2017 Fourth International Conference on Image Information Processing (ICIIP)

A Neuromorphic Approach To Image Processing And

Machine Vision
Arvind Subramaniam
Birla Institute of Technology & Science (BITS) Pilani, Hyderabad Campus
Hyderabad, India
[email protected]

Abstract— Neuromorphic engineering is essentially the the foreground or relevant information present in the image.
development of artificial systems, such as electronic analog Previous studies have proposed the use of neuromorphic
circuits that employ information representations found in circuits [1], which can perform object-based selection and
biological nervous systems. Despite being faster and more segmentation. Recent studies have applied the concept of
accurate than the human brain, computers lag behind in synchronized oscillations, where photo-detector cells receiving
recognition capability. However, it is envisioned that the similar light intensities from an object oscillate together in a
advancement in neuromorphics, pertaining to the fields of synchronous manner [2], [3]. Visual tasks such as object-based
computer vision and image processing will provide a considerable selection and segmentation have been implemented with the
improvement in the way computers can interpret and analyze
help of specialized sensory processing functions called
information. In this paper, we explore the implementation of
visual tasks such as image segmentation, visual attention and
Neuromorphic Vision sensors. Several publications have
object recognition. Moreover, the concept of anisotropic diffusion appeared in recent years documenting the use of neuromorphic
has been examined followed by a novel approach employing vision sensors and pre-processors to integrate imaging, color
memristors to execute image segmentation. Additionally, we have segmentation and color-based object recognition [4]–[6]. These
discussed the role of neuromorphic vision sensors in artificial sensors consist of analogue electronic circuits operating in the
visual systems and the protocol involved in order to enable sub-threshold regime, interfaced to digital processing systems
asynchronous transmission of signals. Moreover, two widely that execute machine vision algorithms in order to realize
accepted algorithms that are used to emulate the process of selective attention, object recognition, etc. in artificial systems.
object recognition and visual attention have also been discussed. Recent developments in the field of neuroscience have led to a
Throughout the span of this paper, we have emphasized on the renewed interest in the development of computational models
employment of non-volatile memory devices such as memristors and algorithms for auditory and visual perceptions in the
to realize artificial visual systems. Finally, we discuss about human brain. Among these models, the Saliency based model
hardware accelerators and wish to represent a case in point for and the HMAX model have been extensively used for visual
arguing that progress in computer vision may benefit directly attention and object recognition respectively. The aim of this
from progress in non-volatile memory technology. paper is to provide a comprehensive exploration of some of the
recent techniques in the field of neuromorphics, which have
Keywords—Object recognition, Memristor, Image segmentation, been employed to implement visual tasks such as selective
Neuromorphic vision sensors, Address Event Representation, attention, gesture recognition, object recognition and tracking.
Random walker algorithm, Anisotropic Diffusion, Dynamic Vision
Sensor. The remainder of the paper is organized into the following
sections. Section II outlines a few techniques and approaches
I. INTRODUCTION adopted to execute scene segmentation. In Section III, we
The field of neuromorphics is a relatively young one examine the types and mechanisms of neuromorphic sensors.
considering the amount of active research that has been Section IV describes the models and algorithms used to better
invested into it. Despite decades of efforts in the field of implement the visual challenges such as object recognition and
artificial vision, no computer algorithm has been able to match selective attention in artificial systems. Furthermore, a modest
the performance and robustness of the primate visual system. introduction to hardware accelerators has also been presented.
Moreover, the proposed algorithms tend to be computationally Finally, Section V draws conclusions.
intensive, and their hardware implementations have
encountered problems such as high power consumption and II. IMAGE SEGMENTATION
unreliability. It has been discussed that neuromorphic systems
offer considerable advantages over conventional computing There are several popular image segmentation algorithms
systems in several areas and introduce the biologically realistic based on a variety of techniques such as thresholding, edge
possibility of implementing collective computation and detection, clustering, partial-differential equations (PDEs)
memory storage simultaneously. Autonomous visual and and artificial neural networks [7], [8].These methods can be
auditory processing systems can benefit greatly from the low- categorized into supervised, semi-supervised and
power features of the neuromorphic hardware technology. unsupervised. The Random walker algorithm is a semi-
An important processing stage in many visual processing supervised algorithm that models the image as a graph where
systems is image segmentation. It is the process of dividing an pixels correspond to nodes and are connected to neighboring
image into multiple parts, thereby making it easier to identify pixels through weighted edges. Recent studies have also

