FR3117239A1 - Training and use of an impulse neural network for the characterization of movements - Google Patents

Abstract

The subject of the invention is a system comprising: a camera configured to emit pulses each representing a change in light intensity at a given pixel; at least one calculation unit configured to: obtain pulses from said camera; on the occurrence of a pulse at a camera pixel and a time, provide a pulse neural network having at least two inputs for each camera pixel: a first pulse, at a first input for the pixel, said time; a second pulse, at a second input for the pixel, at a predefined delay after said time; upon discharge of a neuron from said pulsed neural network, reducing the activation potential of neurons belonging to the same layer and sharing the same receptive field as said neuron, for a predefined period. Figure for the abstract: Fig. 1

Description

Training and use of an impulse neural network for the characterization of movements

Field of invention

The present invention relates to the field of computer vision, and more particularly to the field of the use of neural networks for the characterization of movements of objects from a video.
Previous state of the art

Motion analysis is a constantly evolving field, serving as the basis for many applications, such as the prediction of object movement in a video. Recent developments in this field make it possible to envisage new applications, such as autonomous vehicle driving. However, the analysis and prediction of the movement of an object from a video remains an extremely complex problem.

Conventionally, videos are produced, in real time or delayed by frame cameras, producing, at a predefined frequency (for example, 25 times per second) frames encoding the entire visual field of the camera in the form of pixels. The video is therefore presented as a sequence of images independent of each other, representing the visual field of the camera at regular intervals.

More recently, so-called event cameras have been used. These cameras produce a signal representing, not successive and independent images of its visual field, but pulses (or in English, spike) representing variations of it. This allows the sensors to generate extremely sparse, meaningful data and dramatically reduce response latency.

The pulses from these cameras can be processed by spiking neural networks (SNN). These neural networks can thus take the pulses produced by the camera as input, and extract relevant information from them.

However, the use of these neural networks does not currently make it possible to finely discriminate movements within the video. In particular, the selectivity of neurons to the characteristics of a movement, such as its direction, its speed, or the shape of the object at the origin of the movement remains limited.

There is therefore a need for an impulse neural network whose neurons have a strong selectivity to the characteristics of a movement, in order to analyze in a finer way the movements of objects in the field of vision of a camera.

To this end, the subject of the invention is a system comprising: a camera configured to emit pulses each representing a change in light intensity at a given pixel; at least one computing unit configured to: obtain pulses from said camera; on the occurrence of a pulse at a pixel of the camera and a time, providing a pulse neural network having at least two inputs for each pixel of the camera: a first pulse, at a first input for the pixel, said time; a second pulse, at a second input for the pixel, at a predefined delay after said time; upon discharge of a neuron from said pulsed neural network, reducing the activation potential of neurons belonging to the same layer and sharing the same receptive field as said neuron, for a predefined period.

Advantageously, the pulsed neural network is formed of LIF neurons.

Advantageously, the reduction in the activation potential of neurons belonging to the same layer and sharing the same receptive field as said neuron consists in subtracting from their membrane potential an instantaneous inhibition value at said time, then a temporal inhibition value increasing then decreasing during said predefined duration.

Advantageously, the instantaneous inhibition value and the temporal inhibition values depend on the activation potentials of the neurons of the receptive field at said time.

Advantageously, the at least one calculation unit is further configured to predict, from a pulse in the pulse neural network, a future position of an object whose movement is at the origin of the pulses .

Advantageously, said prediction consists in adding to the position of the pulse at least one offset value characteristic of a filter having generated the pulse.

The invention also relates to a method comprising: the reception of pulses emitted by a camera, each representing a change in light intensity at a given pixel; upon occurrence of a pulse at a camera pixel and a time (t), providing a pulse neural network having at least two inputs for each camera pixel; a first pulse, at a first input for the pixel, at said time; a second pulse, at a second input for the pixel, at a predefined delay after said time; upon discharge of a neuron from said pulsed neural network, the reduction of the activation potential of neurons belonging to the same layer and sharing the same receptive field as said neuron, for a predefined duration.

The invention also relates to a method for training a pulse neural network, comprising: obtaining a training database comprising: pulses emitted by at least one camera each representing a change in luminous intensity at a given pixel; the occurrence of a pulse at a pixel of the camera and a time resulting in the supply to a pulsed neural network having at least two inputs for each pixel of the at least one camera: of a first pulse, at a first entry for the pixel, at said time; a second pulse, at a second input for the pixel, at a predefined delay after said time; applying a peak-dependent plasticity rule to the spiking neural network; upon discharge of a neuron from said pulsing neural network during said application, reducing the activation potential of neurons belonging to the same layer and sharing the same receptive field as said neuron, for a predefined duration.

Advantageously, when applying the peak-dependent plasticity rule to the pulsed neural network, the weight of a connection between a first neuron of a first layer, and a second neuron of the following layer is: increased if , when a pulse is emitted by the second neuron, a previous pulse had been emitted by the first neuron for a predefined duration before said pulse; decreased, otherwise.

The invention also relates to a computer program comprising program code instructions recorded on a computer-readable medium, said program code instructions being configured, when said program operates on a computer to execute a method according to the one of the embodiments of the invention.

Other characteristics, details and advantages of the invention will become apparent on reading the description given with reference to the appended drawings given by way of example and which represent, respectively:

a system in one set of embodiments of the invention;

an architecture of a neural network in one set of embodiments of the invention;

an example of inhibiting the activation potential of neurons upon activation of a neuron sharing the same receptive field, in one set of embodiments of the invention;

an example of a rule for temporal inhibition of the activation potential of neurons upon activation of a neuron sharing the same receptive field, in a set of embodiments of the invention;

a method of using a pulsed neural network in one set of embodiments of the invention;

a method of training a pulsed neural network in one set of embodiments of the invention;

an example of updating synaptic weights when applying a peak-dependent plasticity rule for training a spiking neural network in one set of embodiments of the invention;

an example of calculating the selectivity to the direction of filters of a pulsed neural network, in a set of embodiments of the invention;

an example of calculating the selectivity at the rate of filters of a spiking neural network, in a set of embodiments of the invention;

an example of calculation of the selectivity to the form of filters of a pulsed neural network, in a set of embodiments of the invention;

an exemplary video of a tennis match, in an exemplary embodiment of the invention for predicting a landing point of a tennis ball;

an example of evolution of an average prediction error of the arrival position of a tennis ball, in a set of embodiments of the invention as a function of the layer at which the pulse took place, in a set of embodiments of the invention;

an example of evolution of an average prediction error of the arrival position of a tennis ball, in a set of embodiments of the invention as a function of the layer at which the pulse took place, and a path presentation percentage, in one set of embodiments of the invention.

