Proceedings of the 39th Annual Meeting of the Cognitive Science Society (2017). London, pp.

829-834

Multitasking Capability Versus Learning Efficiency

in Neural Network Architectures
Sebastian Musslick1,⇤ , Andrew M. Saxe2 , Kayhan Özcimder1,3 ,
Biswadip Dey3 , Greg Henselman4 , and Jonathan D. Cohen1
1 Princeton
Neuroscience Institute, Princeton University, Princeton, NJ 08544, USA.
2 Center
for Brain Science, Harvard University, Cambridge, MA 02138, USA.
3 Department of Mechanical and Aerospace Engineering, Princeton University, Princeton, NJ 08544, USA.
4 Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104, USA.
⇤ Corresponding Author: [email protected]

Abstract a time. In this view, the constraints on the number of control-

dependent tasks that can be executed at one time reflect an
One of the most salient and well-recognized features of human
goal-directed behavior is our limited ability to conduct mul- intrinsic property of the control system itself. However, alter-
tiple demanding tasks at once. Previous work has identified native (“multiple-resource“) accounts (Allport, 1980; Meyer
overlap between task processing pathways as a limiting fac- & Kieras, 1997; Navon & Gopher, 1979; Salvucci & Taat-
tor for multitasking performance in neural architectures. This
raises an important question: insofar as shared representation gen, 2008) have suggested that multitasking limitations arise
between tasks introduces the risk of cross-talk and thereby lim- from local processing bottlenecks. That is, if two tasks share
itations in multitasking, why would the brain prefer shared task the same local resources (i.e. representations required to per-
representations over separate representations across tasks? We
seek to answer this question by introducing formal considera- form the tasks), then executing them simultaneously can lead
tions and neural network simulations in which we contrast the to cross-talk and degraded performance. It has been argued
multitasking limitations that shared task representations incur that the very purpose of cognitive control is to prevent such
with their benefits for task learning. Our results suggest that
neural network architectures face a fundamental tradeoff be- cross-talk by limiting the number of active task processes
tween learning efficiency and multitasking performance in en- engaged (Cohen, Dunbar, & McClelland, 1990; Botvinick,
vironments with shared structure between tasks. Braver, Barch, Carter, & Cohen, 2001). In this view, con-
straints in multitasking reflect the consequences of control
Keywords: multitasking; cognitive control; capacity con- doing its job, rather than limitations intrinsic to the mecha-
straint; learning; neural networks nisms of control itself. This line of argument suggests that,
to better understand the conditions under which multitasking
Introduction is and is not possible, it is necessary to understand the extent
Our limited capability to execute multiple tasks at the same to which the task processes involved share representations,
time highlights one of the most fundamental puzzles con- and are thus subject to potential interference and the inter-
cerning human processing, which must be addressed by any vention of control to limit processing. This, in turn, raises
general theory of cognition (Shenhav, Botvinick, & Cohen, the question of whether there are general principles of neural
2013; Kurzban, Duckworth, Kable, & Myers, 2013; Ander- architectures that determine the use of shared representation,
son, 2013): Why, for some tasks, is the human mind capa- and how these interact with learning and processing.
ble of a remarkable degree of parallelism (e.g., navigating a One may argue that the constraints that shared represen-
crowded sidewalk while talking to a friend), while for others tations impose on multitasking are negligibly small in a pro-
its capacity for parallelism is radically limited (e.g., conduct cessing system as large as the human brain. However, simula-
mental arithmetic while constructing a grocery list)? tion studies (Feng, Schwemmer, Gershman, & Cohen, 2014),
Early theories of cognition, that have continued to be followed by analytic work (Musslick et al., 2016) have found
highly influential, assert that the ability to multitask – that is, that the multitasking capability of a network can drop precip-
to carry out a set of tasks concurrently1 – can be understood itously as a function of overlap between task processes (i.e.
in terms of a fundamental distinction between automatic and number of shared representations), and that this effect is rel-
controlled processing, with the former relying on parallel pro- atively insensitive to the size of the network.
cessing mechanisms (that can support multitasking) and the
The findings above suggest that maximal parallel process-
latter assumed to rely on a serial processing mechanism with
ing performance is achieved through the segregation of task
limited capacity (Posner & Snyder, 1975; Shiffrin & Schnei-
pathways, by separating the representations on which they
der, 1977) that can only support processing of a single task at
rely. This raises an important question: insofar as shared rep-
1 Multitasking can, in some situations, be achieved by rapid se- resentation introduces the risk of cross-talk and thereby lim-
quential processing (e.g., switching between asynchronous serial itations in parallel processing performance, why would the
processes, as is common in computers), rather than through true brain prefer shared task representations over separate ones?
synchronous processing. Here, our focus is on forms of multitask-
ing that reflect truly concurrent processing, sometimes referred to as Insights gained from the study of learning and representation
perfect timesharing or pure parallelism. in neural networks provide a direct answer to this question:
Shared representations across tasks can support inference and weight projections from each task unit can act as control sig-
generalization (Caruana, 1997). These benefits are strongly nals that bias processing towards task-relevant stimulus infor-
linked to the ability of neural networks to carry out “inter- mation represented at the associative and output layer.
active parallelism“, that is, the ability to learn and to pro- In order to represent the task environment described below,
cess complex representations by simultaneously taking into the stimulus layer is compromised of 45 input units (features)
account a large number of interrelated and interacting con- and the task layer of nine task units. The output layer con-
Network
straints (McClelland, Rumelhart, & Hinton, 1986). sists of 15 units and is organized into three response dimen-
In this study, we examine the tension between interactive sions (with five units per response dimension.). The number
parallelism that promotes learning efficiency through use of of units in the associative layer is set to 100.
shared representations, on the one hand, and “independent Output
parallelism“ (i.e. the ability to carry out multiple processes 
yo
independently), on the other hand. That is, we are interested 1
in studying biases that promote shared representations over neto = ∑ woh yh +∑ wot xt + θ o yo =
1+ e−neto
multitasking performance. We first demonstrate that the well- h t
who
recognized (and valued) emergence of shared representations Associative (Hidden)
woh
(Hinton, 1986) in response to extrinsic biases (i.e. shared 
yh …
structure in the task environment) leads to constraints in mul-
1
titasking performance. In the second part, we introduce a neth = ∑ whs xs +∑ wht xt + θ h yh =
formal characterization of a tradeoff between learning effi- s t 1+ e−neth wht
ciency and multitasking performance and examine how in- Stimulus
whs Task
 