978-1-5090-6734-3/17/$31.00 ©2017 IEEE 68

 2017 Fourth International Conference on Image Information Processing (ICIIP)

employed unsupervised random walk approaches to execute manually change the resistance of the device connected across
image segmentation. The edges are weighted in accordance two nodes or pixels.
with the similarity between the corresponding pixels, and the
edge weights are equal to electrical conductance. III. NEUROMORPHIC OSCILLATOR NETWORKS
In the first- among other interpretations of the algorithm-the Previous research has documented the use of oscillatory
subject labels a small number of pixels commonly typified as correlation to segment and represent an image [13], [14]. In
seeds. Each unlabelled pixel is assumed to release a random such a scheme, different objects present in the image are
walker, after which the probability that a random walker will represented by different groups of synchronized oscillators. As
first arrive at each seed is calculated. After the probability for a result, no two distinct objects correspond to the same set of
every pixel is calculated, each pixel is assigned to the synchronized oscillations. Furthermore, the locally excitatory
respective seed which has the maximum probability of globally inhibitory oscillator network (LEGION) has been
receiving a random walker from the corresponding pixel. shown to provide a viable and effective framework to solve
Segmentation of the image is complete when all the unlabelled the problem of image segmentation [15], [16]. A major
nodes have each been assigned a seed. The random walker has drawback of the LEGION algorithms employed in previous
been shown to achieve segmentation in multiple ways such as research has been their sensitivity to noise. Executing the
solving a discrete Dirichlet problem [9] and employing an algorithm without any modifications would lead to the
iterative method to solve a discrete PDE [10]. The former problem of fragmentation since noisy images may result in a
approach, however, has proven to be computationally large number of fragments. However, this problem can be
intensive for real time applications. In an alternative approach, solved by suppressing the oscillators corresponding to the
the nodes corresponding to each pixel are assigned to a seed noisy regions of the image.
based on the effective conductance (edge weights) between Previous studies have indicated that memristors have
the nodes and the foreground or background seeds. After a few proven to be useful in designing non-linear oscillators [17]–
iterations, pixels with intensities higher than a certain [19] and contribute to the synchronization of coupled
threshold value are assigned to foreground seeds, else treated neuromorphic systems to perform visual segmentation [20]–
as background. A method which is slightly different from the [22]. However, there are a few challenges to the application of
aforementioned approaches involves viewing the graph of the memristors in designing oscillators. The resistance offered by
image as an electrical circuit and establishing an electrical the memristor is a non-linear function of time. However, this
potential at each node associated with the foreground. The might be resolved by using a memristor-transistor pair in the
magnitude of these potentials would be equal to the circuit.
probability that a random walker dropped at a node will reach
a foreground seed before reaching a background seed. IV. NEUROMORPHIC VISION SENSORS
However, in case of iterative algorithms, the conductance Neuromorphic vision sensors have been in the forefront of
between nodes/pixels must change appropriately after every the development of artificial visual systems, as they offer
iteration. This is not practically realizable in conventional visual perception at a lower computational load. Several
CMOS, since one would have to manually change the studies have discussed the emulation of the human retina with
resistance offered by the device connecting adjacent the help of analog electronic circuits [23]–[25].These circuits
nodes/pixels. are parallel and operate in the sub-threshold domain. As a
Yet another interesting approach to the Random walker result, two or more computational problems can be solved
algorithm involves the concept of anisotropic diffusion [11]. simultaneously, while consuming extremely low amounts of
Anisotropic diffusion is a method of denoising images by power. This permits the use of circuits with high
smoothing the pixel values only on one side of the boundary, computational density. Furthermore, the circuits are
as opposed to Isotropic diffusion, which averages pixel values asynchronous i.e. the transition from one state to another is the
across edges. The latter technique is applied in local averaging result of a change in their primary inputs. Despite issues
filters and median filters. The main drawback of the isotropic regarding instability and inefficiency, they are preferred over
diffusion is the loss of weak boundaries and the undesirable synchronous circuits due to the continuous nature of the
smoothing of all edges. There have also been modifications to sensory input. Moreover, the employment of synchronous
anisotropic diffusion such as the introduction of an edge systems would lead to misinterpretation of sensory
seeking diffusion coefficient (varies with image position) that information as a result of aliasing.
prohibits diffusion across strong edges [12]. The concept of neuromorphic sensors in not new. Since
A novel approach would be to employ memristor-based the 1990s, several studies have explored and widely proposed
crossbar arrays to represent edges between a node and a two types of neuromorphic sensors, namely Silicon retinas
foreground/background seed, as a memristor can change its [26], [27] and silicon cochleae [28], [29]. These systems make
effective conductance based on the history of current that had use of a protocol called Address Event Representation (AER)
previously flown through it. As a result, the variable resistance to transmit signals asynchronously. In AER, a variable number
of the memristor would allow for changes in the edge weights of lines (bus) are used to transmit data. This data is usually an
when there is a change in the potential difference across nodes address that has been assigned to each analog element present
after every iteration. Hence, this would eliminate the need to on the sending device [30]. The ACK and REQ lines are