There represents system in one set of embodiments of the invention.

System 100 includes a camera 110 configured to emit pulses each representing a change in light intensity at a given pixel.

This camera can be a pulse camera (or "spiking camera"), that is to say a digital camera natively generating pulses (or "spikes" in English during significant changes in the brightness of a pixel and/or of its environment. A pulse is therefore located at the level of a pixel on which a change in light intensity takes place. In practice, the camera emits, as soon as a change is detected in the light field of a pixel, a pulse.The pulse is associated with a time, representing the time of occurrence of the pulse, relative to a fixed initial time.In one set of embodiments, the camera may emit a pulse corresponding to the onset of significant changes of luminosity (we can then for example speak of an “ON” pulse), and a pulse at the end of the significant modifications of luminosity (we can then for example speak of an “OFF” pulse).

This camera can also be a digital camera emitting successive frames of pixels of light intensity, in which differences between corresponding pixels of successive frames make it possible to generate pulses.

According to different embodiments, many types of cameras can be used. In particular, the cameras can have varying resolutions and frequencies. For example, a resolution of 128 by 120 pixels, and a frame rate of 240 images per second (or “fps”, from English “Frames per second”, having for literal translation “frames per second”) can be used.

In one set of embodiments, one or more filters are implemented by the camera for pulse generation. For example, a noise reduction filter can be used.

In one set of embodiments of the invention, an edge detection filter may also be used by the camera. For example, this filter can be a difference of Gaussians filter (also called in English DoG for “Difference of Gaussians”). This type of filter makes it possible to return at the output of the camera only the pulses defining a border, that is to say the pulses characterizing the limit of a moving object.

In one set of embodiments, pulses are computed to a camera natively capturing frames of pixel light intensities, not a camera natively generating pulses. The camera can then be configured to calculate the occurrence of pulses from the differences between successive frame pixels. Unlike cameras generating frames of pixels which generate frames at a given frequency (for example 25 fps), native pulse cameras are asynchronous, and generate pulses, associated with a time, as soon as the emission conditions of the impulse are united. A so-called “latency intensity” rule can then be applied, to assign each pulse a time between two frames. This rule consists initially in determining an "intensity" of the pulse, which is generally proportional to the absolute difference of the pixels between two successive frames, and assigning to each pulse a time between the two frames corresponding to a latency of as much weaker as the intensity of the impulse is high.

For example latency of the pulse between two frames, and the time of occurrence of the impulse can be calculated as:

• F : Index de la trame actuelle ;
• FPS : nombre de trames par seconde ;
• Δt_{s, rel}: Temps relatif du spikesà la frameF ;
• I_s: Intensité inversée de l’impulsions ;
• I_min: représente une intensité minimum sur l’ensemble des impulsions, par exemple un minimum absolu, ou l’intensité la plus faible parmi toutes les impulsions si cette donnée est disponible ;
• I_max: intensité maximum sur l’ensemble des impulsions par exemple un maximum absolu, ou l’intensité la plus faible parmi toutes les impulsions si cette donnée est disponible ;
• t_s: temps d’occurrence de l’impulsions.

With :

• F: Index of the current frame;
• FPS: number of frames per second;
• Δt _{s, rel} : Relative time from spike s to frame F;
• I _s : Inverted intensity of pulse s;
• I _min : represents a minimum intensity over all the pulses, for example an absolute minimum, or the lowest intensity among all the pulses if this data is available;
• I _max : maximum intensity over all the pulses, for example an absolute maximum, or the lowest intensity among all the pulses if this data is available;
• t _s : time of occurrence of pulse s.

It should be noted that, in these embodiments, the notion of intensity of the pulse is only used to determine the occurrence time t _s of the pulse. Once this has been determined, the pulse is only associated at the input of the neural network with a location (ie pixel where it is emitted), and an emission time, without any notion of intensity. These embodiments are given by way of example only: according to different embodiments of the invention, other methods could be implemented to associate with a pulse from a camera returning frames of pixels a time comprised between the occurrence time of the frame at which it occurs and the occurrence time of the following frame.

The system 100 also includes at least one calculation unit 120. The at least one calculation unit 120 can be any type of calculation unit capable of performing computer calculations. For example, the calculation unit can be a processor configured with machine instructions, a microprocessor, an integrated circuit, a microcontroller, a programmable logic circuit, or any other calculation unit capable of being programmed to perform calculation operations. In order to facilitate the reading of this document, the at least one calculation unit may be described as the “calculation unit”, even in embodiments involving several calculation units.

According to different embodiments of the invention, the at least one calculation unit can be a single calculation unit, or a plurality of calculation units, located in one or more calculation devices. In particular, according to different embodiments of the invention, the at least one calculation unit can be located in the camera 110 itself, or in an external calculation device, for example a computer or a server. In the case where several calculation units are used, they can be distributed between the camera and an external calculation device.

The at least one computing unit 120 is configured 121 to obtain pulses from the camera 110. As indicated above, each pulse is associated with a location (usually the pixel at which it occurred), and a time, the time being determined from a predefined origin, for example the initialization of the system. In one set of embodiments of the invention, two types of pulse can be emitted: an "ON" pulse in the event of an increase in brightness at a given pixel, and an "OFF" pulse in the event of a decrease in brightness. the luminosity.

• une première entrée pour les impulsions aux temps t auxquels elles sont émises ;
• une deuxième entrée pour les impulsions, avec un délai prédéfini δt après chaque temps t.

The at least one calculation unit is also configured to execute a pulse neural network 130. The pulse neural network 130 comprises, for each pixel of the camera, at least two inputs:

• a first input for the pulses at the times t at which they are emitted;
• a second input for the pulses, with a predefined delay δt after each time t.

In other words, the input layer of the neural network 130 comprises a number of neurons equal to at least twice the number of pixels: for each pixel, a first neuron corresponds to the initial impulse, and a second neuron corresponds to the pulse shifted by the predefined delay.

In one set of embodiments of the invention, the input to the neural network includes multiple inputs for pulses shifted by multiple predefined delays. For example, it can comprise for each pixel a second entry corresponding to an offset δt, and a third entry for a higher offset, for example 2*δt.

In one set of embodiments of the invention, the input layer of the neural network may comprise, for each pixel and for each shift delay, an input corresponding to an "ON" pulse, and an input corresponding to a “OFF” impulse. In this case, if a single predefined delay δt is used in addition to the non-time-shifted input, each pixel will be associated with 4 (2 * 2) inputs; if two predefined delays are used in addition to the non-timeshifted input, each pixel will be associated with 6 (3 * 2) inputs, etc.