trinsic biases of the network toward the use of shared rep- xs … xt …
resentations can expose this tradeoff in neural network sim-
ulations. The source code for all simulations is available at Figure 1: Feedforward neural network used in simulations.
github.com/musslick/CogSci-2017. The input layer is composed of stimulus vector ! xs and task
vector !xt . The activity of each element in the associative
Neural Network Model layer yh 2 !yh is determined by all elements xs and xt and their
respective weights whs and wht to yh . Similarly, the activity
For the simulations described in the paper we focus on a net-
of each output unit yo 2 ! yo is determined by all elements yh
work architecture that has been used to simulate a wide array
and xt and their respective weights woh and wot to yo . A bias
of empirical findings concerning human performance (e.g.
of q = 2 is added to the net input of all units yh and yo .
Cohen et al., 1990; Botvinick et al., 2001), including recent
Blue shades in the input and output units (circles) correspond
work on limitations in multitasking (Musslick et al., 2016).
to unit values of > 0 and illustrate an example input pattern
In this section we lay out the architecture of this network, its
with its respective output pattern: The second task requires
processing, as well as the task environments used to train it.
the network to map the vector of values in the first five stim-
Network Architecture and Processing ulus input units to one out of five output units (yellow shade).

The network consists of two input layers, one of which repre- Task Environment
sents the stimulus presented to the network and another that Each task is defined as a mapping between a subspace of five
encodes the task that the network is instructed to perform on stimulus features (referred to as a task-relevant stimulus di-
the stimulus. Stimulus input features can take any real value mension) onto five output units of a task-specific response
between 0 and 1 and can be grouped into stimulus dimen- dimension, so that only one of the five relevant output units is
sions that are relevant for a particular task. The network is permitted to be active (see Fig. 1). The value of each stimu-
instructed to perform a single task by clamping the corre- lus feature is drawn from a uniform distribution U[0, 1]. The
sponding task unit in the task layer to 1 while all other task rule by which 5 relevant stimulus features of any task-relevant
units are set to 0. These stimulus and task input values are stimulus dimension are mapped onto one of the 5 output units
multiplied by a matrix of connection weights from the re- of the task-relevant response dimension corresponds to a non-
spective input layer to a shared associative layer, and then linear function that was randomly generated2 with a sepa-
passed through a logistic function to determine the pattern of rate “teacher“ network (cf. Seung, Sompolinsky, & Tishby,
activity over the units in the associative layer. This pattern is 1992), and is the same across tasks. However, tasks are con-
then used (together with a set of direct projections from the sidered to be independent in that they differ which stimulus
task layer) to determine the pattern of activity over the output dimension is linked to which response dimension.
layer. The latter provides a response pattern that is evaluated The task environment across all simulations encompasses
by computing its mean squared error (MSE) with respect to nine tasks. As illustrated in Fig. 2 groups of three tasks map
the correct (task-determined) output pattern. Similar to stim- 2 Note that it is ensured that, for the uniform distribution U[0, 1]
ulus features, output units can be grouped into response di- of stimulus unit activations in the task-relevant set of input units, ev-
mensions that are relevant for a particular task. Note that the ery relevant output unit is equally likely to be required for execution.
onto the same response dimension. However, similarity be- no relevant stimulus features (cf. Fig. 2a), all five stimu-
tween tasks could be varied by manipulating the overlap be- lus features (cf. Fig. 2b) or any whole number of features
tween their relevant stimulus dimensions. At the extremes, in between, resulting in 6 different task environments. We
task environments can be generated such that tasks of differ- Task
trained 100 networks in each of the environments. The net-
Structure
ent response dimensions relate to separate stimulus features works were trained on all nine tasks with the same set of 50
(no feature overlap, Fig. 2a), or the same stimulus features stimulus samples until the network achieved an MSE of 0.01.
(full feature overlap, e.g. tasks 1-3 in Fig. 2b). Feature O
(a) Feature Overlap Between Tasks (b)

Multitasking Accuracy (%)