69
 2017 Fourth International Conference on Image Information Processing (ICIIP)

active low lines which enable the synchronization of data V. NEUROMORPHIC VISION ALGORITHMS FOR ATTENTION AND
between the sender and the receiver. A major advantage of OBJECT RECOGNITION
using AER is that the power consumed for the transmission of
sensory information is significantly reduced, since AER A. Visual Attention
sensors transmit signals based on the activity of each The mammalian visual system has an inherent ability to
individual pixel, contrary to CMOS sensors. The activity of a detect the salient and important sections of an image and filter
pixel is represented by a stream of digital pulses produced by out redundant visual details through the selective attention
the neuromorphic sensor. mechanism. A number of models and algorithms of the human
Asynchronous devices that respond only to a change in the visual system have been proposed [38]. The focus of these
brightness of a pixel, such as the Dynamic Vision Sensor systems is to direct attention to regions of interest in an image.
(DVS) have been proposed, their key advantages being a As a result, further processing to sub-regions of the image is
significant reduction in data storage and computational restricted and the amount of data for complex processing tasks
complexity. Several studies have employed Spike-timing- such as object recognition is reduced. The locations to be
dependent plasticity (STDP) to extract correlated features analyzed are selected with the help of two guiding influences:
(temporally) from dynamic vision sensors [31], [32]. STDP is the goal-oriented, top-down attention and the image-driven,
a biologically-realistic learning mechanism based on the bottom-up attention [39].
relative timing of the post- and pre-synaptic spikes.

Vision
Sensors

Synchronous Asynchronous
or or
Frame-Based Event-Based
Figure 2

Top-down attention (TD), also known as sustained or

endogenous attention involves the voluntary allocation of
Silicon Silicon attention to certain features or regions in space for sustained
Retina Cochlea periods of time. As a result, it is slower and requires more
effort to engage than the bottom-up attention, in which the
resources are involuntarily captured [40], [41]. Top-down or
sustained attention would be utilized when the subject
intentionally directs attention to a task such as driving, etc.
Dynamic Asynchronous Alternatively, an unexpected event such as an accident in the
Time-based adjacent lane, might inadvertently capture the subject’s
Vision Image
Sensor Sensor attention, as it is significantly different from the surrounding
events. This involuntary allocation of attention is influenced
by the bottom-up attention (BU). Recent studies have
Figure 1. Types of vision sensors
documented two types of top-down attention (TD)
mechanisms, namely the volitional top-down selection process
Several studies have emphasized on the utilization of and the mandatory top-down selection process [42]. The
memristors as a modulating synapse between neurons, in order former mechanism is an attentional TD process that exerts
to realize Spike-timing-dependent plasticity in neuromorphic influence on sensory processing through acts of will, such as
systems [33]–[37], since the conductance of a memristor can wilfully shifting attention a particular region of space (top-
be modified by controlling the charge through it. The most down spatial attention) in the middle or to all black items (top-
recent silicon retina prototype is the Asynchronous time-based down feature attention) [43], [44].
Image Sensor (ATIS) which has the feature of conditional Alternatively, the mandatory TD process influences sensory
exposure measurement in addition to measuring a change in processing in an automatic and persistent manner. Fig.2
brightness. presents the mental impression of the stimulus, which can be
wilfully changed from a scene to a face and vice-versa,
demonstrating TD modulation of sensory processing in a
dynamic manner.

