every atomic nucleus in the molecule. This can be compared to the
number of pixels in a digital image.”[4] This meant that calculations on large
molecules would have taken a computer years to finish using extreme
amounts of data.
Karplus, Levitt, and Warshel discovered a method for simulating chemical re-
actions using both quantum physics and classical physics together. Prior to
this, computer simulations were only able to consider either quantum phys-
ics or classical physics but not both at once. This discovery allows for a much
more accurate prediction of chemical reactions. It also opens up the ability to
simulate reactions without the extreme time constraints and on a larger
scale than was previously capable.
The original work on these simulations began in the early 1970’s. Since then
not only has our understanding and ability to simulate these reactions im-
proved but the computers we use to simulate them have become signifi-
cantly more powerful. The next step is meshing our knowledge of modeling
chemical reactions with our knowledge of modeling the human body.
5 The Combination
You may be wondering how cognitive computing can play a role in drug de-
velopment. Cognitive computing is already playing a role in medical diagno-
ses. This can be observed with IBM’s new Watson Health technology. Wat-
son is design to assist doctors by analyzing a patient’s medical information
and making appropriate treatment suggestions. One thing that makes Wat-
son so groundbreaking is that all the information is processes is available for
every computer running the software. As a result, Watson can constantly im-
prove its knowledge from many different locations at once. Doctors can add
their own input to Watson’s suggestions further improving its knowledge.
Watson is a cognitive device because of its ability to teach itself. This is
known as machine learning1. Machine learning is a computer’s ability to build
off of its existing knowledge to improve its ability in regards to a specific
topic. For example, self-driving cars2 learn from their mistakes and improve
upon their ability to stay on course.
The concept behind Watson Health can be extended past medical diagnoses
into the field of drug creation. The first step would be linking together VPH’s
with the technology to simulate chemical reactions. This will allow us to cre-
ate new chemicals and immediately test them on virtual humans. Medical
data can be gathered from large groups of people of all ages and health
backgrounds and tested against these new drugs. We would no longer be
limited to long clinical trials or small test groups. It could take weeks to see
results in a human but with cognitive computing we would only be limited by
the speed of our computers. It is predicted that pharmaceutical drugs will be
tested against computer simulations within 5 years [6].
These new computers can learn how certain compounds effect the body
through testing and human assistance. They can observe a specific drug’s ef-
fect on the body, note the effects and make appropriate adjustments in the
drug’s composition. Due to the computer’s ability to simulate chemical reac-
tions, a mastery of these reactions will be developed. Additionally, the VPH
can have specific ailments programmed into it for testing. Testing does not
need to strictly occur on healthy humans. For example, we will be able to
test drugs on theoretical cancer patients or patients with multiple sclerosis.
Self-contained. The use of both VPH and reaction simulation technology al-
lows this device to be completely self-contained. The drugs can be created
within the system and then immediately tested without having to waste time
physically creating the drug to be tested on humans. Using simulations to
test in a predictive manner allows us to tackle medical problems before they
affect people rather than current methods of attempting to treat illnesses af-
ter they have affected people3.
Much like Watson Health, we want to have all of the computers running this
technology pooling their information together. This is a crucial piece of the
puzzle. A method for comparing data will be essential because it will allow
for better, more frequent improvements.
Databanks. Information can be stored at a central location for all of the de-
vices to access. This can be accomplished through cloud services or any alter-
native data storage technique. “A depository of patient data would be help-
ful for model designers… and model users (in certain fields, e.g. drug devel-
opment). An important milestone to make progress in these directions is the
development of these large databases”[11]. By storing information from vari-
ous patient models all over the world, these computers will always be im-
proving their understanding by making comparisons from hundreds or more
locations at a time.
3 “It is interesting to note that Virtual environment training systems are at the same stage of
development as airplane simulators were in the late 1930’s, and airplane simulators were
not accepted as valid training devices until 1955”[11].
