TECHNICAL FIELD OF THE INVENTION

The present invention concerns a neuromorphic circuit implementing an oscillatory neural network.

BACKGROUND OF THE INVENTION

The development of many applications necessitates a large amount of data process with complex processing methods carried out by calculation units.

Autonomous vehicle is a striking example of such applications. Indeed, such specific application also requires that the physical implementation of the calculation units and the data storage means adapted to carry out the processing methods to be close to the sources of input data, such as the sensors.

One way to fabricate such calculation units is to use a CMOS (Complementary metal- oxide-semiconductor) implementation.

Indeed, innovations in CMOS technology and the continuous scaling roadmap of transistors outlined by Moore’s prediction have enabled today’s powerful computers and handheld devices. Mere miniaturization of devices was initially sufficient to reduce the area and power requirements of transistors, yet for sub-100 nanometers (nm) technology nodes, this was not enough. Two main paths were taken 1 ) to change the device materials to reduce its parasitics, and 2) change the device geometry for better channel control. First, the metal gates and high-K oxides were introduced to improve the transistor performance. Such material level improvement eventually led to a structural change and the introduction of new gate geometry such as fin field-effect transistors (FinFET). FinFETs allowed controlling the channel from three sides. Recently, TMSC has announced volume production as of 2020 of its 5nm gate-all-around FinFET transistors made available by EUV process technology and with a target to start production of 3 nm node by 2022.

Nevertheless, despite the advancements in the transistor device and fabrication technologies, CMOS is facing physical barriers - as scaling is approaching a fundamental physical limit with the transistor channel length becoming comparable to the size of a handful of atoms. Such channel lengths lead to significant leakage currents and suffer from lower yield due to high process variations. Consequently, this would translate to more power consumption and more expensive chips, that would be an overkill to what Moore’s law has been promising so far. At this point, the scientific and industrial communities have focused on developing novel devices that go beyond CMOS transistors. Emerging memories such as magnetic and phase change (PCRAM, RRAM, STT-RAM), and new transistor technologies such as tunnel, negative capacitance and 1 D/2D channel material (TFET, NC- FET, CNT-FET/MOS2-FET) are being investigated as potential solutions to extend the performance and capacity of Von Neumann computing paradigm.

Despite the on-going research on novel device geometries and channel materials, there is a tremendous effort on exploring innovative non-Von Neumann computing architectures to meet the requirements of data-centric applications. In the classical von Neuman architecture, data moves from memory to the processor, which for processing large datasets becomes infeasible as a large amount of power is consumed in data movement, hence, arises the memory-wall problem.

Non-Von Neumann architectures like brain-inspired architectures based on neural networks have drawn a lot of interest as more understanding of how the brain and neurons work is gained. Neural networks aim to mimic the parallelism of the brain and their implementation in resource-intensive hardware such as GPUs have revolutionized Al applications. For example, current CMOS implementations of neural networks such as Google’s Tensor Processing Unit can offer up 86X more computations per watt. Even though these systems are more power-efficient compared to a CPU due to their architecture, the CMOS implementations of neural networks will eventually face the problems described earlier. Neuromorphic circuits appear to be the solution to go beyond Von Neuman architecture. These systems are based on brain architecture with artificial neural networks made of synapses and neurons. Many artificial neural network algorithms for machine learning are already used on software, such as spiking neural networks, convolutional neural networks, Hopfield neural networks. Their integration into hardware appeared in the last decade and revolutionized the world of artificial intelligence by enabling parallel architecture.

An alternative computing approach based on artificial neural networks uses oscillators to compute or oscillatory neural networks. Such an approach differs from classical CMOS and classical von Neumann where building blocks are analog and perform computations efficiently. Moreover, data is encoded on the oscillator signals phase, which is a departure from the classical voltage level-based data encoding (such as amplitude voltage to represent a logical bit T or ’O’). Oscillatory neural networks can perform computations efficiently and can be used to build a more extensive neuromorphic system.

SUMMARY OF THE INVENTION

There is therefore a need for a compact neuromorphic circuit adapted to implement an oscillatory neural network.