70
 2017 Fourth International Conference on Image Information Processing (ICIIP)

B. Object Recognition

Visual Object recognition is the ability of the mammalian

brain to perceive and identify an object's physical properties.
Therefore, it is a crucial feature of any artificial visual
processing system. Despite several attempts to practically
implement object recognition, it has proven to be an extremely
complex computational problem.
The computational load lies in the fact that each object can
project infinitely many dissimilar 2-D images as the size,
orientation and background of the object vary with respect to
the subject [49]. Among the various models proposed, the
Hierarchical Model and X (HMAX) model has been
extensively used for image classification with applications in
object and gesture recognition [50], [51].
The HMAX model is a four-level feed forward architecture
Figure 3
fundamentally made up of the following Simple (S) and
Complex (C) layers, which perform convolution and max-
On the other hand, in Fig.3, which presents the convex and
pooling operations to build complex features:
concave sides of a mask, the subject would be unable to
distinguish the concave side (inside) of the mask as it would
still seem to be convex. This inherent bias in identifying faces The first layer, commonly known as the S1 layer, is the result
as convex rather than concave is an illuminative example of of processing 12 two-dimensional Gabor filters of dimensions
mandatory TD processing. 11x11, with different orientations at every possible pixel.
Next, an image pyramid corresponding to each Gabor filter is
The Bottom-up attention (BU), also typified as transient or created. Each pyramid consists of a set of 12 scales, where
exogenous attention, is instrumental in the realization of each scale is smaller than the previous scale by a factor of 21/4.
involuntary attention in artificial visual systems. Central to the The 12 image pyramids (144 scales) are then fed into the
discipline of the bottom-up attention has been the saliency- second layer.
based model of visual attention [45]. The model employs
The second layer, referred to as the C1 layer, convolves the
feature maps to detect regions of interest in an image based on
spatial discontinuities in features such as orientation, color, image pyramids having the same orientation with a Max filter,
intensity, etc. and combines them into a unique topographical to create scale and position invariance over local regions i.e.,
saliency map. The saliency map is then scanned sequentially in the system is unchanged to changes in scale and position.
order of decreasing saliency. This breaks down the complexity Further, the C1 layer also subsamples the pyramids by a factor
of the image and enables real-time performance. Therefore, it of 5. As a result, C1 pyramids are smaller despite having the
be rapidly implemented as opposed to top-down attention, same number of orientations.
which requires successive operations on sensory input [46]. The third layer, commonly typified as the S2 layer, is one of
the most important and computationally intensive layers and
However, despite the real-time performance demonstrated accounts for over 80% of the total execution time. This arises
by the bottom-up influence, there is a need to combine both the
as a result of convolving every scale of each multi-scaled
influences in order to successfully execute tasks such as visual
pyramid from the C1 layer. In other words, the S1 layer
surveillance, where the detection of goal-specific targets as
well as unexpected visual events is critical. processes the results of the C1 layer by matching it against a
set of pre-learned feature prototypes. These feature prototypes
Certain studies have revealed the integration of both the are considerably more detailed than the previously mentioned
afore-mentioned guiding influences to optimize detection speed features detected by the S1 layer.
[47], [48]. First, the bottom-up influence determines the visual Finally, the C2 layer removes all scale and position
salience of each object in a number of dissimilar feature maps. information by pooling the obtained results across different
Next, the top-down guiding influence makes use of learnt scales and orientations to produce feature vectors.
knowledge of the aforementioned features of the object in Simply put, the objective of the aforementioned model is
order to enhance the relative weights of the bottom-up maps.
to obtain feature vectors from gray scale images, and classify
This maximizes the overall salience of the visual target with
those vectors with the help of a Support Vector Machine
respect to the surrounding. Therefore, the target is more salient
than other visual distractions and target detection speed is (SVM).
maximized. The classification of the afore-mentioned However, there have been a number of problems
neuromorphic vision approaches has been illustrated in Fig.4. accompanying the implementation of the HMAX model such
‘Both’ refers to the integration of the Top-down and Bottom-up as low energy efficiency and excessive power consumption.
guiding influences. Recent research has documented the use of hardware
accelerators, which are in the form of programmable cores or
programmable logic such as FPGA, ASIC and GPU, to