A spiking neural network (or, in English, "Spiking Neural Network", abbreviated "SNN") is a type of artificial neural network (or, in English, "Artificial Neural Network", abbreviated "ANN"), in which the exchange between neurons is based on the integration of impulses and the descent of activation. In a spiking neural network, each neuron is characterized by a membrane potential, also called membrane potential or activation potential, and when the activation potential of a neuron exceeds a threshold, it “fires” , that is, it itself emits an impulse to the neurons of the next layer to which it is connected, and its activation potential returns to an initial value.

• une première impulsion, à une première entrée pour le pixel, audit temps (t) ;
• une deuxième impulsion, à une deuxième entrée pour le pixel, à un délai prédéfini (t+δt) après ledit temps.

The at least one calculation unit 120 is more precisely configured to provide 122 as input to the impulse neural network 130, on the occurrence of an impulse at a pixel of the camera and a time (t):

• a first pulse, at a first input for the pixel, at said time (t);
• a second pulse, at a second input for the pixel, at a predefined delay (t+δt) after said time.

In practice, for each pulse at time t, the pulse is inputted at time t, and at least one pulse is inputted with an offset according to a predefined delay. The number of pulses provided as input, and their delays, depend on the inputs provided in the neural network. As indicated above, the pulse can be provided on more than two different inputs, at the original time and with more than one offset according to more than one predefined delay.

The supply, for the same pixel and the same initial pulse, of the initial pulse and of at least one shifted pulse on two different outputs allows the neural network 130 to learn the spatio-temporal correlations between the pulses, and therefore to carry out, as the layers of the network progress, a selection of the relevant characteristics of the movement.

In the embodiments according to which "ON" and "OFF" pulses are provided, respectively for increases and decreases in the light intensities on the pixels, the "ON" and "OFF" pulses will be sent, with or without offset, on the inputs provided for “ON” and “OFF” pulses respectively.

Using more than two predefined time-shift delays, and/or “ON” and “OFF” pulses, allows for more complex spatio-temporal inferences, and thus further improves spatio-temporal predictions performed by the neural network.

During execution of the spiking neural network 130, when a neuron fires, the at least one computing unit is configured to perform spatial and temporal inhibition of neurons sharing the same receptive field.

This spatial and temporal inhibition consists in reducing the activation potential of neurons belonging to the same layer, and sharing the same receptive field as the neuron having discharged.

This makes it possible to increase the selectivity of the neurons to the different characteristics of a movement (such as the direction, the speed, or the shape of the object at the origin of the movement), because the discharge of a neuron on a movement corresponding to a given direction greatly reduces the probability that a neuron sharing the same receptive field will fire, for the same movement.

The system 100 therefore allows a fine detection of the characteristics of movements, in which each neuron of the network of impulse neurons is very sensitive to the different characteristics of the movement.

In one set of embodiments of the invention, the at least one calculation unit is further configured to predict 124, from a pulse in the pulse neural network, a future position of an object whose the movement is at the origin of the impulses.

Indeed, each impulse is linked to a position in the video, represented by the neuron having emitted the impulse, and characteristics of the movement, defined by the characteristic, or filter, having emitted the impulse.

Knowledge of the location of the pulse and the characteristics of the movement thus makes it possible to use the information available for the movement in order to predict the future position of the object at the origin of the movement.

In one set of embodiments of the invention, said prediction consists in adding to the position of the pulse at least one offset value characteristic of a filter having generated the pulse.

In essence, the neuron having received the impulse indicates the position of the impulse, while the characteristics of the movement, defined by the characteristic, or filter, having emitted the impulse make it possible to define an offset value, specific to each filter. , making it possible to predict a future position of the object at the origin of the movement.

It is for example possible to associate with a characteristic, or filter, a shift value of Δx pixels, for a time shift Δt. It then suffices to add this offset value Δx to the position x of a pixel to predict that the object will be located at a position x+ Δx at a time Δt.

In one set of embodiments of the invention, this time shift value will have been previously determined during a learning phase, for example by learning, for a given filter of the neural network, whether the exact positions of the objects at the at the end of the delay Δt are known in the learning phase, the offset values Δx making it possible to minimize a prediction error of these positions. For example, learning can be done by linear or polynomial regression.

There depicts an architecture of a neural network in one set of embodiments of the invention.

In one set of embodiments of the invention, the neural network 200 is formed of LIF (Leaky Integrate and Fire) neurons.

An LIF neuron is a neuron characterized by an activation potential, the potential of which increases on receipt of an impulse, and decreases with time in the absence of an impulse.

More specifically, when a neuron receives an incoming impulse, its membrane potential value increases in proportion to the weight of the connection to the presynaptic neuron from which the impulse originates. Then, this value of membrane potential leaks (or decreases) over time in the absence of reception of an impulse.

For example, the membrane potential can decrease according to the following formula:

• est le potentiel d’activation d’un neurone i à un instant ;
• est le potentiel de repos. On a en général ;
• est le temps (précédent) auquel a eu lieu la dernière impulsion pour le neurone i ;
• est une constante de temps, représente la vitesse à laquelle le potentiel de membrane fuit ;
• est le poids de la connexion entre le neurone post-synaptique i et le neurone pré-synaptique j d’où l’impulsion est issue.

In which :

• is the activation potential of a neuron i at a time ;
• is the resting potential. We have in general ;
• is the (previous) time at which the last pulse occurred for neuron i;
• is a time constant, represents the rate at which the membrane potential leaks;
• is the weight of the connection between postsynaptic neuron i and presynaptic neuron j from which the impulse originated.

The neuron fires when its activation potential exceeds a threshold value . The impulse then propagates to the next layer, and more specifically to the neurons to which neuron i is connected in the next layer.

According to different embodiments, different numbers and sizes of layers can be used.

In the example of the , an input 204 and 3 layers 210, 220, 230 are used, with forward propagation connections (known as “feed-forward” connections). Thus, the impulses are transmitted from layer to layer with the activation of the neurons. For example, pulses 201, 202, and 203 are respectively transmitted from input 200 to first layer 210, from first layer 210 to second layer 220, and from second layer 220 to third layer 230.

In one set of embodiments of the invention, the neurons are retinotopically organized, i.e. a neuron in one layer receives inputs from an area of size s x s x d (i.e. that is, an area of identical width and height s, and of depth d) of the previous layer. This area is called its receptive field (or, in English “Receptive Field”, abbreviated “RF”). This allows the neurons to be connected to the local part of the neurons of the previous layer. The other neurons in the layer each have a different receptive field, an offset of a neuron, or pixel in a layer corresponding to an offset of a neuron, or pixel, from its receptive field.

