Compact Early Vision Signal Analyzers in Neuromorphic Technology

Valentina Baruzzi

, Giacomo Indiveri

2 a

and Silvio P. Sabatini

1 b

Department of Informatics, Bioengineering, Robotics and Systems Engineering, University of Genoa, Genoa, Italy

Institute of Neuroinformatics, University of Z

urich and ETH Z

urich, Z

urich, Switzerland

Keywords:

Early Vision, Gabor Filters, Receptive Fields, Neuromorphic Engineering, Event-based Sensors, Bioinspired

Vision, Harmonic Representations.

Abstract:

Reproducing the dynamics of biological neural systems using mixed signal analog/digital neuromorphic cir-

cuits makes these systems ideal platforms to implement low-power bio-inspired devices for a wide range of

application domains. Despite these principled assets, neuromorphic system design has to cope with the limited

resources presently available on hardware. Here, different spiking networks were designed, tested in simula-

tion, and implemented on the neuromorphic processor DYNAP-SE, to obtain silicon neurons that are tuned to

visual stimuli oriented at speciﬁc angles and with speciﬁc spatial frequencies, provided by the event camera

DVS. Recurrent clustered inhibition was successfully tested on spiking neural networks, both in simulation

and on the DYNAP-SE board, to obtain neurons with highly structured Gabor-like receptive ﬁelds (RFs);

these neurons are characterized by tuning curves that are sharper or at least comparable to the ones obtained

using equivalent feed-forward schemes, but require a signiﬁcantly lower number of synapses. The resulting

harmonic signal description provided by the proposed neuromorphic circuit could be potentially used for a

complete characterization of the 2D local structure of the visual signal in terms of phase relationships from all

the available oriented channels.

1 INTRODUCTION

The goal of early vision is to extract as much in-

formation as possible about the structural properties

of the visual signal. Such a process must be ef-

ﬁcient, providing reliable features of high informa-

tive content in short time, in order not to compro-

mise the efﬁcacy of subsequent processing stages, and

without being, at the same time, an unbearable bot-

tleneck. Recent asynchronous event-driven cameras

combined with brain-inspired spiking neuromorphic

processors can be a solution. They, indeed, constitute

reconﬁgurable systems of silicon neurons that operate

on mixed (analog/digital) signal to achieve sophisti-

cated real-time visual processing. Due to the imple-

mentation limit for connectivity patterns between sili-

con neurons, vision front-ends typically restrict them-

selves to be image edge and moving object detectors.

Sometimes, neurons operate exclusively on temporal

contrast events, disregarding the spatial contrast, that

is the local spatial structure of the visual signal, ob-

tained by integrating visual events over spatial neigh-

https://orcid.org/0000-0002-7109-1689

https://orcid.org/0000-0002-0557-7306

borhoods (e.g., (Osswald et al., 2017)). Sometimes,

they implement simple (e.g., binary) feature match-

ing, by composing local receptors outputs through re-

ceptive ﬁelds (RFs) with minimal and simple weight-

ing proﬁles (e.g., (M

uggler et al., 2017) ). Speciﬁc

matched operators extract informative (symbolic) el-

ements of an image, such as points and lines, but in-

escapably discard part of the signal. A more sophis-

ticated visual signal analysis would require highly

structured RFs, e.g., with 2D wavelet-like proﬁles

to extract local amplitude, phase, and orientation in-

formation in a given frequency sub-band (cf. linear

visual cortical cell responses, see e.g. (Jones and

Palmer, 1987)). In general, for many image process-

ing tasks, it is commonly used to represent an image

by oriented spatial-frequency channels in which some

properties of the image are better represented than in

image space. The spatio-temporal properties of the

resulting harmonic components have been proved to

be critically important for extracting primary early vi-

sion information. Indeed, in general, as evidenced

in several studies (e.g., see (Fleet and Jepson, 1993)

(Ogale and Aloimonos, 2007) (Sabatini et al., 2010)),

by using harmonic patterns for matching instead of

530

Baruzzi, V., Indiver i, G. and Sabatini, S.