71
 2017 Fourth International Conference on Image Information Processing (ICIIP)

primarily speed up the S2/C2 stage as it is the most time- mechanisms and types of neuromorphic vision sensors have
consuming stage. It has been indicated that FPGA provides the also been explored. We have laid emphasis on the use of non-
best functional configurability while ASIC demonstrates the volatile memory devices such as memristors to enhance
highest efficiency. performance and energy efficiency during the execution of the
Recent studies have employed memristive Neuromorphic aforementioned tasks. The application of memristors in the
Computing Accelerators (NCA) consisting of memristor-based implementation of image segmentation, neuromorphic sensors
crossbar arrays, as opposed to conventional accelerators which and hardware accelerators has proven to be more energy
are based on systolic arrays. Memristor-based models of the efficient than previous methods, which used conventional
Boltzmann Machine, a massively parallel computational CMOS.
model used for solving combinatorial optimization problems, Throughout the span of the paper, we have tried to highlight
have been proposed to accelerate neural computation tasks in the importance of using memristors to better solve challenges
an energy efficient manner. The memristive hardware pertaining to image processing and computer vision.
accelerator had an appreciable improvement of 6.89x and 5.2x In addition to the employment of memristors to overcome
on performance and power consumption respectively, as visual challenges, future research directions should focus on
compared to a standard RRAM based memory [52]. However, the application of memristive systems to solve issues related to
the proposed memristor-based accelerator has not been shown upcoming fields such as the Internet of Things (IoT) and
to solve problems of higher computational complexity yet. quantum computing.