• entrée 204 : 120 x 120 x 6 ;
• 1^ecouche 210 : 116 x 116 x 60 ;
• 2^ecouche 220 : 112 x 112 x 80 ;
• 3^ecouche 230 : 106 x 106 x 100 ;
• Sortie 231 : pour chacun des 106 x 106 pixels de la 3^ecouche, un vecteur de sortie 231 peut être extrait, comprenant 231 éléments correspondant aux 100 caractéristiques et issu des 100 neurones de profondeur de la 3^ecouche.

In the example of the , the height, width and depth of each layer are as follows (width x height x depth):

• input 204: 120x120x6;
• ^1st layer 210: 116 x 116 x 60;
• ^2nd layer 220: 112 x 112 x 80;
• ^3rd layer 230: 106 x 106 x 100;
• Output 231: for each of the 106 x 106 pixels of the ^3rd layer, an output vector 231 can be extracted, comprising 231 elements corresponding to the 100 characteristics and coming from the 100 depth neurons of the ^3rd layer.

For example, the input has a height and width of 120, and a depth of 6, while the first layer has a height and width of 116, and a depth of 60.

• Impulsion « ON », temps t de l’impulsion ;
• Impulsion « ON », temps t de l’impulsion + premier délai prédéfini δt ;
• Impulsion « ON », temps t de l’impulsion + deuxième délai prédéfini δt’ ;
• Impulsion « OFF », temps t de l’impulsion ;
• Impulsion « OFF », temps t de l’impulsion + premier délai prédéfini δt ;
• Impulsion « OFF », temps t de l’impulsion + deuxième délai prédéfini δt’.

Regarding the input, the height and width of 120 correspond to an input image resolution of 120 * 120 pixels, and the depth of 6 to the delivery of a pulse at three different times (time t of the pulse, and two predefined offset times), for “ON” pulses and “OFF” pulses. There are therefore in this example 86400 (120 * 120 * 6) neurons in the input layer, i.e., for each pixel of the camera, 6 neurons:

• “ON” pulse, time t of the pulse;
• “ON” pulse, time t of the pulse + first predefined delay δt;
• “ON” pulse, time t of the pulse + second predefined delay δt';
• “OFF” pulse, time t of the pulse;
• “OFF” pulse, time t of the pulse + first predefined delay δt;
• “OFF” pulse, time t of the pulse + second predefined delay δt'.

This example is given by way of non-limiting example only, and other diaper sizes can be used.

It will nevertheless be noted that, in general, between each layer, the height and the width decrease, while the depth increases. Indeed, the size of the receptive field slightly decreases the number of pixels between each layer. For example, in the example of the , a receptive field of width and height 5 is used between each layer: the number of pixels in height in width and height is therefore reduced by 4 between each layer. In parallel, the depth increases between each layer, meaning that the amount of information collected for each pixel increases between each layer.

In one set of embodiments of the invention, the neurons are affected by a weight sharing mechanism (or, in English, “weight sharing” or “WS”). This means, in practice, that N_RF neurons (N_RFcorresponding to the depth of the layer) connected to the same receiving field have respectively the same connection weights with the previous layer as N_RFother neurons connected to another receptive field. We therefore have, for a given layer of depth N_RF, NOT_RFfilters defined by their connection weight matrix, each filter encoding a feature. Retinotopic organization and weight sharing classify this spiking neural network as a convolutional spiking neural network.

As indicated above, the pulses received at the input of the neural network 200 correspond, depending on their position in the input layer, either to an emission directly emitted by the camera at a time t, or to the shifted pulse of a predefined delay δt, δt'. This means, in practice, that in the input layer 200, for each pixel, a set of neurons distributed over the depth of the layer represent the pulses at the different times (initials or time offsets).

Thus, by optimizing the weights of the filters between the different layers, the training of the neural network 200 leads to the detection of the spatio-temporal coherences between the pulses in the form of increasingly high-level characteristics. Thus, in output on layer 3, each pixel is associated with a vector of characteristics 231 (of dimension 100 in this example), where each element of the vector corresponds to a characteristic of the output layer, representing a description of movement of height level.

There depicts an example of inhibiting the activation potential of neurons upon activation of a neuron sharing the same receptive field, in one set of embodiments of the invention.

In this example, deliberately simplified, two receptive fields 311 and 312 of a first layer 300 have been represented. Each disc in layer 300 represents a neuron.

Three filters F1, F2 and F3 are applied from this layer, i.e. three weight matrices are applied to each receptive field to arrive at 3 neurons in the next layer corresponding to this receptive field. Due to the weight sharing mechanism, the same three filters are applied to each receptive field.

• L’application du filtre F1 au champ récepteur permet de transmettre les impulsions issues du champ récepteur 311 au neurone 3111 ;
• L’application du filtre F2 au champ récepteur permet de transmettre les impulsions issues du champ récepteur 311 au neurone 3112 ;
• L’application du filtre F3 au champ récepteur permet de transmettre les impulsions issues du champ récepteur 311 au neurone 3113 ;
• L’application du filtre F1 au champ récepteur permet de transmettre les impulsions issues du champ récepteur 312 au neurone 3121 ;
• L’application du filtre F2 au champ récepteur permet de transmettre les impulsions issues du champ récepteur 312 au neurone 3122 ;
• L’application du filtre F3 au champ récepteur permet de transmettre les impulsions issues du champ récepteur 312 au neurone 3123.

In this example:

• The application of the filter F1 to the receptive field makes it possible to transmit the impulses coming from the receptive field 311 to the neuron 3111;
• The application of the filter F2 to the receptive field makes it possible to transmit the impulses coming from the receptive field 311 to the neuron 3112;
• The application of the filter F3 to the receptive field makes it possible to transmit the impulses coming from the receptive field 311 to the neuron 3113;
• The application of the filter F1 to the receptive field makes it possible to transmit the impulses coming from the receptive field 312 to the neuron 3121;
• The application of the filter F2 to the receptive field makes it possible to transmit the impulses coming from the receptive field 312 to the neuron 3122;
• The application of the filter F3 to the receptive field makes it possible to transmit the impulses coming from the receptive field 312 to the neuron 3123.

When a neuron fires, i.e. when its activation potential exceeds a threshold and it itself emits an impulse, the neurons belonging to the same layer and sharing the same receptive field are inhibited (i.e. their activation potential is reduced). For example, a discharge of neuron 3112 causes inhibition of neurons 3111 and 3113 sharing the same receptive field 311; a discharge of neuron 3121 leads to an inhibition of neurons 3122 and 3123 sharing the same receptive field 312.

• une inhibition instantanée consiste à soustraire, à la décharge d’un neurone, une valeur prédéterminée aux potentiels d’activation des neurones du même champ récepteur ;
• une inhibition temporelle consiste à diminuer le potentiel d’activation des neurones du même champ récepteur durant une période de durée prédéfinie.