To this end, the specification describes a neuromorphic circuit implementing an oscillatory neural network, the neuromorphic circuit comprising :

• neuron units, each neuron unit being an analog oscillator adapted to convert an incoming analog signal having a specific waveform on an analog input of the neuron unit into a digital output signal, and

• synapse units, each synapse unit being adapted to multiply a digital output from a neuron unit with a respective synaptic weight to output a weighted signal sent to the analog input of a neuron unit.

According to further aspects of the neuromorphic circuit, which are advantageous but not compulsory, the neuromorphic circuit might incorporate one or several of the following features, taken in any technically admissible combination:

• each neuron unit comprises:

• an hysteresis block having an input and an output, the output of the hysteresis block being a digital output signal, and

• a shaper block converting the digital output signal of the hysteresis block into said analog signal having a specific waveform, the shaper block having an input and an output, and • a connection between the input of the hysteresis block and the output of the shaper block wherein said analog signal is injected.

• the hysteresis block is a regenerative comparator with hysteresis.

• the hysteresis block is a Schmitt trigger.

• the shaper block comprises a commanded pull-up current source and a commanded pull-down current source.

• the specific waveform is triangular.

• each synapse unit comprises at least one of a synaptic weight amplitude block and a synaptic phase block.

• the synaptic weight amplitude block comprises at least one capacitor.

• the synaptic weight amplitude block comprises several programmable capacitors.

• the synaptic phase block comprises a delay unit.

• each synapse unit further comprises a control unit adapted to convert the synaptic weight in an amplitude value and a phase value, to send the amplitude value to the synaptic weight amplitude block and to send the phase value to the synaptic phase block.

• the neuromorphic circuit comprises digital connections between the synapse units and the neuron units.

• the neuromorphic circuit is a programmable circuit, such as an ASIC.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention will be better understood on the basis of the following description which is given in correspondence with the annexed figures and as an illustrative example, without restricting the object of the invention. In the annexed figures:

• figure 1 is a schematic representation of a neuromorphic circuit implementing an oscillatory neural network with enlarged block diagrams of respectively a neuron unit and a synapse unit,

• figure 2 is a schematic representation of an example of circuit of a neuron unit according to figure 1 , and

• figure 3 is a schematic representation of an example of circuit of a synapse unit according to figure 1.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

A neuromorphic circuit 10 is represented in figure 1 .

The neuromorphic circuit 10 is a circuit implementing an oscillatory neural network. The abbreviation ONN is often used to designate the oscillatory neural network.

A neural network is a mathematical function made of a set of neurons linked by synapses.

A synaptic weight is associated with each synapse. It is often a real number, which takes both positive and negative values. In some cases, synaptic weight is a complex number.

A neural network is an oscillatory neural network when the neurons are oscillators.

Unlike any other neural network such as the most commonly used spiking neural network, in oscillatory neural networks, the information is computed in the phase domain rather than the time domain. By describing neurons as oscillators, it is the phase difference between oscillating neurons that enables to encode information rather than the rate of spikes as in spiking neural network.

This means that oscillatory neural networks are coupled oscillators with distinctive phase differences. The output is encoded on the phase differences to represent either in- phase (i.e. logic value 0) or out-of- phase (i.e. logic value 1 ).

Distinctive phase relations are obtained by the synchronization of the coupling network dynamics. Phase differences correspond to the memorized patterns in the network.

The neuromorphic circuit 10 is a programmable circuit, such as an ASIC.

ASIC stands for an application-specific integrated circuit and designates an integrated circuit (IC) chip customized for a particular use, rather than intended for general-purpose use.

The neuromorphic circuit 10 comprises neuron units 12 and synapse units 14.

In the present example, four neuron units 12 are represented on figure 1 and 16 synapse units 14.

The 16 synapse units 14 are arranged in an array, whose lines and column are respectively indexed by an integer m and an integer n.

The first neuron unit 12 is linked to each synapse units 14 whose line is indexed by an integer m = 1 .

However, this is only for an illustrative purpose, the number of neuron units 12 and synapse units 14 being as desirable for the application.