REFERENCES

[1] G. Crebbin, Image Segmentation using Neuromorphic Integrate-and-Fire

VISION Cells, in: Information, Commun. Signal Process. 2005 Fifth Int. Conf.,
ALGORITHMS 2005: pp. 305–309.
[2] M. G. Quiles, D. Wang, L. Zhao, R. A. F. Romero, and D. S. Huang,
“Selecting salient objects in real scenes: An oscillatory correlation
model,” Neural Networks, vol. 24, no. 1, pp. 54–64, 2011.
Visual Object [3] M. G. Quiles, D. L. Wang, L. Zhao, R. A. F. Romero, and D. S. Huang,
Attention Recognition “An oscillatory correlation model of object-based attention,” in
Proceedings of the International Joint Conference on Neural Networks,
2009, pp. 2596–2602.
[4] T. Serre, T. Poggio, A neuromorphic approach to computer vision,
Commun. ACM. 53 (2010) 54. doi:10.1145/1831407.1831425.
Bottom- Both Top- HMAX [5] A. Vanarse, A. Osseiran, A. Rassau, A review of current neuromorphic
Up Down approaches for vision, auditory, and olfactory sensors, Front. Neurosci.
10 (2016). doi:10.3389/fnins.2016.00115.
[6] A. Mishra, R. Ghosh, A. Goyal, N. V. Thakor, S.L. Kukreja, Real-Time
robot tracking and following with neuromorphic vision sensor, in: Proc.
IEEE RAS EMBS Int. Conf. Biomed. Robot. Biomechatronics, 2016:
Saliency- pp. 13–18. doi:10.1109/BIOROB.2016.7523451.
based
Mandatory Volitional [7] C. Li and Y. Li, “Fast and robust image segmentation by small-world
model of
neural oscillator networks,” Cogn. Neurodyn., vol. 5, no. 2, pp. 209–220,
visual 2011.
attention
[8] N. M. Zaitoun and M. J. Aqel, “Survey on Image Segmentation
Techniques,” in Procedia Computer Science, 2015, vol. 65, pp. 797–
806.
[9] D. P. Onoma, S. Ruan, I. Gardin, G. A. Monnehan, R. Modzelewski, and
Figure 4. Classification of Neuromorphic vision algorithms P. Vera, “3D random walk based segmentation for lung tumor
delineation in PET imaging,” in Proceedings - International Symposium
on Biomedical Imaging, 2012, pp. 1260–1263.
VI. CONCLUSION [10] T. Pätz and T. Preusser, “Segmentation of stochastic images with a
stochastic random walker method,” IEEE Trans. Image Process., vol.
21, no. 5, pp. 2424–2433, 2012.
This paper focuses on the techniques and algorithms
[11] P. Perona, Anisotropic Diffusion Processes in Early Vision, in: Proc.
employed to practically implement critical visual tasks such as 6^{\;th} Multidimens. Signal Process. Work., 1989: p. 68.
image segmentation, selective attention and object doi:“10.1109/MDSP.1989.97028.”
recognition, by making use techniques pertaining to the [12] P. Perona, J. Malik, Scale-space and edge detection using anisotropic
emerging field of neuromorphics. We have also recommended diffusion, IEEE Trans. Pattern Anal. Mach. Intell. 12 (1990) 629–639.
a novel approach which involves utilizing memristor-based doi:10.1109/34.56205.
crossbar arrays to act as edges between a node and a [13] K. Chen, D.L. Wang, X. Liu, Weight adaptation and oscillatory
correlation for image segmentation, IEEE Trans. Neural Networks. 11
foreground/background seed and hope to instigate new (2000) 1106–1123. doi:10.1109/72.870043.
direction for further research integrating neuromorphics, [14] K. Chen, D.L. Wang, A dynamically coupled neural oscillator network
image processing and non-volatile memory devices. Apart for image segmentation, Neural Networks. 15 (2002) 423–439.
from the realization of the aforementioned tasks, the doi:10.1016/S0893-6080(02)00028-X.

72
 2017 Fourth International Conference on Image Information Processing (ICIIP)