1
(a) No Overlap 75
1
 Response Units

Multitasking P(Correct)
Between Task- 0.75 70
0.8

Feature Overlap
Relevant Stimulus yo 0.8

Feature Overlap
0.7 0.6
65
Features 0.65 0.6
60 0.4
!" !*
!$ !) 0.6 0.4 55
!# ! !& !' !( 0.2
% 0.55
 0.2 50
xs 0.5
0 0.5 1
0
0
Stimulus Features 0.4 0.6 0.8 1 Learned Task Correlation
Learned Weight Correlation
(b) Full Overlap  Response Units Figure 3: Effects of task similarity. (a) Networks were trained
Between Task- yo in task environments with varying degrees of feature overlap.
Relevant Stimulus
Yellow and red shades highlight task-relevant stimulus fea-
Features !" !& !*
tures for two tasks involving different response dimensions.
!$ !)
!# !% !' !(
(b) Final multitasking accuracy of the network as a function
 of the learned similarity between tasks involving different
xs
Stimulus Features response dimensions. Colors indicate the degree of feature
Figure 2: Task environments. For each task, the network was overlap present in the task environment as illustrated in (a).
trained to map a subset of 5 stimulus features onto a subset of In order to assess the similarity of learned task representa-
5 output units. At the extremes tasks that were mapped onto tions we focus our analysis on the weights from the task units
different response dimensions (e.g. tasks 1-3) could either (a) to the associative layer, insofar as these reflect the compu-
rely on separate stimulus features or (b) completely overlap tations carried out by the network required to perform each
in terms of their relevant stimulus features. task. For a given pair of tasks we compute the learned repre-
Networks are initialized with a set of small random weights sentational similarity between them as the Pearson correlation
and then trained on all tasks using the backpropagation algo- of their weight vectors to the associative layer.
rithm3 (Rumelhart & Geoffrey E. Hinton, 1986) to produce We measured multitasking performance for pairs of tasks
the task-specified response for each stimulus. (of different stimulus and response dimensions) by activating
two task units at the same time and evaluating the concurrent
Multitasking Limitations Due to Shared processing performance in the response dimensions relevant
Structure in the Task Environment to the two tasks. The accuracy of a single task ASingle can be
A key feature of neural networks is their ability to discover computed as ac
Asingle = 5 (1)
latent structure in the task environment, exploiting similarity Âi=1 ai
between stimulus features in the form of shared representa-
tions (Hinton, 1986; Saxe, McClelland, & Ganguli, 2013). where ai is the activation of the ith output unit of the task-
In this section we explore how the emergence of shared rep- relevant response dimension and ac is the activation of the
resentations as a function of structural similarities between correct output unit. The multitasking accuracy is simply the
tasks can impact the multitasking performance of a network. mean accuracy of both engaged single tasks.
The simulation results confirm well-known explorations in
Simulation Experiment 1: Shared Task neural networks (Hinton, 1986; McClelland & Rogers, 2003;
Representations as a Function of Feature Overlap Saxe et al., 2013) that task similarities in the environment
In order to investigate the effect of structural similarities be- can translate into similarities between learned task represen-
tween tasks we generated task environments with varying tations. Critically, this extrinsic bias toward the learning of
overlap between task-relevant stimulus features. We define shared representations negatively affected multitasking per-
feature overlap as the number of relevant stimulus features formance (Fig. 3b). To illustrate this, consider the simultane-
that are shared between any pair of tasks linked to different ous execution of tasks 1 and 5 in an environment as depicted
response dimensions (see Fig. 3a). That is, two tasks in- in Fig. 2b. If the network learns similar representations at the
volving two different response dimensions could either share associative layer for tasks 1 and 2 (note that both tasks rely on
3 All reported results were obtained using gradient decent to min- the same stimulus features), then executing task 1 will implic-
imize the MSE of each training pattern. However, we observed the itly engage the representation of task 2 which in turn causes
same qualitative effects using the cross-entropy loss function. interference via its link to the response dimension of task 5.
 Multitasking Limitations due to Intrinsic !$ !"
(a) !#
Output gating
signal
Learning Biases
Associative gating
In addition to environmental biases that shape the learning of signal
shared task representations there may be factors intrinsic to #
&'(
*
&'(