This inhibition can be of two types:

• instant inhibition consists in subtracting, from the discharge of a neuron, a predetermined value from the activation potentials of the neurons of the same receptive field;
• a temporal inhibition consists of reducing the activation potential of neurons in the same receptive field for a predefined period of time.

This temporal inhibition, in combination with the presence of pulses corresponding, in addition to the initial pulse, to one or more predefined delays within the input layer, makes it possible to prevent neurons corresponding to different characteristics from unload for the same type of movement. This increases the selectivity of the neurons to the direction. During the training of the neural network, each characteristic therefore specializes in detecting a type of movement, in particular a direction and a speed.

Temporal inhibition can be of different types. For example, a fixed inhibition value can be subtracted from neurons in the same receptive field throughout the inhibition period.

In one set of embodiments of the invention, a variable temporal inhibit value is subtracted from the neurons of the same receptive field, this value being increasing then decreasing during the period of inhibition.

This allows the inhibition to be smoothed out over time. In particular, this avoids excessive inhibition at the start of the delay (the instantaneous inhibition being already applied).

There shows an example of a rule for temporal inhibition of the activation potential of neurons upon activation of a neuron sharing the same receptive field, in a set of embodiments of the invention.

The graph 400 represents the evolution of a temporal inhibition as a function of time, in a set of embodiments of the invention.

The horizontal axis 420 represents the time elapsed from the moment of neuron firing. This time can for example be expressed in seconds, in milliseconds, etc.

The vertical axis 430 represents the inhibition value. It should be noted here that the axis is inverted, to represent the effect on the membrane potential of the neurons of the receptive field: the lower one is on the axis, the stronger the inhibition, and the greater the activation potential. receptive field of neurons will be low.

• , la valeur du potentiel de la membrane excitatrice du neurone à un instant . est une valeur positive : plus sa valeur absolue est élevée, plus le neurone sera excité et aura de chance d’émettre une impulsion ;
• , la valeur du potentiel de la membrane inhibitrice du neurone à un instant . est une valeur négative, plus sa valeur absolue est élevée, plus le neurone sera inhibé et moins il aura de chance d’émettre une impulsion.

In this model, the excitation of an index neuron at a moment is the sum of two terms:

• , the value of the excitatory membrane potential of the neuron at a moment . is a positive value: the higher its absolute value, the more the neuron will be excited and will have a chance of emitting an impulse;
• , the value of the inhibitory membrane potential of the neuron at a moment . is a negative value, the higher its absolute value, the more the neuron will be inhibited and the less likely it will be to emit an impulse.

We will therefore say that the inhibition value is increasing when the absolute value of is increasing. We could say, in an equivalent way, that we subtract this absolute value, or that we add a negative value.

In this example, the time inhibit value increases then decreases during the inhibition period.

More specifically, in a set of embodiments of the invention, it increases linearly during a first half of the inhibition period to reach a maximum value at a time , then decreases linearly for a second half of the inhibition period, to return to 0 at a time 2 x .

• si , ;
• si , .

The temporal inhibit value is therefore in this example such that:

• whether , ;
• whether , .

In one set of embodiments, the instantaneous inhibition value and the temporal inhibition values depend on the activation potentials of the receptive field neurons at that time.

In other words, when a neuron fires, the more the other neurons in the receptive field have a high activation potential, the stronger the instantaneous and temporal inhibitions will be.

This makes it possible to generate strong inhibitions at the beginning of the learning phase and thus prevents neurons from learning similar patterns. Indeed, as the weights are random at the beginning of the learning phase, all the neurons can be solicited by any stimulus. Their average membrane potential at the time of the first pulse is high. During learning, selectivity becomes stronger. When one neuron generates an impulse, most other neurons connected to the same receptive field have a lower membrane potential value because they are not selective to that stimulus, resulting in lower inhibition.

For example, instantaneous inhibition values , and the maximum temporal inhibition value can both be proportional to an activation value depending on the activation potentials of the neurons of the receptive field:

• est le nombre de neurone en entrée d’un filtre où une impulsion s’est produite, c’est-à-dire dans le même champ récepteur que le neurone ayant déchargé. Il convient de noter que cela comprend le neurone ayant déchargé lui-même, et que sa valeur est ici prise avant décharge, c’est-à-dire avant propagation de l’impulsion et remise à zéro des potentiels et pour ce neurone ;
• représente l’indice du premier neurone connecté au même champ récepteur que le neurone ayant généré l’impulsion (dans une nomenclature où tous les neurones de la couche sont représentés par des indices croissants) ; les neurones du champ récepteur vont alors de l’indice à l’indice ;
• est la valeur, positive, du potentiel de la membrane excitatrice du neurone à un instant ;
• est la valeur, négative, du potentiel de la membrane inhibitrice du neurone à un instant .

Or :

• is the number of neurons at the input of a filter where an impulse has occurred, that is to say in the same receptive field as the neuron having discharged. It should be noted that this includes the neuron having discharged itself, and that its value is here taken before discharge, that is to say before propagation of the impulse and reset of the potentials And for this neuron;
• represents the index of the first neuron connected to the same receptive field as the neuron having generated the impulse (in a nomenclature where all the neurons of the layer are represented by increasing indices); the neurons of the receptive field then go from the index at the index ;
• is the value, positive, of the potential of the excitatory membrane of the neuron at a moment ;
• is the negative value of the inhibitory membrane potential of the neuron at a moment .

According to different embodiments of the invention, the maximum instantaneous and temporal inhibition values are equal. We then have, for example, . We can also have two values And different, by multiplying by two different factors And respectively. We then have , And .

There depicts a method of using a spiking neural network in one set of embodiments of the invention.

The method 500 includes a first step 510 of receiving pulses emitted by a camera each representing a change in light intensity at a given pixel.

The method 500 then comprises, the occurrence of a pulse at a camera pixel and a time (t), a second step 520 of providing a pulse neural network having at least two inputs for each camera pixel : first pulse, at a first input for the pixel, at said time (t); a second pulse, at a second input for the pixel, at a predefined delay (t+δt) after said time.

The method 500 then comprises a third step 530 of reduction, upon discharge of a neuron of said pulsed neural network, of the activation potential of neurons belonging to the same layer and sharing the same receptive field as said neuron, for a predefined duration.

All of the embodiments discussed in Figures 1 through 4 are respectively applicable to the 500 method.

The order presented in the is provided as an illustrative example only. In practice, many pulses can be received, discharges carried out, in variable orders.

There shows a method of training a pulsed neural network in one set of embodiments of the invention.

The 600 method is used to train a pulsed neural network such as the 130 neural network.