6 Conclusion
7 References
1. U.S. Food and Drug Administration. (2014, Nov. 10). How Drugs are Developed and Ap-
proved [Online]. Available: http://www.fda.gov/Drugs/DevelopmentAp-
provalProcess/HowDrugsareDevelopedandApproved/
2. M. Karplus, M. Levitt, A. Warshel. (2013, Oct. 9). The Nobel Prize in Chemistry 2013
[Online]. Available: http://www.nobelprize.org/nobel_prizes/chemis-
try/laureates/2013/
4. M. Karplus, M. Levitt, A. Warshel. (2013, Oct. 9). The Nobel Prize in Chemistry 2013
Popular Information [Online]. Available: http://www.nobelprize.org/no-
bel_prizes/chemistry/laureates/2013/popular-chemis-
tryprize2013.pdf
7. S. Peters. (2014, Nov 18). Cost to Develop and Win Marketing Approval for a New Drug
Is $2.6 Billion [Online]. Available: http://csdd.tufts.edu/news/com-
plete_story/pr_tufts_csdd_2014_cost_study
8. S. Paul, et al. (2010, March). How to Improve R&D Productivity: The Pharmaceutical In-
dustry’s Grand Challenge [Online]. Available: http://www.na-
ture.com/nrd/journal/v9/n3/full/nrd3078.html
9. M. Herper. (2013, Aug 11). The Cost of Creating a New Drug Now $5 Billion, Pushing
Big Pharma to Change [Online]. Available:
http://www.forbes.com/sites/matthewherper/2013/08/11/how-
the-staggering-cost-of-inventing-new-drugs-is-shaping-the-
future-of-medicine/
11. N. Ayache, et al. (2005, Nov). Towards Virtual Physiological Human: Multilevel Model-
ling and Simulation of the Human Anatomy and Physiology [Online]. Available:
http://www.vph-institute.org/upload/file517569145f61b.pdf
12. C. Sansom, M. Mendes, P. Coveney, “Modelling the Virtual Physiological Human,” Bio-
Technologia, Poland, Vol. 92(3), 2011
Management of existential Risks of Applications
leveraged through Cognitive Computing.
Robert Richter
1 Introduction
During the past half of an decade digital technologies made huge advancements
in fields which were preserved only to humans. Although this is not an indicator
for the possibility of intelligent machines which capabilities could match those
of humans, scientists agree those machines will be put into existence someday.
Therefore it is necessaryto research risks those concepts might bring. This paper
will discuss possible technical and organizational methods to improve the fu-
ture coexistence of machine and man and obtain the best assistance for humans
through these technologies.
The focus of the next chapter (chapter 2) is on the advancements of digital
technologies which were mentioned above as well as possible paths to computers
with a general intelligence. After the state of the art and future developments
are discussed, chapter 3 concentrates on the necessity of technology assessments
and why it is about time to discuss possible risks of cognitive computing. Risks
of AI names several kinds of risks connected to cognitive computing like eco-
nomic ones but concentrates on the existential kind which are the focus of this
paper. It further explains the need for technology assessment before chapter 4
II
presents possible measures to reduce existential risks which were introduced be-
fore. Finally, before the conclusion 6, viewpoints will be displayed that don’t see
harmful risks in the design of intelligent machines in chapter 5.
3 Risks of AI
Before we further examine, what risks the development of an human-like level
of general intelligence could involve, we want to define something we previously
revered to as existential risks and thereby delimit which kind of risks we want
to further discuss in this paper.
As machines will become more and more capable of tasks currently exclusive to
humans they might suppress humans in those fields. The possible consequences
for the labour market or other economic risks are not subject of this paper. Also
threats to privacy or human dignity which could result as machines might replace
people in positions that require care and respect like a nursemaid for the elderly
are just stated for the sake of completeness. This paper limits itself to existential
risks as those give the most urgent motivation to discuss possible measurements
to develop an intelligent agent which is aligned to human interests. Existential
risks, as stated by Bostrom [1] are risks where “ an adverse outcome would either
annihilate Earth-originating intelligent life or permanently and drastically curtail
its potential”. This definition also holds scenarios where humanity doesn’t get
extinct but loses the control over their own development and is not capable of
making any further necessary progress. Figure 1 [1] shows different kinds of risks
to better delimit which type is discussed in this paper. It further describes an
existential risk as terminal and transgenerational, meaning that it will not only
affect every currently living human being but also all generations to come.
the first ultra-intelligent machine is the last invention that man need ever make
[10].