Compact Early Vision Signal Analyzers in Neuromorphic Technology.

DOI: 10.5220/0009171205300537

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 4: VISAPP, pages

530-537

ISBN: 978-989-758-402-2; ISSN: 2184-4321

image luminance measures, the resulting perception

is more reliable (i.e., stable), denser and immune to

lighting conditions.

Since a direct implementation of such wavelet RFs

on neuromorphic hardware is hampered by the lim-

ited wiring capability between silicon neurons, de-

signing and validating efﬁcient architectural solutions

to obtain compact visual signal analyzers with min-

imal resource consumption becomes a key issue. In

this paper, we propose an economic way to imple-

ment spike-based early-vision detectors of oriented

features in given spatial frequency bandwidths that

mimic the known properties of Gabor-like simple

cells RFs in the primary visual cortex (V1) (Jones

and Palmer, 1987). In particular, we developed a re-

current neural network architecture based on a model

of the retinocortical visual pathway to obtain neurons

highly tuned to oriented visual stimuli along a spe-

ciﬁc direction and with a speciﬁc spatial frequency,

with Gabor-like RFs. The computation performed by

the biological retina is emulated by an asynchronous

event-driven Dynamic Vision Sensor (DVS) (Licht-

steiner et al., 2008), which only indicates luminance

temporal changes in the image impinging on the pho-

todiode array. Its output feeds a neuromorphic pro-

cessor (DYNAP-SE) (Moradi et al., 2018) with re-

conﬁgurable silicon neurons that comprises adaptive

integrate-and ﬁre neurons and dynamic synapses. We

show how the network implemented on this device

gives rise to neurons tuned to speciﬁc orientations

and spatial frequencies, independent of the temporal

frequency of the visual stimulus. Compared to al-

ternative feed-forward schemes, the model proposed

produces highly structured RFs of any phase symme-

try with a limited number of synaptic connections,

thus optimizing hardware resources. We validate the

model and approach with experimental results using

both synthetic and natural images.

2 MATERIALS AND METHODS

The system setup (see Fig. 1), consists of a DVS,

representing the retinal stage, whose output is fed to

a DYNAP-SE board, representing the cortical stage.

The sensor is composed of pixels that respond asyn-

chronously to relative changes in light intensity, gen-

erating a stream of ‘events’ that constitute its out-

put. Each event encodes the timestamp at which it

occurred, the position of the pixel that detects it, and

its polarity (ON and OFF for positive and negative

intensity changes, respectively). Sensor activity is

reproduced by a population of spiking ‘virtual neu-

rons’, implemented digitally by a FPGA module on

the DYNAP-SE board, which acts as spike genera-

tor for the physical silicon neurons on the chip. One

of the chips on the DYNAP-SE board is programmed

so to have a speciﬁc conﬁguration of synaptic con-

nections, described in the following, that gives rise to

silicon neurons with well-structured Gabor-like RFs.

These neurons are thus tuned on a speciﬁc orienta-

tion and spatial frequency, reproducing the computa-

tion carried out by simple cells in the primary visual

cortex. The ‘membrane voltage’ of selected silicon

neurons can be observed using an oscilloscope and the

spiking activity of the entire chip can be monitored in

real time on a computer through the CTXCTL Primer

interface (aiCTX, 2018).

2.1 Network Scheme

In order to obtain neurons with Gabor-like RFs that

act as early vision feature extractors, while deal-

ing with the limited number of synaptic connec-

tions available on the chip, we adopted the recurrent

clustered inhibition approach, described in (Sabatini,

1996). Ideally, the network is thus composed of a ﬁrst

population of neurons representing the retinal stage

(’retina layer’) and a second population representing

the cortical stage (’V1 layer’). Each neuron of the

V1 layer receives excitatory afference from a group

of neurons in an elongated region on the retina layer

(feed-forward kernel) and inhibitory afference from

two groups of neurons of the same layer (recurrent

or feed-back kernel), in regions that are displaced

orthogonally to the major axis of the feed-forward

kernel (see left side of Fig. 1). For both feed-

forward and recurrent contributions we adopt Gaus-

sian weighting proﬁles. Each neuron with such a

connectivity scheme will have an induced RF with a

spatial proﬁle described by a two-dimensional Gabor

function, as a result of the combination of the feed-

forward excitatory and the recurrent inhibitory con-

tributions.