Method 600 includes a first step 610 of obtaining a training database.

This training base is made up of data of the same type as in the inference phase, i.e. pulses from one or more cameras similar to camera 110.

The training base therefore comprises pulses emitted by at least one camera, each representing a change in light intensity at a given pixel.

• d’une première impulsion, à une première entrée pour le pixel, audit temps (t) ;
• d’une deuxième impulsion, à une deuxième entrée pour le pixel, à un délai prédéfini (t+δt) après ledit temps.

The occurrence of a pulse at a pixel of a camera and a time t results in the supply to a pulse neural network having at least two inputs for each pixel of the at least one camera:

• a first pulse, at a first input for the pixel, at said time (t);
• of a second pulse, at a second input for the pixel, at a predefined delay (t+δt) after said time.

The neural network is trained by applying a Spike Timing Dependent Plasticity (STDP) rule to the spiking neural network.

The STDP rule is for example described by Bichler, O., Querlioz, D., Thorpe, SJ, Bourgoin, JP, & Gamrat, C. (2012). Extraction of temporally correlated features from dynamic vision sensors with spike-timing-dependent plasticity. Neural networks , 32 , 339-348., or Masquelier, T., & Thorpe, SJ (2007). Unsupervised learning of visual features through spike timing dependent plasticity. PLoS Comput Biol , 3 (2), e3. The STDP consists in modifying the weights of the connections between the neurons according to the time of the impulses, so that the neural network learns temporal patterns. The STDP rule is particularly effective for learning temporal patterns.

As during the inference phase, during learning, the method 600 comprises, when a neuron is discharged, a step 630 of reducing the activation potential of the neurons belonging to the same layer and sharing the same field receptor than said neuron for a predefined period.

This allows, in combination with STDP, neurons to become particularly selective to one type of movement, since, when one neuron fires on a movement, the reduction in the activation potential of other neurons considerably reduces their probability of emit an impulse for the same movement.

The temporal nature of the inhibition also avoids the over-learning of certain types of movement having generated a lot of impulses, and therefore allows learning to effectively discriminate all types of movements.

All the embodiments described with reference to FIGS. 1 to 5 are respectively applicable to method 600. In particular, weight sharing can be applied during learning.

As explained above, the learning phase can be followed by the training of a predictor of a future prediction of an object having emitted a pulse, from the occurrence of the latter.

There depicts an example of updating synaptic weights when applying a spike-dependent plasticity rule for training a spiking neural network in one set of embodiments of the invention.

• augmenté si, à l’émission d’une impulsion par le deuxième neurone, une impulsion précédente avait été émise par le premier neurone pendant une durée prédéfinie avant ladite impulsion ;
• diminué, dans le cas contraire.

In one set of embodiments of the invention, when applying the peak-dependent plasticity rule to the spiking neural network, the weight of a connection between a first neuron of a first layer, and a second neuron of the next layer is:

• increased if, when a pulse is emitted by the second neuron, a previous pulse had been emitted by the first neuron for a predefined duration before said pulse;
• decreased, otherwise.

In other words, the weight of a connection between two neurons will be reinforced if an impulse from the second neuron of the next layer is frequently preceded by an impulse from the first neuron, and diminished otherwise.

This has the effect of making it possible to strengthen the temporal correlations between the pulse emissions between the different layers. Indeed, over the course of learning, the weights of the connections will be reinforced when nearby pulses are detected.

In the example of figure 7, the horizontal axis 710 represents the time, and the vertical axis 720 the variation with a connection weight between a first neuron i and a second neuron j.

At time 730, an impulse is emitted by the second neuron j.

The so-called LTP window (from the English “Long Term Potentiation”, in French “Potentialisation à Long Terme”) 740 corresponds to an increase in the weight of the connection: if an impulse had been previously emitted from the first neuron i for a duration (corresponding to the width of the window) before time 730, i.e. if the time of the last impulse emitted by the first neuron i, and the time 730 of the current impulse emitted by the second neuron j are such that , the weight of the connection between the first and the second neuron is reinforced: , Or is a weight gain factor.

Otherwise, if , the weight of the connection is reduced, we then apply an “LTD” (from English “Long Term Depression”, in French “Dépression à Long Terme”): , Or is a weight reduction factor, generally lower than .

This learning rule is therefore characterized by a few parameters, in particular , And . These parameters can be defined manually, or determined automatically to minimize a cost function.

In one set of embodiments of the invention, these parameters are determined automatically by a genetic algorithm, to minimize the cost function. This makes it possible to quickly identify an optimal combination of parameters, while avoiding falling into a local optimum. Different cost functions can be used. For example, the parameters can be determined to minimize an average error prediction.

There shows an example of calculating the selectivity to the direction of neurons of the final layer of a spiking neural network, in a set of embodiments of the invention.

One of the objectives of the invention is to train a neural network with impulses, so that the filters of the last layer provide information on the movement of an object on the video. In particular, each element of the output vector for each pixel of the last layer (for example, vector 231) must encode a distinct characteristic of the movement, characterized in particular by a direction. Direction selectivity represents how well each feature in the output layer represents a given direction of motion.

To test this property, the following experiment was conducted. A training game was generated, filming tennis matches with a camera such as the 110 camera, therefore including very strong movements of the balls. In parallel with the video capture, the 3D trajectories of the bullets were recorded, using infrared targets placed on the bullets, and specific cameras measuring the 3D position of the targets at a frequency of 200 Hz. 6 Hz has been applied to the positions. The position series allows you to directly obtain the direction and speed of a ball at any time.

This therefore made it possible to determine a matrix Vk of positions Xvk and Yvk, speeds Svk, directions θvk, which can be compared to the output of the neural network to characterize the selectivity of the filters: V _k = [Xv _k Yv _k θv _k Sv _k ].

Thus, each pulse on a filter can be associated with a movement defined by a direction: when a pulse is detected at a point of coordinates X, Y, at a time t, it is possible to identify in the matrix V _k the corresponding movement, and associate the speeds and directions to the filter. Thus, a filter can be associated with a set of movement directions having led it to generate an impulse.

There thus represents the range of directions generating a pulse for each filter: for each filter, an average direction θ is calculated; all the directions associated with this filter are normalized around θ, and represented by a vector on the . For example, if the average of the directions for the filter is 275, and a vector associated with this filter has a direction of 277, we will have for this filter θ = 275, and for this vector a direction θ + 2.

There shows that the directions are very little dispersed, for each filter, around the mean direction θ. This means that the filters of the trained neural network have good direction selectivity: the pulses emitted by the same filter will have been generated by a set of movements having almost identical directions.