The neuromorphic circuit 10 further comprises digital connections 18 between the synapse units 14 and the neuron units 12.

The neuron units 12 are here analog circuits, wherein phase computation occurs while Digital synaptic signals propagate through the synapse units 14 in a feedforward manner from the output of a neuron unit 12 to the input of other neuron units 12. This implies that the propagation is carried out with digital connections 18 and achieved in the digital domain.

This means that the neuromorphic circuit 10 is a mixed circuit with calculation in the analog domain and propagation in the digital domain.

This notably enables to benefit from known and efficient digital circuit design to achieved bufferization or multiplexing. This also facilitates the integration of the neuromorphic circuit 10 with other CMOS circuits.

The neuron units 12 are achieving the role of neurons in the ONN implemented by the neuromorphic circuit 10.

In the present case, the neuron unit 12 has the role of ensuring that phase computing is achieved via the processing of an incoming analog signal having a specific waveform.

The neuron unit 12 also has the role of generating output digital signals. Such digital signals will be used to propagate the calculated phase information through the ONN both by a feedforward propagation and a recurrent signal propagation.

This means that the neuron unit 12 has an analog input 12 ¹ and a digital output 12°.

In what follows, each block or unit with the reference sign X has an input labelled X ¹ and an output labelled X°.

For this, in the present case, each neuron unit 12 is an analog oscillator adapted to convert an incoming analog signal having the specific waveform into a digital output signal.

Each neuron unit 12 is thus a relaxation oscillator.

Many specific waveforms can be considered, such as sine wave, sawtooth wave, or triangular wave.

Preferably, the specific waveform is triangular.

Accordingly, the incoming analog signal is analog triangular input oscillations.

In what follows, as a specific example, it is assumed that the specific waveform is triangular.

As apparent in the enlarged block diagram of figure 1 , each neuron unit 12 comprises a hysteresis block 20, a shaper block 22 and a connection unit 24 connecting the two blocks 20 and 22.

The hysteresis block 20 holds the neuron state and commands the shaper block 22.

The output of the hysteresis block 20 on its output 20° is a digital output signal as apparent in the figures.

An example of a specific hysteresis block 20 can be found in figure 2.

In this figure 2, the hysteresis block 20 is a regenerative comparator with hysteresis.

For this, the hysteresis block 20 comprises a series of transistors Q1 to Q11 arranged to achieve such function. Notably, the transistors Q7 and Q8 apply the positive feedback responsible for the hysteresis behavior. The inverter formed by Q5 and Q1 1 produces the digital output 20°.However, other configurations achieving a similar function are known.

Thus, in a more general way, the hysteresis block 20 is a Schmitt trigger.

A Schmitt trigger is a comparator circuit with hysteresis implemented by applying positive feedback to the noninverting input of a comparator or differential amplifier.

It can also be noted here that the hysteresis block 20 further comprises a transistor Q12 enabling to start the oscillation.

The transistor Q12 therefore serves as a switch with two states, an “OFF-state” wherein the oscillation is present and an “ON-state” wherein the oscillation is absent.

Thus, by adjusting the moment when the transistor Q12 turns off, the phase between the neuron units 12 is initialized. In other words, the gate of the transistor Q12 constitutes an initialization entry of the neuron unit 12.

The shaper block 22 is adapted to convert the digital output signal of the hysteresis block 20 into the analog signal having the triangular waveform.

In other words, driven by the hysteresis block 20, the shaper block 22 is adapted to produce an analog signal on its output 22°, which is fed back to the hysteresis block 20 via an electrical connection 34 to produce oscillations.

The output 22° of the shaper block 22 is the input 12 ¹ of the neuron unit 12.

According to the case of figure 2, the shaper block 22 comprises a commanded pull- up current source 26 and a commanded pull-down current source 28.

Each current source 26 or 28 comprises a transistor 30 and a current generator 32.

The gate of each transistor 30 is connected to the output of the hysteresis block 20, so that each transistor 30 is controlled by the digital output signal of the hysteresis block 20.

The source of the transistor 30 of the pull-up current source 26 is connected to the alimentation source and thus put at a potential equal to V _d d.