So, I. J. Good observed that the design of an AGI might create a positive feed-
back loop leading to an intelligence explosion, meaning that human researchers
don’t have to design an super-intelligent agent from scratch, but rather assign-
ing this task, intended or not, towards smarter-then-human systems. The In-
telligence Explosion is by [3] also referred to as technological singularity in a
sense that from our point of view and knowledge of today, we cannot see past
it. That’s why it is important to distinguish between propositions we can make
from this point of view and those which a uncertain, to evaluate possible risks.
This chapter will discuss possible technical measures and research directions to
improve the future coexistence of intelligent machines and humans and obtain
the best assistance for humans through these technologies. Most measures have
in common that researchers can work on them today which is necessary, as once
an ASI exists they have to be already in place to alter the outcome of super-
intelligent agents in a positive way [1]. Researchers organised at MIRI 3 (Machine
Intelligence Research Institute) further argue, that there are many examples for
3
A non-profit organization which researches safety issues related towards AGI
VII
is a concept of designing those agents in a way that they reason as if they were
incomplete and potentially awed in dangerous ways so that they stay amenable
for correction. It adds error tolerance so that human errors that might and prob-
ably will happen throughout the design process can be corrected.
Even if we manage to implement all the previous mentioned methods in time
it is not assured that an ASI will have a positive and beneficial impact. It still
depends on how appropriate the goals, that the ASI is told to achieve, are de-
fined. A trivial example for a ill defined goal could be: “Find a cure for HIV.”
With no further restrictions, an ASI with sufficient resources could simply try
to kill everybody with the virus in order to eliminate it, which was certainly not
what the operator wanted. What is needed is a design of the super-intelligent
agent that considers the preferences of its operators. This is called the Value
learning problem. One first approach is to let it learn inductively from train-
ing data. Problems which are faced here include what dataset provides useful
information in a way that it gives the smart agent the opportunity to fully learn
the complexities of values and how to model the volition of the operator.
The strategies stated in this chapter are not complete. Nevertheless it gives an
overview about the challenges which have to be faced to make sure that once ASI
happens it is possible for humanity to control and restrain it [8]. The selection
is based on two principles. They cover the mayor challenges which have to be
fulfilled in order to ensure an intelligent agent is aligned and the necessary basic
knowledge is already there on which future research can be build. Table 1 lists
the fields of research stated in this chapter and the problem space they cover to
construct a minimal reliable generally intelligent system.
5 Critics
Regarding the possibility of an artificial general intelligence exist three major
viewpoints. The one represented in this paper argues that general intelligent
agents will be put into existence in the not so far away future and that it
holds existential risks to humanity without proper preparations. Critics from
IX
another viewpoint dispute that those agents will ever be constructed and chal-
lenge the usefulness of such an discussion. Sociologists like Dickel talk about
super-intelligent agents as utopias, which won’t become reality in the way they
are predicted by AGI-critics like Bostrom [8] or AGI-proponents like Kurzweil
[12]. He also states, that the scientific work of organisations with focus on the
research of aligned systems like the machine intelligence research institute is in-
significant and delimited to public relations work [13]. However, he admits that
cognitive computing will experience advances in narrow tasks, but actors like the
Defense Advanced Research Projects Agency (DARPA) will play a bigger role
in those advances as it spends billions for the neurological armament of soldiers.
As Dickel [13] sees no possible development towards AGIs the corresponding
risks would be of another nature then focused here. Researchers in the field of
cognitive computing who support this viewpoint are very view and chapter 3.2
already stated reasons which counter these notions. The last viewpoint acknowl-
edges the possibility of super-intelligent agents in the near future but reasons
that it will happen in a fundamentally human-friendly way with an positive
impact by default. One of the most important representatives of this camp is
R. Kurzweil, who is well-known for his accurate predictions, like his prediction
of the explosive growth in worldwide internet use in the early 90’s. In [12] he
examines the possibilities of artificial super-intelligence together with nanotech-
nology and genetics and even considers potential dangers of these technologies.
He justifies his opinion about the positive impact in saying: “[ASI] is emerging
from many diverse efforts and will be deeply integrated into our civilisations in-
frastructure. Indeed, it will be intimately embedded in our bodies and brains. As
such, it will reflect our values because it will be us.”. Again, chapter 3.2 contains
arguments which object this statement. Also, even in applications of weak ai the
dangers and absence of the reflection of our values was demonstrated as provable
through the 2010 flash crash [1].