Directional selectivity can also be expressed by a metric. For example, as explained above, we can for a filter n define all the vectors of V_k about this filter. We will then have vectors Xfn, Yfn, θfn, Sfn, representing respectively all the X, Y positions, directions and speeds of a ball for all the pulses of the filter n.

The selectivity in the direction DSn can then be estimated, for a filter n, as the standard deviation of the vector θfn. It is then possible to calculate a weighted average of the selectivities of the filters, by weighting the selectivity in the direction DSn of a filter n by the number of pulses Cn emitted for this filter. We then obtain a weighted average:

In a first test, this selectivity was calculated as equal to 0.3 rad (17.5 degrees).

In a second test, only pulses located at a distance of less than 10 pixels from the position of a ball were used. Indeed, even if, within the framework of this test, the movements of the balls are the most important, the filters can also generate impulses for other movements (arms of the tennis players, etc). Considering only impulses very close to a bullet, the average selectivity becomes, in this example and this simulation, 0.258 rad.

This example demonstrates that a spiking neural network according to the invention allows excellent direction selectivity.

There shows an example of calculating the selectivity at the rate of filters of a spiking neural network, in a set of embodiments of the invention.

Within the framework of the same experiment as that described with reference to the , it is possible to test the selectivity of the filters at speed.

Indeed, the knowledge of the real trajectories of bullets makes it possible to associate, with each pulse of a filter, a bullet trajectory associated with a speed.

The curves 910, 920, 930 represent, for three filters, a probability density distribution of the speeds of bullets having generated pulses.

The horizontal common axis 940 represents the ball speed, as indicated by the infrared camera. The vertical axes 912, 922, 923 represent the probability density of ball velocities having generated a pulse for each of the three filters respectively.

The three filters chosen for this representation have a similar direction. In addition, to complete the representation, for each neuron, a random selection of ball trajectories having generated an impulse for a direction substantially equal to that of the neuron was carried out. The distributions of the speeds of the trajectories of these random selections have been represented for each neuron, respectively by the curves 911, 921, 931.

These representations show that, for the three filters, the distributions 911, 921 and 931 of the random selections are substantially equivalent for the three filters, and relatively homogeneous over the speed range.

On the contrary, the distributions 910, 920 and 930 of the trajectories having generated impulses for each of the three neurons are much “tighter”, and each centered on a different speed.

This means that the training of the neural network allows them to code not only directions, but also different speeds: even with an overall equivalent direction, the three neurons here have coded significantly different speeds.

This example shows the ability of the invention to discriminate not only the direction, but also the speed of movements.

There represents an example of calculation of the selectivity to the form of filters of a pulsed neural network, in a set of embodiments of the invention.

In the context of the experiment described in Figures 8 and 9, another element of interest is the determination of the object that generated the pulses. Indeed, if, within the framework of a video of a tennis match, the majority of the impulses are logically generated by movements of the ball, other movements, such as movements of rackets, or of the arms of tennis players , can generate pulses.

For all the impulses generated by each filter of the final layer of the neural network, it is possible to determine a distance, in number of pixels, between the neuron for which this impulse takes place, and the position of the center of the ball. when this impulse occurs. Then, the average of the distances to the ball, for the pulses emitted by each filter, is here determined.

A threshold of 10 pixels was defined, to characterize the impulses generated by a movement of the ball, and the impulses emitted by other movements.

The graph 1000 represents the distribution of the average distances to the ball for the various filters. The horizontal axis 1011 represents the average distance, in pixels. The vertical axis 1012 represents the number of filters having said average distance. The distances are grouped by group of 5 pixels. For example, the bar 1020 indicates that 35 filters have an average distance between [0 pixels; 5 pixel [ ; bar 1021 indicates that 27 filters have an average distance in the interval [5 pixels; 10 pixels [, etc.)

Bars 1020 and 1021 represent the filters associated with an average distance to the ball of less than 10 pixels, whose pulses are considered to be generated by movements of the ball. The other bars represent filters associated with an average distance to the ball greater than 10 pixels, whose pulses are considered to be generated by other objects.

The positions of the pulses generated by the various filters were also represented, superimposed, on an image of the video of the tennis match. This shows that some filters encode specific motion phases, such as an upward phase of the ball to the right.

These examples show that the filters of a neural network trained according to the invention are also shape-selective, since some filters correspond to movements of the ball, and other filters to movements of other objects.

There depicts an exemplary video of a tennis match, in an exemplary embodiment of the invention for predicting a receiving point of a tennis ball.

Image 1100 represents a frame from a video of a tennis match. The video is captured in a plane parallel to the two players, i.e. the trajectories of the ball are substantially parallel to the camera. Trail 1110 represents bullet trajectories captured by the camera. The planes 1111 and 1112 represent the planes in which the two players respectively recover the balls.

The objective of this experiment is to predict, from an impulse on the last layer of the neural network, the position at which the ball will be received. Indeed, each pulse is linked to a filter and a given neuron. The neuron corresponds to the location of a pixel in the video, and the filter to a motion characteristic. Knowing the movement of the ball at a given location, it is possible to predict the location of the reception of the ball.

For this, a polynomial regression is trained. For example, the following regression can be used:

• représente la position verticale entre le neurone ayant généré l’impulsion et la position verticale prédite du point de réception de la balle, c’est-à-dire la position verticale sur l’un des axes 1111 (dans le cas où le mouvement va de droite à gauche) et 1112 (dans le cas où le mouvement va de gauche à droite). Cette position permet de connaître la position finale de réception de la balle, en s’additionnant à la position verticale du neurone pour lequel l’impulsion a été générée ;
• et représentent respectivement les positions verticales et horizontales du mouvement pour le filtre ayant généré l’impulsion.

In which :

• represents the vertical position between the neuron having generated the impulse and the predicted vertical position of the point of reception of the ball, that is to say the vertical position on one of the axes 1111 (in the case where the movement goes from right to left) and 1112 (in the case where the movement is from left to right). This position makes it possible to know the final position of reception of the ball, by being added to the vertical position of the neuron for which the impulse was generated;
• And respectively represent the vertical and horizontal positions of the movement for the filter having generated the pulse.

At the training step, the final receiving position is known. A training game can therefore be constituted for each filter by associating, for each pulse of the filter, a difference observed between the vertical position of the impulse and the vertical position of reception of the ball. The settings , , , , And are thus determined for each filter by linear regression.

In the inference phase, the values , , , , And associated with the filter make it possible to directly determine the estimated difference in vertical position , and therefore the position of the ball.

There represents an example of evolution of an average prediction error of the arrival position of a tennis ball, in a set of embodiments of the invention as a function of the layer at which the impulse took place , in a set of embodiments of the invention.