The source of the transistor 30 of the pull-down current source 28 is connected to the ground.

Each current generator 32 is connected to the drain of a respective transistor 30 on one end and connected to each other on the other end.

The middle point between the current generators 32 corresponds to the output 22° of the shaper block 22.

The connection unit 24 is a physical interface circuit between the input of the hysteresis block 20 and the output of the shaper block 22 wherein the analog signal is injected. In the present example, the connection unit 24 comprises the electrical connection 34 and a capacitor 36.

The electrical connection 34 is connected at one end to the input 20 ¹ of the hysteresis block 20 and at the other end to the output 22° of the shaper block 22.

The capacitor 36 is connected in parallel with the electrical connection 34.

The capacitor 36 is connected at one end to the ground.

The synapse units 14 are adapted to implement a synapse function.

In neural networks, a synapse makes the multiplication between its input and the synaptic weight.

Each synapse unit 14 is adapted to multiply a signal emitted by the digital output 12° from a neuron unit 12 with a respective synaptic weight, resulting in a weighted signal sent to the analog input 12 ¹ of the neuron unit 12

The synaptic weight can be written as:

Where:

• K _mn is the synaptic weight of the synapse unit 14 corresponding to the line m and the column n,

• C _mn is the synaptic weight amplitude, that is the amplitude of the synaptic weight mn’

• j is the complex number, and

• 6 _mn is the synaptic weight phase, that is the phase of the synaptic weight K _mn .

According to the enlarged block diagram of figure 1 , each synapse unit 14 comprises a synaptic weight amplitude block 38 and a synaptic phase block 40.

Each synapse unit 14 further comprises a control unit 42.

An example of implementation of a synaptic weight amplitude block 38 is represented on figure 3.

The synaptic weight amplitude block 38 implements the synaptic weight amplitude C

The synaptic weight amplitude block 38 comprises several programmable capacitors 44, the programmable capacitors 44 being in parallel.

As an example, each programmable capacitor 44 is a combination of a switch 46 and a capacitor 48 in series.

Preferably, the switches 46 and the capacitors 48 are the same components.

There is, in this specific example, the number of programmable capacitors 44 is equal to 4, so that the switches 46 are named BO to B3 and the capacitors 48 are named CO to C3 on figure 3. With 4 programmable capacitors 44, the synaptic weight amplitude C _mn is coded with a 4-bit resolution.

Of course, this number being non limitative, the resolution depending from the use of neuromorphic circuit 10.

The synaptic weight amplitude block 38 is connected via the synaptic phase block 40 to the input 14 ¹ of the synapse unit 14.

The synaptic weight amplitude block 38 thus implements the multiplication between its input voltage corresponding in amplitude to the input 14 ¹ of the synapse unit 14 and its capacitance value, the capacitance value depending from the state of the switches BO to B3 and the values of the capacitors CO to C3.

The capacitors CO to C3 thus converts the digital input waveform to a charge injected to the oscillator analog input, resulting in phase shift. In other words, a synapse unit 14 acts as a digital to analog phase converter.

Such signal corresponds the output signal of the synapse unit 14 since the output 38° of the synaptic weight amplitude block 38 is connected to the output 14° of the synapse unit 14

In variant, the synaptic weight amplitude block 38 may be any number of capacitors 48, and notably one.

The synaptic phase block 40 is adapted to implement a synaptic weight phase 9 _mn .

The synaptic phase block 40 comprises a delay unit 52.

The delay unit 52 is adapted to delay the signal injected at the input 14 ¹ by a given quantity.

In the present example, the quantity is r _mn = T-^ where T is the period of oscillation.

An example of implementation of a delay unit with 1 -bit of precision for the synaptic weight phase 9 _mn is represented in figure 3, where 9 _mn = 0 or 9 _mn = it.

The delay unit 52 comprises an inverter that delays the synapse input signal 14 ¹ by half an oscillation period.

Hence, the digital output 52° is phase-shifted by 180° with respect to the digital input 14', and corresponds to the case where 9 _mn = it.