[15] J. Yuan, D. Wang, B. Wu, L. Yan, and R. Li, “LEGION-based [36] A. Afifi, A. Ayatollahi, F. Raissi, STDP implementation using
automatic road extraction from satellite imagery,” IEEE Trans. Geosci. memristive nanodevice in CMOS-Nano neuromorphic networks, IEICE
Remote Sens., vol. 49, no. 11 PART 2, pp. 4528–4538, 2011. Electron. Express. 6 (2009) 148–153. doi:10.1587/elex.6.148.
[16] N. Shareef, D. L. Wang, and R. Yagel, “Segmentation of medical [37] [44] A. Pantazi, S. Woźniak, T. Tuma, E. Eleftheriou, All-memristive
images using LEGION.,” IEEE Trans. Med. Imaging, vol. 18, no. 1, pp. neuromorphic computing with level-tuned neurons, Nanotechnology. 27
74–91, 1999. (2016) 355205. doi:10.1088/0957-4484/27/35/355205.
[17] M. Itoh, L.O. Chua, Memristor Oscillators, Int. J. Bifurc. Chaos. 18 [38] V. Mnih, N. Heess, A. Graves, K. Kavukcuoglu, Recurrent Models of
(2008) 3183–3206. doi:10.1142/S0218127408022354. Visual Attention, Nips-2014. (2014) 1–9. doi:ng.
[18] F. Corinto, A. Ascoli, M. Gilli, Nonlinear dynamics of memristor [39] O. Le Meur, P. Le Callet, D. Barba, D. Thoreau, A coherent
oscillators, IEEE Trans. Circuits Syst. I Regul. Pap. 58 (2011) 1323– computational approach to model bottom-up visual attention, IEEE
1336. doi:10.1109/TCSI.2010.2097731. Trans. Pattern Anal. Mach. Intell. 28 (2006) 802–817.
[19] A. Talukdar, A.G. Radwan, K.N. Salama, Non linear dynamics of doi:10.1109/TPAMI.2006.86.
memristor based 3rd order oscillatory system, Microelectronics J. 43 [40] T. Liu, S.T. Stevens, M. Carrasco, Comparing the time course and
(2012) 169–175. doi:10.1016/j.mejo.2011.12.012. efficacy of spatial and feature-based attention, Vision Res. 47 (2007)
[20] C.K. Volos, I.M. Kyprianidis, I.N. Stouboulos, E. Tlelo-Cuautle, S. 108–113. doi:10.1016/j.visres.2006.09.017.
Vaidyanathan, Memristor: A new concept in synchronization of coupled [41] S. Ling, M. Carrasco, Sustained and transient covert attention enhance
neuromorphic circuits, J. Eng. Sci. Technol. Rev. 8 (2015) 157–173. the signal via different contrast response functions, Vision Res. 46
[21] W. Sun, C. Li, J. Yu, C. Berkeley, S. Williams, A memristor based (2006) 1210–1220. doi:10.1016/j.visres.2005.05.008.
chaotic oscillator, 2009 Int. Conf. Commun. Circuits Syst. (2009) 955– [42] F. Baluch, L. Itti, Mechanisms of top-down attention, Trends Neurosci.
957. doi:10.1109/ICCCAS.2009.5250351. 34 (2011) 210–224. doi:10.1016/j.tins.2011.02.003.
[22] J. Cosp, J. Madrenas, Scene Segmentation Using Neuromorphic [43] S.L. Bressler, W. Tang, C.M. Sylvester, G.L. Shulman, M. Corbetta,
Oscillatory Networks, IEEE Trans. Neural Networks. 14 (2003) 1278– Top-down control of human visual cortex by frontal and parietal cortex
1296. doi:10.1109/TNN.2003.816364. in anticipatory visual spatial attention., J Neurosci. 28 (2008) 10056–61.
[23] C. Tan, S. Lallee, G. Orchard, Benchmarking neuromorphic vision: doi:10.1523/JNEUROSCI.1776-08.2008.
Lessons learnt from computer vision, Front. Neurosci. 9 (2015). [44] B. Giesbrecht, M.G. Woldorff, A.W. Song, G.R. Mangun, Neural
doi:10.3389/fnins.2015.00374. mechanisms of top-down control during spatial and feature attention,
[24] S.C. Liu, T. Delbruck, G. Indiveri, A. Whatley, R. Douglas, Event-based Neuroimage. 19 (2003) 496–512. doi:10.1016/S1053-8119(03)00162-9.
neuromorphic systems, 2014. doi:10.1002/9781118927601. [45] N. Ouerhani, R. von Wartburg, H. Hugli, Empirical validation of the
[25] S.C. Liu, T. Delbruck, Neuromorphic sensory systems, Curr. Opin. saliency-based model of visual attention, Electron. Lett. Comput. Vis.