2.2 Simulated Network

To test the efﬁcacy of the recurrent clustered inhibi-

tion approach for the the spiking neuron implementa-

tion, the network was ﬁrst simulated by using Brian2

(Stimberg et al., 2019) and its toolbox teili (Milde

et al., 2018). This simulator for spiking neural net-

works implements the model of the physical silicon

neurons on the DYNAP-SE board. The simulated net-

work has the structure described above, and discrete

nature.

The use of sinusoidal gratings as visual input allowed

us to extract the spatial frequency and orientation tun-

Compact Early Vision Signal Analyzers in Neuromorphic Technology

531

Figure 1: General structure of the network and connectivity scheme. (Left) The target neuron, indicated as a black bar on the

V1 layer, receives feed-forward excitation from neurons of the retina layer within an elongated region, and recurrent inhibition

from V1 neurons located in two circular clusters displaced symmetrically along the orientation selectivity bias provided by

feed-forward afferent connections. The direction of the bar indicates the orientation to which the neuron will be eventually

sensitive according to such connectivity scheme. The same pattern of connections is repeated for every neuron of the V1

layer. (Right) The overall system setup detailing how the model network has been physically mapped on the DYNAP-SE

board. The DVS output is reproduced by a population of spiking ‘virtual neurons’ that acts as spike generator for the physical

silicon neurons on the DYNAP-SE chips. The board can be connected to an oscilloscope, to observe the ‘membrane voltage’

of selected silicon neurons, and to a computer, through the CTXCTL Primer interface, to monitor the spiking activity of the

four chips in real time. For more details see text.

ing curves of the neurons of interest in the V1 layer,

offering a way to verify the efﬁcacy of the recurrent

clustered inhibition in molding the neurons’ tuning on

speciﬁc values of these features. Moreover, the sim-

ulation process allowed us to verify the assumption

of linearity of the network’s behaviour and to tune

the signiﬁcant parameters of the connectivity scheme,

i.e. distance between the recurrent Gaussian clusters,

size of the clusters and strength of the inhibition to

obtain the narrowest tuning curves. For the sake of

simplicity, the size of the feed-forward kernel and the

strength of the excitation are kept constant.

This process leads to RFs for the neurons of the V1

layer that can be considered feature extractors for ori-

entation and spatial frequency. The preferred orienta-

tion selectivity can be controlled by rotating the con-

nectivity scheme, while the peak spatial frequency

can be varied by scaling the displacement of the in-

hibitory kernels with respect to their size. However,

due to the network structure through which they were

created, they all have an even symmetry, i.e. they

all exhibit the same (zero) phase. To overcome this

limitation and obtain feature extractors with arbitrary

phase values, we adopted a method similar to the

one described in (Raffo et al., 1998). Considering

the one-dimensional projections of the spatial pro-

ﬁles of the RFs, obtained as the Fourier transform

of their spatial frequency tuning curves, they can be

summed in a convenient way to obtain a proﬁle with

the desired phase value. In particular, we considered

3 neurons of the V1 layer, the central one in posi-

tion n and the lateral ones at the positions n − k and

n + k, where k is chosen to be approximately the dis-

tance d, in terms of neuron index, between a target

neuron of the V1 layer and the centers of the Gaus-

sian clusters from which it receives recursive inhibi-

tion. This way, the maximum of the spatial proﬁles of

the RFs of the lateral neurons will lay in correspon-

dence with the minima of the spatial proﬁle of the RF

of the central neuron. By weighting the sum of the

three proﬁles with α = − sinψ− 0.5· cos ψ, β = cos ψ,

γ = sinψ − 0.5 · cos ψ we can obtain a spatial proﬁle

with any desired phase value (ψ).