the neural system that can regulate the degree to which such
%# %$ %"
representations are exploited in learning. In this section we
Group 1 Group 2
introduce a formal analysis of how such biases can affect the (b)

!# !$ !"
tradeoff between learning efficiency and multitasking perfor- Output gating signal
mance. We then use weight initialization as a learning bias Group 1 Group 2
in simulations to establish a causal relationship between the
use of shared representations on the one hand, and resulting *,$
&'(
#,$
&'( *,# Associative
#,# &'(
effects on learning and multitasking, on the other hand. &'( gating
signal

Formal Intuitions on the Tradeoff between Learning %# %$ %"

Efficiency and Multitasking Capability Figure 4: Gating model used for formal analysis. (a) Task
information directly switches on or off task-relevant dimen-
To gain formal intuition into the tradeoff between multitask-
sions in the output and associative layers. This allows input-
ing ability and learning speed, we consider a stripped-down
to-hidden weights to be shared across the M different tasks
version of the introduced network model that is amenable to
corresponding to differentpresponse dimensions, increasing
analysis. In the full model, nonlinear interactions between
learning speed by a factor M. However, two tasks that rely
the task units and the stimulus units occur in the associative
on different input dimensions cannot be multitasked due to
layer. Here we assume a gating model in which these non-
crosstalk at the output (convergent red and green arrows). (b)
linear interactions are carried out through gating signals that
Multitasking ability can be improved by separating response
can zero out parts of the activity in the associative and output
dimensions into Q groups, each with a dedicated set of units
layers, or pass it through unchanged. The choice of which
in the associative layer. Gating now permits one task from
parts of each layer are gated through on each input is left to
each group to operate concurrently (red and green arrows no
the designer (not learned, as in the full model).
longer converge). However, weightpsharing is limited to the
We study the scheme depicted in Fig. 4 consisting of M
group, yielding a learning speed of M/Q.
input and response dimensions with full feature overlap (cf.
Fig. 2b). For the output layer, we assume that the gating learning speed and multitasking ability. Let t be the number
variables automatically zero all but the task-relevant response of iterations required to learn all tasks, Q the maximum num-
dimensions. For the associative layer, we separate the hid- ber of concurrently executable tasks, and M the number of
den units into dimensions, one for each input dimension, and input/response response dimensions. Then
make the gating variables zero all representations except the
t 2 µ Q/M (2)
one coming from the task-relevant input dimension (Fig. 4a).
Crucially, when the gating structure is known on a specific where the proportionality constant is related to the statisti-
example, the output of the network is a linear function of the cal strength of the input-output association for one task, the
neurons that are on. Given this setting, the learning dynamics learning rate, and the error cut-off used to decide when learn-
can be solved exactly using methods developed by Saxe, Mc- ing is complete (Saxe, Musslick, & Cohen, 2017).
Clelland, and Ganguli (2014). The key advantage afforded by Due to the tradeoff in Eqn. (2), gating schemes that share
the gating scheme is depicted in Fig. 4a: the input-to-hidden more structure will learn more quickly. Hence generic, ran-
weights for one input dimension can be shared by all tasks p domly initialized nonlinear networks will tend to favor shared
that rely on that input dimension. This leads to a factor M representations, as shown in Simulation Experiment 1.
speedup in learning relative to learning a single task by itself
(proof omitted due to space constraints). Simulation Experiment 2: Effects of Learning
However, with this gating system, multitasking is not pos- Biases for Shared Representations
sible: gating another task through to the output will lead to In Simulation 2 we focus on a bias intrinsic to the neural sys-
interference. To counteract this, the gating scheme must be tem, i.e. the initialization of the weights from the task layer.
changed: response dimensions can be divided into Q groups, We use this factor to systematically examine how the use of
each with a dedicated set of hidden units (Fig. 4b). This al- shared representations facilitates the discovery of similarity