Neurobiol. 20 (2010) 288–295. doi:10.1016/j.conb.2010.03.007. Image Anal. 3 (2004) 13–24.
[26] C. Posch, T. Serrano-Gotarredona, B. Linares-Barranco, and T. [46] C.E. Connor, H.E. Egeth, S. Yantis, Visual attention: Bottom-up versus
Delbruck, “Retinomorphic event-based vision sensors: Bioinspired top-down, Curr. Biol. 14 (2004). doi:10.1016/j.cub.2004.09.041.
cameras with spiking output,” Proc. IEEE, vol. 102, no. 10, pp. 1470– [47] V. Navalpakkam, L. Itti, An integrated model of top-down and bottom-
1484, 2014. up attention for optimizing detection speed, in: Proc. IEEE Comput. Soc.
[27] T. Delbrück, B. Linares-Barranco, E. Culurciello, and C. Posch, Conf. Comput. Vis. Pattern Recognit., 2006: pp. 2049–2056.
“Activity-driven, event-based vision sensors,” in ISCAS 2010 - 2010 doi:10.1109/CVPR.2006.54.
IEEE International Symposium on Circuits and Systems: Nano-Bio [48] T. Ogawa, H. Komatsu, Target selection in area V4 during a
Circuit Fabrics and Systems, 2010, pp. 2426–2429. multidimensional visual search task, J. Neurosci. 24 (2004) 6371–6382.
[28] M. Yang, C. H. Chien, T. Delbruck, and S. C. Liu, “A 0.5V 55μW doi:10.1523/jneurosci.0569-04.2004.
64×2-channel binaural silicon cochlea for event-driven stereo-audio [49] N. Pinto, D.D. Cox, J.J. DiCarlo, Why is real-world visual object
sensing,” in Digest of Technical Papers - IEEE International Solid-State recognition hard?, PLoS Comput. Biol. 4 (2008) 0151–0156.
Circuits Conference, 2016, vol. 59, pp. 388–389. doi:10.1371/journal.pcbi.0040027.
[29] T. J. Hamilton, C. Jin, A. van Schaik, and J. Tapson, “An active 2-D [50] W. Gu, C. Xiang, H. Lin, Modified HMAX models for facial expression
silicon cochlea,” IEEE Trans. Biomed. Circuits Syst., vol. 2, no. 1, pp. recognition, in: 2009 IEEE Int. Conf. Control Autom. ICCA 2009, 2009:
30–43, 2008. pp. 1509–1514. doi:10.1109/ICCA.2009.5410482.
[30] G. Indiveri, Neuromorphic engineering, in: Springer Handb. Comput. [51] C. Liu, F. Sun, HMAX model: A survey, in: Proc. Int. Jt. Conf. Neural
Intell., 2015: pp. 715–725. doi:10.1007/978-3-662-43505-2_38. Networks, 2015. doi:10.1109/IJCNN.2015.7280677.
[31] O. Bichler, D. Querlioz, S.J. Thorpe, J.P. Bourgoin, C. Gamrat, [52] M. N. Bojnordi and E. Ipek, “Memristive Boltzmann machine: A
Extraction of temporally correlated features from dynamic vision hardware accelerator for combinatorial optimization and deep learning,”
sensors with spike-timing-dependent plasticity, Neural Networks. 32 in Proceedings - International Symposium on High-Performance
(2012) 339–348. doi:10.1016/j.neunet.2012.02.022. Computer Architecture, 2016, vol. 2016–April, pp. 1–13.
[32] O. Bichler, D. Querlioz, S.J. Thorpe, J.P. Bourgoin, C. Gamrat,
Unsupervised features extraction from asynchronous silicon retina
through spike-timing-dependent plasticity, in: Proc. Int. Jt. Conf. Neural
Networks, 2011: pp. 859–866. doi:10.1109/IJCNN.2011.6033311.
[33] W. Chan, J. Lohn, Spike timing dependent plasticity with memristive
synapse in neuromorphic systems, in: Proc. Int. Jt. Conf. Neural
Networks, 2012. doi:10.1109/IJCNN.2012.6252822.
[34] G.S. Snider, Spike-timing-dependent learning in memristive
nanodevices, in: 2008 IEEE/ACM Int. Symp. Nanoscale Archit.
NANOARCH 2008, 2008: pp. 85–92.
doi:10.1109/NANOARCH.2008.4585796.
[35] C. Zamarreño-Ramos, L.A. Camuñas-Mesa, J.A. Perez-Carrasco, T.
Masquelier, T. Serrano-Gotarredona, B. Linares-Barranco, On spike-
timing-dependent-plasticity, memristive devices, and building a self-
learning visual cortex, Front. Neurosci. (2011).
doi:10.3389/fnins.2011.00026.

73