2.3 Network Implementation on the

DYNAP-SE Board

Based on the results obtained through simulations, we

implemented the network on the DYNAP-SE board.

It is worth noting that, to overcome the restrictions

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

532

posed by the chip, the network structure had to be

modiﬁed. In particular, since every neuron can have

at most 64 afferent connections, it was necessary an

extra layer of neurons, the relay layer, between the

retina layer and the V1 layer, to increase the number

of available synapses. The relay layer receives ex-

citation from the feed-forward kernels on the retina

layer and projects one-to-one connections to the V1

layer, where the inhibitory recursion is then applied.

The weights were adjusted so that that the network

with this new structure behaves in an equivalent way

to the simulated one. Moreover, each chip is di-

vided into four cores, and all the neurons belong-

ing to a core share the same biases, including the

synaptic weights of the afferent connections. Also,

only two types of excitatory synapses and one type

of inhibitory synapses were available for each neu-

ron. Due to these constraints, the weights of the con-

nections that form the kernels could not be assigned

by sampling Gaussian functions. Instead, we adopted

a constant value for the weights of the recurrent in-

hibitory connections, and two values for feed-forward

excitatory connections (higher at the center and lower

at the periphery of the kernel).

The DYNAP-SE board was interfaced with the com-

puter through the software CTXCTL Primer, which

provides a Python console that can be used to send

commands to deﬁne the connectivity of the silicon

neurons, to set biases to speciﬁc values, to send in-

puts to the neurons, and to record their activity.

2.4 Visual Stimuli Dataset

We choose as input sinusoidal gratings, widely used

in the literature to investigate the response of cells

in the primary visual cortex. Gratings with different

orientations and spatial frequencies allow us to ob-

tain the tuning curves that characterize a neuron’s be-

havior. Speciﬁcally, the open source platform for be-

havioral science experiments PsychoPy (Peirce et al.,

2019), was used to generate moving sinusoidal grat-

ings with speciﬁc orientation, measured in degrees,

spatial frequency, measured in cycles/deg, and tem-

poral frequency, measured in Hertz. The moving grat-

ings were displayed on a screen and recorded by the

DVS. jAER (Delbruck, 2007), an open-source Java-

based framework, was used to set the camera biases

and record the data. Figure 2 shows how the sinu-

soidal grating is perceived by the DVS sensor. Since

the DVS is sensitive to local contrast changes, bands

of ON and OFF events are generated where the si-

nusoidal proﬁle is steep enough. Conversely, where

the proﬁle is almost ﬂat, contrast differences are too

small to be detected by the sensor and no events are

generated, resulting in bands without events, which

are wider or narrower according to the contrast sensi-

tivity threshold. The spatial frequency information is

preserved anyway, encoded in the period of the bands

of events, not in their width; it is worth noting that the

phase is shifted by 90

◦

For a given distance D between the display and the

DVS, the angular and the linear ﬁeld of views of the

sensor, AFOV and LFOV, respectively, can be ex-

pressed as a function of the focal length f of the lens

and the size W of the pixel array:

AFOV = 2 arctan



2 f



LFOV =

For the DVS128, W is 128 pixels × 40 µm/pixel =

5.12 mm.

Figure 2: How the sinusoidal grating is converted in events

by the DVS. Upward arrows represent ON events and down-

ward arrows represent OFF events. Beige stripes are regions

without events.

3 RESULTS

3.1 Linearity Test and Feature Tuning

Characterization

First, we tested the linearity assumption. If the net-

work behaves in a linear way, the ﬁring rate of the

neuron of interest should be modulated by the same

temporal frequency of the grating used as input. Fig-

ure 3 shows that this condition is fulﬁlled for a wide

range of temporal and spatial frequencies of the stim-

ulus, both when the inhibitory recursion is turned off

and when it is applied.

The simulations showed that recurrent clustered in-

hibition does indeed elicit the tuning of the neurons

in the V1 layer on speciﬁc values of orientation and

spatial frequency. The best results, i.e., the narrowest

tuning curves, are obtained when the size of the re-

current inhibitory clusters and their distance from the

target neuron are both comparable to the width of the

feed-forward excitatory kernel.