lows tasks that use response dimensions in different output structure while diminishing multitasking performance. To do
groups to be performed simultaneously. Hence a maximum so, we focus initially on a training environment in which tasks
of Q tasks can be performed simultaneously, but weight shar- are maximally similar, as this is the condition in which there
ing is reduced across tasks by a factor Q, slowing learning. is most opportunity for exploiting shared representations. We
This analysis provides, at least in a simplified system, a then examine environments with 80% and 0% feature overlap
quantitative expression of the fundamental tradeoff between between tasks, to test the generality of the observed effects.
 To manipulate the bias towards shared task representations, sources of this capacity constraint associated with control re-
we initialized the weights from the task units to the associa- main largely unexplored. Here, we build upon the observation
tive layer, varying the similarity among the weight vectors that multitasking limitations can arise from shared representa-
across tasks with the rationale that greater similarity should tions between tasks (Feng et al., 2014; Musslick et al., 2016),
produce a greater bias toward the use of shared representa- and use a combination of formal analysis and neural network
tions in the associative layer. Weight vectors for tasks relying simulations to examine biases towards shared representations
on the same stimulus input dimensions were randomly ini- that incur such costs in multitasking.
tialized to yield a correlation coefficient of value r. The cor- In the first part of this study, we build upon early insights of
relation value r was varied from 0 to 0.975 in steps of 0.025 connectionism that shared representations emerge as a func-
and was used to constrain initial weight similarities for 100 tion of task similarities in the environment and demonstrate
simulated networks per initial condition. The weight vectors the deleterious consequences for multitasking performance.
for tasks of non-overlapping stimulus dimensions were un- It has been shown that networks are capable of extracting sim-
correlated. Finally, all task weights to the associative layer ilarities from a hierarchically structured input space (Hinton,
were scaled by a factor of 5 to enhance the effects of different 1986). Recent analytic and empirical work in the domain of
initial task similarities. The networks were trained using the semantic cognition paints a similar picture: neural systems
same parameters as reported for Simulation Experiment 1. may gradually discover shared structure in the task environ-
Simulation results indicate that networks with a higher ment with a bias towards the initial formation of shared, low-
similarity bias tend to develop more similar representations dimensional representations (Saxe et al., 2013; McClelland &
at the associative layer for those tasks (in terms of their fi- Rogers, 2003). Our simulation results are in line with these
nal weight vector correlations), whereas a lower similarity observations showing that shared task representations emerge
bias leads to more distinct task representations at this layer. as a function of high stimulus feature overlap between tasks
In environments with high feature overlap between tasks, and furthered the insight that such similarities in the task en-
stronger initial biases toward shared representations lead to vironment lead to multitasking limitations.
increased learning speed (i.e. less iterations required to train
In the second part, we examined how intrinsic learning bi-
the network), as similarities between tasks can be exploited
ases towards shared or separate representations (by means
(Fig. 5a). Critically, this comes at the cost of multitasking
of weight initialization) can be used to expose a tradeoff
performance. Learning benefits gained from shared represen-
between learning efficiency and multitasking performance.
tations are less prevalent in environments with less feature
Early& work in machine learning suggests that learning bi-
Learning Speed
overlap between tasks. However, effects of weight similarity Generalization
ases towards a particular representation can be understood
biases on multitasking impairments remain (Fig. 5b).
as biases of the learner’s hypothesis space (Baxter, 1995),
(a) 100% Feature Overlap (b) Effects Across Task Environments that is, the set of all hypotheses a learner may use to ac-
quire new tasks. We formalized this hypothesis space in terms
Multitasking Accuracy (%)