The case where 9 _mn = 0 occurs when the synapse input signal 14 ¹ is directly fed to the weight amplitude block 38 without any delay.

The multiplexer 54 is used to select one of the two digital signals 52° or 14 ¹ , namely if 9 _mn = n or 9 _mn =0, respectively. In other words, the multiplexer 54 sets the sign of the synaptic weight in the example of figure 3. More precision on the synaptic weight phase 0 _mn can be obtained by adding more inverters to the delay unit 52 and more multiplexer entries.

The synaptic phase block 40 thus applies a delay r _mn to its input signal corresponding to the input 14 ¹ of the synapse unit 14, the delay value being obtained from the control unit 42.

The control unit 42 is adapted to convert the synaptic weight K _mn in an amplitude value C _mn and a phase value 0 _mn and to send the amplitude value C _mn to the synaptic weight amplitude block 38 and to send the phase value 0 _mn to the synaptic phase block 40.

For instance, an electrical signal is sent to the switches 48 so that they change their state (open or closed) in accordance with the amplitude value C _mn .

Similarly, an electrical signal is sent to the multiplexer 54 of the synaptic phase block 40 to select the multiplexer entry corresponding to the synaptic phase value 0 _mn , as schematically illustrated by the element 60 in figure 3.

Figure 3 illustrates schematically a possible implementation of such control unit 42 as a set of registers 56.

Each synapse unit 14 is therefore adapted to code for any complex synaptic weight K _mn with any value of precision.

Alternatively, each synapse unit 14 is adapted to code for a synaptic weight either limited to a unique amplitude or a unique phase.

In such cases, the synapse unit 14 comprises only one block among the synaptic weight amplitude block 38 and the synaptic phase block 40 and the control unit 42.

In operating, it can be shown that such specific waveform leads to phase dynamics fulfilling the following equation:

Where:

• L> ₀ is the oscillating frequency,

• AK is the amplitude of the input triangular oscillation,

• En=i sums the contributions of all neurons to the neuron m,

• C _L is the capacitance load of the neuron input 12 ¹ and corresponds to the capacitor 36 in figure 2,

• C _mn is the capacitance value that implements the synaptic amplitude K _mn , and r — 1, if 0 < X < n

• square(X) = j + 1, if n < X < 2 .

( 0, if e {0, nr} By implementing the neuromorphic circuit 10 with such specific waveforms, its phase dynamics are determined by the compact ordinary differential equation previously defined. Thus, there is no need to perform time-consuming circuit-level simulations to check the neuromorphic circuit functionality, and only solving the ordinary differential equation suffices to obtain the computation result.

This advantage should be added to the other advantages provided by the neuromorphic circuit 10, as shown by the experiments carried out by the Applicant.

Notably, such ONN architecture is mix-signal architecture with analog phase computing and digital propagation. Indeed, oscillators compute in analog domain to enlarge the solution space whereas synaptic signals propagate in a digital way to enable efficient, automated large-scale ONN design and integration in heterogeneous systems. In particular, the neuromorphic circuit 10 is of very-low consumption.

In addition, the decoupling between the input and the output of the neuron unit 12 allows both feedforward and recurrent type of signal propagation. This implies that the neuromorphic circuit 10 can implement a large variety of different neural networks.

This renders such neuromorphic circuit 10 particularly adapted to solving optimization tasks, notably NP-hard problems, for transportation, logistics, machine scheduling, robotics...

As a specific example, the Applicant has shown that the neuromorphic circuit 10 is able to solve efficiently a weighted MAX-CUT problem, that is finding a subset of vertices such that the weights of edges between the subset and the rest of the vertices be maximum.

As other specific examples, it can be cited the resolution of MAX-CLIQUE problem, NP-hard Graph-coloring problem or NP-hard Travelling Salesman problems.

The neuromorphic circuit 10 can also be used for many other applications and notably in image processing. For instance, it has been demonstrated that such neuromorphic circuit 10 can be advantageously used for pattern recognition.

Another application can be the processing of any sensory data in connected devices, the associated sensor being for instance an image sensor, an audio sensor, a temperature sensor or a pressure sensor.