Moreover, using recurrent clustered inhibition gives

rise to well-structured Gabor-like RFs with a substan-

tially lower number of synaptic connections than that

Compact Early Vision Signal Analyzers in Neuromorphic Technology

533

Figure 3: Instantaneous ﬁring rate oscillations during 1s simulation. Black and red curves refer to the central neuron of V1

layer and the blue curve to four sample corresponding afferent neurons of the retina layer. In the ﬁrst row the response to

moving gratings with spatial frequency of 0.26 c/deg and temporal frequencies that vary as indicated; in the second row the

response to moving gratings with temporal frequency of 3.16 Hz and spatial frequencies that are indicated.

Figure 4: (Left) The spatial frequency tuning curves and their Fourier transforms for three sample neurons of the V1 layer

(central neuron and two nearby neurons at distance d from it). The input grating used to obtain the curves had a temporal

frequency of 3.16 Hz. (Right) The weighted sum of the spatial proﬁles to obtain a ﬁlter with an arbitrary phase value. The

value of k can be chosen as being approximately equal to d.

required when using an exclusively feed-forward ap-

proach. To put this claim in numbers, we can con-

sider a network in which both the feed-forward ker-

nel and the recurrent clusters are 5 neurons wide. In

this case, each neuron of the V1 layer receives 101

afferences. We can compare this result with equiva-

lent networks in which the spatial proﬁle of a V1 neu-

ron’s RF is obtained through exclusively feed-forward

excitatory and inhibitory connections from the retina

layer, deﬁned by sampling a Gabor function whose

central lobe is 5 neurons wide. To obtain RFs with

3 and 5 subregions, 127 and 241 afferences to each

neuron of the V1 layer are required, respectively. Be-

sides, the orientation tuning curves of such neurons

are comparable to the ones of the neurons of the recur-

rent network and the spatial frequency tuning curves

are slightly worse.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

534

Figure 5: Filters with different phase values resulting from the weighted sum of responses of nearby neurons.

3.2 Spatial Proﬁles with Arbitrary

Phase Values

As explained before, by linear weighting of neighbor-

ing RFs we can obtain higher-order RF proﬁles of any

phase. For the sake of clarity, results are shown for

the central 1D cross section of a neuron’s RF. The

left side Fig. 4 shows the spatial frequency tuning

curves and their Fourier transforms for three sample

neurons, placed in central region of the V1 layer. We

observe that the neurons are tuned to a speciﬁc spatial

frequency value, and that the ﬁlters’ spatial proﬁles

present a central positive lobe and two negative side-

bands, resembling a Gabor function with even sym-

metry.

The right side of Fig. 4 shows how the weighted sum

of the three RFs yields a resulting proﬁle with a dif-

ferent phase value. Figure 5 showcases a series of re-

sulting spatial proﬁles obtained with this approach for

phase values ranging from 0

◦

(corresponding to even

symmetry) to 90

◦

(corresponding to odd symmetry).

Figure 6: Examples of orientation tuning curves for differ-

ent values of θ. The input grating had a temporal frequency

of 3.16 Hz.

3.3 Functional Validation

The resulting bank of linear ﬁlters can be used as a

minimal and controllable set of operators for extract-

ing early vision features, directly from the spiking

video stream provided by the DVS. Indeed, the spa-

tial structure of the Gabor-like proﬁles allows us to

aggregate ON and OFF temporal events according to

local oriented band-pass spatial frequency channels,

which are frequently used as front-ends of artiﬁcial

vision systems (Dollr et al., 2014) (Luan et al., 2018).

Although several tricks should be considered to efﬁ-

ciently implement a full multichannel representation,

a ﬂavour of the functionality of the proposed network

is presented, for a single scale and four orientation

channels. Figure 8 shows the results for a snapshot

of a DVS recording featuring a moving hand; the pan-

els show the activity of the DVS, the activity of the

retina layer that reproduces ON events as spikes and

the activity of the V1 layer for 4 different values of θ.