52 1 90 1
Multitasking Accuracy (%)

0% Feature Overlap
Initial Task Correlation

of the amount of shared representations between tasks and

Initial Task Similarity

50 0.8 80 0.8

48 0.6 70 showed how this mediates an inverse relationship between

0.6

46 0.4 60
100%
Feature Overlap 80%
learning efficiency and interference-free multitasking. Our
0.4

44 0.2 50 Feature Overlap neural network simulations confirmed these analytical predic-
0.2

42 0 40 0 tions, showing that a weight initialization bias towards shared

70 75 80 85
Iterations Required To Train
60 70 80 90 100 110
Iterations Required To Train
120
representations enables faster learning if shared structure in
the environment can be exploited, but incurs a cost for multi-
Figure 5: Effects of weight similarity bias. Mean multitask- tasking. A promising direction for future research may be to
ing accuracy (for two tasks simultaneously) plotted against explore another prediction: our formalism suggests a role for
the mean number of iterations required to train the network. such biases in regularizing the representational complexity of
Data points represent the mean measures across networks ini- the network, thereby promoting generalization performance.
tialized with the same task similarity (constrained by task
Our analyses indicate that neural learning systems, whether
weight vector correlation) for tasks relying on the same stim-
natural or artificial, are subject to a tension between “in-
ulus dimensions. Effects are shown for (a) environments with
teractive parallelism“ on the one hand, which exploits the
100% feature overlap between tasks, as well as (b) across
fine grained structure of representations and similarity in the
environments with different feature overlap. Different data
service of learning, and “independent parallelism“ that sup-
point clusters correspond to different training environments.
ports concurrent processing of distinct tasks, on the other
hand. A similar tension can be found in the domain of learn-
General Discussion and Conclusion ing and memory. The complementary learning systems hy-
The limited ability to perform multiple control-dependent pothesis proposes two separate learning systems, one system
tasks at the same time is one of the most salient character- that relies on shared representations to support inference, as
istics of human cognition, and is universally considered a well as another system that uses separate representations to
defining feature of cognitive control. Despite these facts, the support independent encoding and retrieval of information
 (McClelland, McNaughton, & O’Reilly, 1995). The latter in the hippocampus and neocortex: insights from the suc-
system supports a form of independent parallelism for asso- cesses and failures of connectionist models of learning and
ciational processes that is similar to the form of independent memory. Psychological Review, 102(3), 419.
parallelism for executional processes described in this paper. McClelland, J. L., & Rogers, T. T. (2003, April). The paral-
Altogether our results suggest that the brain may be con- lel distributed processing approach to semantic cognition.
fronted with balancing multitasking capability against extrin- Nature reviews. Neuroscience, 4(4), 310–322.
sic and intrinsic biases towards shared representations. A ma- McClelland, J. L., Rumelhart, D. E., & Hinton, G. E. (1986).
jor goal for the development of artificial systems may be to The appeal of parallel distributed processing. Cambridge,
systematically configure the balance between interactive and MA: MIT Press.
independent parallelism, as well as to exploit the relative ad- Meyer, D. E., & Kieras, D. E. (1997). A computational theory
vantages of each. Most efforts in complex neural architec- of executive cognitive processes and multiple-task perfor-
tures have focused predominantly on the discovery of shared mance: Part I. Basic mechanisms. Psychological review,
representations for the purpose of inference and generaliza- 104(1), 3.
tion (Bengio, Courville, & Vincent, 2013). However, one of Musslick, S., Dey, B., Özcimder, K., Patwary, M. M. A.,