6 Conclusion
our lack of knowledge about the nature of an intelligence superior to ours makes
it hard to make propositions about the other two possible outcomes. Their is
only agreement that either way, a super-intelligence will have an huge impact
on humanity. It will be either very beneficial in various ways for humanity or
disastrous. After studying the various possible risks and the strategies to com-
pete with them it seems very unlikely that the first super-intelligence will be
beneficial by default. Fundamental knowledge in various topics stated in chapter
4 seem necessary to ensure the profitable coexistence of humans and smart ma-
chines. One argument that theory might precede application as required is that
progress in research fields that provide knowledge to assure aligned intelligent
systems will also have advantages for designing intelligent agents in general. This
will give further motivation to develop this knowledge in time.
On the other hand, significant research efforts are focused on developing and im-
proving AGI’s and very little on the alignment of those systems through topics
mentioned above. Reason for optimism here are teams and organisations like the
machine intelligence research institute or the future of humanity institute that
are just been founded during the last decade and which are dedicated to put
more effort in to the research of aligned systems.
What progress those organisations can make and how they will influence the
development of an general intelligent system might determine the outcome of an
intelligence explosion.
Another characteristic observed during the study of the discussion whether an
ASI will be beneficial or harmful is that most of the researches who don’t see
any risks by super-intelligent systems are computer scientists and involved in
the progress of developing artificial intelligence like Ray Kurzweil. Scientists who
express their concerns about the impact of such systems are mostly from other
disciplines which are somehow connected to computer science like mathematics
or philosophy. Whether this fact can value the opinion of the latter group on
this subject somehow negative or their specific fields of research are those where
scientists who are directly involved in the improvement of cognitive computing
should pay more attention to is open for further discussion.
References
1. Bostrom, N.: Superintelligence - Paths, Dangers, Strategies. Oxford University
Press, Oxford, UK. ISBN 978-0-19-967811-2 (2014)
2. Brynjolfsson, E., Mc Affee, A.: Race Against the Machine - How the Digital Rev-
olution is Accelerating Innovation, Driving Productivity, and Irreversibly Trans-
forming Employment and the Economy. Digital Frontier Press, Lexington, MA,
USA, ISBN 978-0-9844725-11-3 (2014)
3. Brynjolfsson, E., Mc Affee, A.: The Second Machine Age - Work, Progress and
Prosperity in a Time of brilliant Technologies. W.W. Norton & Company, Inc.,
N.Y., USA, ISBN 978-0-393-23935-5 (2014)
4. Georges, T.M.: Digital Soul Intelligent Machines and Human Values. Westview
Press (Perseus Book Group), Oxford, UK, ISBN 0-8133-4266-X (2003)
5. Soares, N., Fallenstein, B.: Aligning Superintelligence with Human Interests: A
Technical Research Agenda. intelligence.org (2014)
XI
6. Muehlhauser, L.: What is AGI?. August 11, 2013, post on Machine Intelligence
Research Institute https://intelligence.org/2013/08/11/what-is-agi/
7. Moor, J.H.: Why We Need Better Ethics for Emerging Technologies. Ethics and
Information Technology September 2005, Volume 7, Issue 3, pp 111-119 (2005)
8. Bostrom, N.: ACM OPINION - You Should Be Terrified of Superintelligent Ma-
chines. by Slate, Article, September 11, 2014
9. Lem, S.: Summa technologiae. Krakow: Wydawnictwo Literackie 1964, Insel Verlag
Frankfurt am Main (1976)
10. Good,I.J.: Speculations Concerning the First Ultraintelligent Machine. Advances
in Computers, vol. 6, 1965
11. Yudkowsky,E.: Artificial Intelligence as a Positive and Negative Factor in Global
Risk. Global Catastrophic Risks New York: Oxford University Press. 2008
12. Kurzweil,R.: The Singularity Is Near: When Humans Transcend Biology Published
by the Penguin Group, Canada. 2005
13. Dickel,S.: Entgrenzung der Machbarkeit? Biopolitische Utopien des Enhancements.
In Bhlemann, Peter, ed. Der machbare Mensch?: moderne Hirnforschung, biomedi-
zinisches Enhancement und christliches Menschenbild. Vol. 13. LIT Verlag Mnster,
2010.