Brighter tones indicate higher ﬁring rates. Only the

neurons of the V1 layer whose orientation sensitivity

Figure 7: Spatial frequency tuning curves and their Fourier

transforms for different temporal frequencies of the input

grating.

Compact Early Vision Signal Analyzers in Neuromorphic Technology

535

Figure 8: A snapshot of the activity of the DVS, the retina layer and the V1 layer with different preferred orientations for a

real scene. The stream of events generated by the DVS is shown in the input image: ’ON’ and ’OFF’ events are represented

as white and black squares overlaid to the corresponding scene acquired with a regular frame-based camera. Brighter tones

indicate higher ﬁring rates. For the sake of simplicity in the visualization, only the neurons of the V1 layer whose ﬁring

rates were over a certain threshold were represented. Cropped DVS240 recording from DVSFLOW16 dataset (Rueckauer and

Delbruck, 2016).

matches the dominant orientation present in the image

show a high ﬁring rate.

3.4 Network Implementation on Chip

The DYNAP-SE board was programmed to imple-

ment the modiﬁed network structure we described

previously. We used one of the 4 chips on the

board, composed of 4 cores with 256 silicon neurons

each. The chip is fabricated using a standard 0.18

um CMOS technology and the neurons and synapses

biophysical behaviors are mimicked through parallel

analog circuits (rather than time-multiplexed digital

ones) which operate in sub-threshold regime to min-

imize the dynamic power consumption and to repro-

duce biologically plausible temporal dynamics. The

computation is asynchronous and the memory ele-

ments (such as capacitors, CAM, and SRAM cells)

are distributed across the computing modules, mak-

ing the chip non Von Neumann. Despite the limita-

tions posed by the hardware, the results in terms of

tuning curves were comparable to the ones obtained

in simulation. Since the behaviour of neurons and

synapses on the DYNAP-SE board is not determinis-

tic due to transistor mismatches, the curves were me-

diated over 10 sessions. Figure 6 shows the polar rep-

resentation of the orientation tuning curves for four

selected values of orientation. The spatial frequency

tuning curves and their Fourier transforms are shown

in Fig. 7; different colors refer to different temporal

frequencies of the gratings used as visual stimuli. It is

worth noting that faster gratings elicit more events on

the DVS and thus higher ﬁring rate of the neurons on

the retina layer, which then project to the relay layer

and ﬁnally to the V1 layer. Nevertheless, the curves

obtained for different temporal frequencies overlay:

this is evidence of the fact that the emergence of ON

and OFF subregions in the RF induced by recurrent

inhibition successfully normalizes the ﬁring rate in in-

put.

4 CONCLUSIONS

Neuromorphic systems are a promising alternative

to conventional von Neumann architectures in terms

of power efﬁciency, computational ﬂexibility, and

robustness. Reproducing the dynamics of biologi-

cal neurons, they represent ideal platforms to im-

plement low-power bio-inspired devices. Here, we

proposed an efﬁcient way to implement spiked-based

early-vision feature detectors on neuromorphic hard-

ware. We took inspiration from the organization of

the retinocortical pathway to obtain silicon neurons

with Gabor-like RFs that are tuned to oriented visual

stimuli with speciﬁc spatial frequency bandwidths. To

achieve that goal while keeping the number of synap-

tic connections low, we successfully used recurrent

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

536

clustered inhibition, proving its efﬁcacy in the con-

text of discrete spiking neural networks. We also ver-

iﬁed that the linearity assumption holds despite the

high non-linearity of spiking neurons. Additionally,

we showed how it is possible to combine such fea-

ture detectors to generate ﬁlters with arbitrary phase

values, effectively implementing a full harmonic rep-

resentation of the image signal. The harmonic sig-

nal description provided by the proposed neuromor-

phic circuit could be potentially used for a complete

characterization of the 2D local structure of the vi-

sual signal in terms of the phase relationships from all

the available oriented channels. This would pave the

way to the implementation of complex bio-inspired

networks for more demanding on-line visual tasks on

neuromorphic hardware.