The graph 1200 represents the prediction error on the reception position of the ball, in number of pixels, within the framework of the experiment described in , depending on the layer where the impulse took place.

• en entrée ;
• sans apprentissage (quelle que soit la couche) ;
• dans la première couche ;
• dans la deuxième couche ;
• dans la troisième couche.

The horizontal axis lists the different possible layers:

• entrance ;
• without learning (regardless of the layer);
• in the first layer;
• in the second layer;
• in the third layer.

The vertical axis represents the number of average error pixels, for each of these layers.

This graph shows that, for a trained neural network, the error decreases as a function of the depth of the layers: it is already lower in the 1 ^st layer than at the input, then decreases sharply for the 2 ^nd and 3 ^rd layers.

This shows that, as the layers progress, the neurons do indeed specialize in coding an increasingly specific movement, which can be associated with a determined movement of an object (in this case, the ball).

There represents an example of evolution of an average prediction error of the arrival position of a tennis ball, in a set of embodiments of the invention as a function of the layer at which the impulse took place , and a percentage of presentation of the trajectory, in a set of embodiments of the invention.

The graph 1300 presents, for each of the conditions mentioned in , the evolution of the prediction accuracy as a function of the percentage of presentation of the trajectory of the ball.

The vertical axis 1310 represents the average error, in pixels.

The horizontal axis 1320 represents the percentage of presentation of the trajectory, that is to say the percentage of trajectory already achieved by the ball on the occurrence of an impulse. For example, a percentage of 50% means that the impulse considered occurs when the ball has already completed half of its trajectory.

• courbe 1330 : évolution de l’erreur moyenne, pour les impulsions émises en entrée ;
• courbe 1331 : évolution de l’erreur moyenne, pour les impulsions émises par un réseau avant apprentissage (quelle que soit la couche) ;
• courbe 1332 : évolution de l’erreur moyenne, pour les impulsions émises par la première couche, après apprentissage ;
• courbe 1333 : évolution de l’erreur moyenne, pour les impulsions émises par la première couche, après apprentissage ;
• courbe 1334 : évolution de l’erreur moyenne, pour les impulsions émises par la première couche, après apprentissage.

Curves 1330, 1331, 1332, 1333 and 1334 represent respectively:

• curve 1330: evolution of the mean error, for the pulses emitted at the input;
• curve 1331: evolution of the average error, for the pulses emitted by a network before learning (regardless of the layer);
• curve 1332: evolution of the mean error, for the pulses emitted by the first layer, after learning;
• curve 1333: evolution of the average error, for the pulses emitted by the first layer, after learning;
• curve 1334: evolution of the mean error, for the pulses emitted by the first layer, after learning.

This graph shows that the precision becomes more important with the percentage of presentation. This result is logical since the closer the impulse is to the reception zone, the more the movement will be predicted in the short term. Interestingly, this is not found on entry 1330, i.e. the information of the input pulse alone does not predict the arrival location of the ball.

On the contrary, the 2 ^nd and 3 ^rd layers, represented by the curves 1333 and 1334 allow, after training, an effective prediction, even descending to a precision of 2 or 3 pixels at the end of the trajectory for the 3 ^rd layer.

This example shows that the neural network according to the invention allows efficient prediction of movements, especially in its last layer.

The examples above demonstrate the ability of the invention to train and use a neural network to achieve strong sensitivity to movement characteristics. However, they are only given by way of example and in no way limit the scope of the invention, defined in the claims below.

Claims (10)

System (100) comprising:

• a camera (110) configured to emit pulses each representing a change in light intensity at a given pixel;
• at least one calculation unit (120) configured for:
  • obtaining (121) pulses from said camera;
  • upon occurrence of a pulse at a camera pixel and a time (t), providing (122) to a pulse neural network (130) having at least two inputs for each camera pixel:
    • a first pulse, at a first input for the pixel, at said time (t);
    • a second pulse, at a second input for the pixel, at a predefined delay (t+δt) after said time;
  • upon discharge of a neuron from said pulsed neural network, reducing (123) the activation potential of neurons belonging to the same layer and sharing the same receptive field as said neuron, for a predefined duration.

A system according to claim 1, wherein the spiking neural network is formed of LIF neurons. System according to one of Claims 1 to 2, in which the reduction of the activation potential of the neurons belonging to the same layer and sharing the same receptive field as the said neuron consists in subtracting from their membrane potential an instantaneous inhibition value at the said time, then an increasing and then decreasing temporal inhibition value during said predefined duration. A system according to claim 3, wherein the instantaneous inhibition value and the temporal inhibition values depend on the activation potentials of the receptive field neurons at said time. System according to one of Claims 1 to 4, in which the at least one calculation unit is further configured to predict (124), from a pulse in the pulse neural network, a future position of an object whose movement is at the origin of the impulses. System according to Claim 5, in which the said prediction consists of adding to the position of the pulse at least one offset value characteristic of a filter having generated the pulse. Method (500) comprising:

• receiving (510) pulses emitted by a camera each representing a change in light intensity at a given pixel;
• upon the occurrence of a pulse at a camera pixel and a time (t), providing (520) to a pulse neural network (130) having at least two inputs for each camera pixel:
  • a first pulse, at a first input for the pixel, at said time (t);
  • a second pulse, at a second input for the pixel, at a predefined delay (t+δt) after said time;
• upon discharge of a neuron from said pulsed neural network, reducing (530) the activation potential of neurons belonging to the same layer and sharing the same receptive field as said neuron, for a predefined duration.

A method (600) of training a spiking neural network (130), comprising:

• obtaining (610) a training database comprising:
  • pulses emitted by at least one camera each representing a change in light intensity at a given pixel;
  • the occurrence of a pulse at a pixel of the camera and a time (t) causing the provision to a pulsed neural network (130) having at least two inputs for each pixel of the at least one camera:
    • a first pulse, at a first input for the pixel, at said time (t);
    • a second pulse, at a second input for the pixel, at a predefined delay (t+δt) after said time;
• applying (620) a peak-dependent plasticity rule to the spiking neural network;
• upon discharge of a neuron from said pulsing neural network during said application (620), reducing (630) the activation potential of neurons belonging to the same layer and sharing the same receptive field as said neuron, for a predefined duration.

A method according to claim 8, wherein, upon applying the peak-dependent plasticity rule to the spiking neural network, the weight of a connection between a first neuron of a first layer, and a second neuron of the next layer is:

• increased if, when a pulse is emitted by the second neuron, a previous pulse had been emitted by the first neuron for a predefined duration before said pulse;
• decreased, otherwise.

A computer program comprising program code instructions recorded on a computer-readable medium, said program code instructions being configured, when said program runs on a computer to execute a method according to one of claims 7 to 9.