the future challenges will be to explore the tension between Willke, T. L., & Cohen, J. D. (2016). Controlled vs. auto-
learning efficiency and multitasking in networks with higher matic processing: A graph-theoretic approach to the analy-
complexity (i.e. deep networks), as well as in more natu- sis of serial vs. parallel processing in neural network archi-
ralistic task environments. We hope that this work will help tectures. In Proceedings of the 38th annual conference of
inspire a proliferation of efforts to further explore this area. the Cognitive Science Society (pp. 1547–1552). Philadel-
phia, PA.
References Navon, D., & Gopher, D. (1979). On the economy of the
Allport, D. A. (1980). Attention and performance. Cognitive human-processing system. Psychological Review, 86(3),
psychology: New directions, 1, 12–153. 214.
Anderson, J. R. (2013). The architecture of cognition. Psy- Posner, M., & Snyder, C. (1975). attention and cognitive con-
chology Press. trol,. In Information processing and cognition: The loyola
Baxter, J. (1995). Learning internal representations. In Pro- symposium (pp. 55–85).
ceedings of the eighth annual conference on computational Rumelhart, D. E., & Hinton, G. E. (1986). Learning represen-
learning theory (pp. 311–320). tations by back-propagating errors. Nature, 323, 533–536.
Bengio, Y., Courville, A., & Vincent, P. (2013). Repre- Salvucci, D. D., & Taatgen, N. A. (2008). Threaded cogni-
sentation learning: A review and new perspectives. IEEE tion: an integrated theory of concurrent multitasking. Psy-
Transactions on Pattern Analysis and Machine Intelli- chological Review, 115(1), 101.
gence, 35(8), 1798–1828. Saxe, A. M., McClelland, J. L., & Ganguli, S. (2013). Learn-
Botvinick, M. M., Braver, T. S., Barch, D. M., Carter, C. S., ing hierarchical category structure in deep neural networks.
& Cohen, J. D. (2001). Conflict monitoring and cognitive In Proceedings of the 35th annual meeting of the cognitive
control. Psychological Review, 108(3), 624. science society (pp. 1271–1276).
Caruana, R. (1997). Multitask learning. Machine learning, Saxe, A. M., McClelland, J. L., & Ganguli, S. (2014). Ex-
28(1), 41–75. act solutions to the nonlinear dynamics of learning in deep
Cohen, J. D., Dunbar, K., & McClelland, J. L. (1990). On the linear neural networks. In Y. Bengio & Y. LeCun (Eds.), In-
control of automatic processes: a parallel distributed pro- ternational conference on learning representations. Banff,
cessing account of the stroop effect. Psychological Review, Canada.
97(3), 332–361. Saxe, A. M., Musslick, S., & Cohen, J. D. (2017). A
Feng, S. F., Schwemmer, M., Gershman, S. J., & Cohen, J. D. formal tradeoff between learning speed and multitasking
(2014). Multitasking vs. multiplexing: toward a norma- ability in a simple neural network. http://www.people
tive account of limitations in the simultaneous execution of .fas.harvard.edu/˜asaxe/multitasking.html. (Re-
control-demanding behaviors. Cognitive, Affective, & Be- trieved May 13, 2017)
havioral Neuroscience, 14(1), 129-146. Seung, H., Sompolinsky, H., & Tishby, N. (1992). Statistical
Hinton, G. E. (1986). Learning distributed representations mechanics of learning from examples. Physical Review A,
of concepts. In Proceedings of the 8th confererence of 45(8), 6056.
the Cognitive Science Society (pp. 1–12). Hillsdale, NJ: Shenhav, A., Botvinick, M. M., & Cohen, J. D. (2013). The
Lawrence Erlbaum Associates. expected value of control: an integrative theory of anterior
Kurzban, R., Duckworth, A., Kable, J. W., & Myers, J. cingulate cortex function. Neuron, 79(2), 217–240.
(2013). An opportunity cost model of subjective effort Shiffrin, R. M., & Schneider, W. (1977). Controlled and auto-
and task performance. The Behavioral and Brain Sciences, matic human information processing: II. Perceptual learn-
36(6), 661–679. ing, automatic attending and a general theory. Psychologi-
McClelland, J. L., McNaughton, B. L., & O’Reilly, R. C. cal Review, 84(2), 127.
(1995). Why there are complementary learning systems

View publication stats