ACKNOWLEDGEMENTS

This project has received funding from the European

Research Council under the Grant Agreement No.

724295 (NeuroAgents).

REFERENCES

aiCTX (2018). https://ai-ctx.gitlab.io/ctxctl/primer.html.

Last checked on Nov 20, 2019.

Delbruck (2007). https://github.com/SensorsINI/jaer. Last

checked on Nov 20, 2019.

Dollr, P., Appel, R., Belongie, S., and Perona, P. (2014).

Fast feature pyramids for object detection. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 36(8):1532–1545.

Fleet, D. and Jepson, A. (1993). Stability of phase infor-

mation. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 15(12):1253–1268.

Jones, J. P. and Palmer, L. A. (1987). An evaluation of the

two-dimensional Gabor ﬁlter model of simple recep-

tive ﬁelds in cat striate cortex. Journal of neurophysi-

ology, 58(6):1233–1258.

Lichtsteiner, P., Posch, C., and Delbruck, T. (2008). A

128x128 120 db 15us latency asynchronous tempo-

ral contrast vision sensor. IEEE journal of solid-state

circuits, 43(2):566–576.

Luan, S., Chen, C., Zhang, B., Han, J., and Liu, J. (2018).

Gabor convolutional networks. IEEE Transactions on

Image Processing, 27(9):4357–4366.

Milde, M., Renner, A., Krause, R., Whatley, A. M., Soli-

nas, S., Zendrikov, D., Risi, N., Rasetto, M., Burelo,

K., and Leite, V. R. C. (2018). teili: A toolbox for

building and testing neural algorithms and computa-

tional primitives using spiking neurons. Unreleased

software, INI, University of Zurich and ETH Zurich.

Moradi, S., Qiao, N., Stefanini, F., and Indiveri, G. (2018).

A scalable multicore architecture with heterogeneous

memory structures for dynamic neuromorphic asyn-

chronous processors (DYNAPs). IEEE transactions

on biomedical circuits and systems, 12(1):106–122.

uggler, E., Bartolozzi, C., and Scaramuzza, D. (2017).

Fast event-based corner detection. In Proceedings of

the British Machine Vis. Conf. (BMVC), pages 1–11.

Ogale, A. and Aloimonos, Y. (2007). A roadmap to the inte-

gration of early visual modules. International Journal

of Computer Vision, 72:9–25.

Osswald, M., Ieng, S., Benosman, R., and Indiveri, G.

(2017). A spiking neural network model of 3d percep-

tion for event-based neuromorphic stereo vision sys-

tems. Scientiﬁc Report, 7:40703.

Peirce, J., Gray, J. R., Simpson, S., MacAskill, M.,

ochenberger, R., Sogo, H., Kastman, E., and Lin-

deløv, J. K. (2019). PsychoPy2: Experiments in

behavior made easy. Behavior Research Methods,

51(1):195–203.

Raffo, L., Sabatini, S. P., Bo, G. M., and Bisio, G. M.

(1998). Analog VLSI circuits as physical structures

for perception in early visual tasks. IEEE Transac-

tions on Neural Networks, 9(6):1483–1494.

Rueckauer, B. and Delbruck, T. (2016). Evaluation of event-

based algorithms for optical ﬂow with ground-truth

from inertial measurement sensor. Frontiers in neu-

roscience, 10:176.

Sabatini, S. P. (1996). Recurrent inhibition and clus-

tered connectivity as a basis for Gabor-like receptive

ﬁelds in the visual cortex. Biological cybernetics,

74(3):189–202.

Sabatini, S. P., Gastaldi, G., Solari, F., Pauwels, K., Hulle,

M. M. V., Diaz, J., Ros, E., Pugeault, N., and Krger,

N. (2010). A compact harmonic code for early vision

based on anisotropic frequency channels. Computer

Vision and Image Understanding, 114(6):681 – 699.

Stimberg, M., Brette, R., and Goodman, D. (2019). Brian

2: an intuitive and efﬁcient neural simulator. BioRxiv,

page 595710.

Compact Early Vision Signal Analyzers in Neuromorphic Technology

537