Generation of physical random values ​​using a resistive memory device

Abstract

Generation of physical random values ​​using a resistive memory device. This description relates to a device configured to generate a random value, the device comprising: a resistive memory device (102) capable of being programmed incrementally; and a control circuit (104) configured to: - perform a first programming operation with first programming parameters in order to generate a first conductance variation of a first polarity in the resistive memory device; - perform a second programming operation with second programming parameters in order to generate a second conductance variation of a second polarity, opposite to the first polarity, in the resistive memory device; and - generate the random value by reading a conductance value from the resistive memory device after the first and second programming operations.Figure for the abridged version: Fig. 1.

Description

Title of the invention: Generation of physical random values ​​using a resistive memory device Technical field

[0001] The present description relates generally to the field of electronic devices, and in particular to the generation of physical random values ​​or numbers, as well as to machine learning using such generation of physical random values ​​or numbers. Prior art

[0002] Random number generation is used in various applications, ranging from cryptography to machine learning. Some solutions for generating random numbers are based on physical devices with variable behavior. However, physical devices proposed in the past for random number generation have a difficulty: they tend to lack scalability for applications in which a relatively large number of random numbers must be generated in parallel, and / or they are limited in terms of tuning the parameters of the probability distribution of the generated random numbers, such as the mean and standard deviation of the probability distribution.

[0003] There is therefore a need in the art for a solution for generating physical random values ​​or numbers which at least partially overcomes one or more drawbacks of the prior art. Summary of the invention

[0004] According to one aspect, there is provided a device configured to generate a random value, the device comprising: a resistive memory device capable of being incrementally programmed; and a control circuit configured to: perform a first programming operation of the resistive memory device with the first programming parameters to generate a first conductance variation of a first polarity in the resistive memory device; perform a second programming operation of the resistive memory device with the second programming parameters to generate a second conductance variation of a second polarity in the resistive memory device, the second polarity being the opposite polarity to the first polarity; and generate the random value by reading a conductance value of the resistive memory device following the first and second programming operations.

[0005] According to one embodiment, the control circuit is further configured to perform an analog-to-digital conversion on the read conductance value in order to generate a random digital value.

[0006] According to one embodiment, the amplitude of the first programming parameters is different from the amplitude of the second programming parameters.

[0007] According to one embodiment, the resistive memory device is a domain wall device, and a pulse width, or number of pulses, applied to the domain wall device during the first programming operation is different from a pulse width, or number of pulses, applied to the domain wall device during the second programming operation.

[0008] According to one embodiment, the control circuit is configured to implement a Bayesian model update operation, in which the magnitude of one of the first or second programming operations is based on the magnitude of a Langevin gradient, and the polarity of said one of the first and second programming operations is based on the sign of the Langevin gradient.

[0009] According to one embodiment, the amplitude of said one of the first or second programming operations is proportional to a programming amplitude equal to:

[0010] programming - - (p ( j1V eÂA ( 9t ) j )

[0011] where [Math.5] is a Langevin gradient, [Math.5] is a learning rate, and [Math.5] <P is a function that converts the noise introduced by applying the Langevin gradient into a programming amplitude.

[0012] According to one embodiment, a value of the first programming parameters is identical to a value of the second programming parameters, the control circuit being further configured to perform a third programming operation of the resistive memory device before the first programming operation or after the second programming operation.

[0013] According to one embodiment, the control circuit is configured to implement a Bayesian model update operation, in which the amplitude of said third programming operation is based on the magnitude of a Langevin gradient, and the polarity of the third programming operation is based on the sign of the Langevin gradient.

[0014] According to another aspect, there is provided a method of generating random values, the method comprising: performing, by a control circuit, a first operation of programming a resistive memory device with first programming parameters to generate a first conductance variation of a first polarity in the resistive memory device, the resistive memory device being incrementally programmable; performing, by the control circuit, a second operation of programming the resistive memory device with the second programming parameters to generate a second conductance variation of a second polarity in the resistive memory device, the second polarity being the opposite polarity to the first polarity;and generating the random value by reading, by the control circuit, a conductance value of the resistive memory device following the first and second programming operations.;

[0015] According to one embodiment, the method further comprises performing an analog-to-digital conversion on the read conductance value in order to generate a random digital value.

[0016] According to one embodiment, the amplitude of the first programming parameters is different from the amplitude of the second programming parameters.

[0017] According to one embodiment, the resistive memory device is a domain wall device, and a pulse width, or number of pulses, applied to the domain wall device during the first programming operation is different from a pulse width, or number of pulses, applied to the domain wall device during the second programming operation.

[0018] According to another aspect, there is provided a method of implementing an update operation of a Bayesian model comprising the above method, wherein the magnitude of one of the first or second programming operations is based on the magnitude of a Langevin gradient, and the polarity of said first and second programming operations is based on the sign of the Langevin gradient.

[0019] According to one embodiment, an amplitude of the first programming parameters is the same as an amplitude of the second programming parameters, the method further comprising executing, by the control circuit, a third programming operation of the resistive memory device before the first programming operation or after the second programming operation.

[0020] According to yet another aspect, there is provided a method of implementing an update operation of a Bayesian model comprising the above method, wherein the magnitude of said third programming operation is based on the magnitude of a Langevin gradient, and the polarity of the third programming operation is based on the sign of the Langevin gradient. Brief description of the drawings

[0021] These characteristics and advantages, as well as others, will be explained in detail in the following description of particular embodiments given without limitation in relation to the attached figures among which:

[0022] [Fig.l] schematically illustrates a device for generating random values ​​according to an exemplary embodiment of the present description;

[0023] [Fig.2] shows graphs illustrating an example of conductance variations of a resistive memory device of the device of [Fig.l];

[0024] [Fig.3] is a graph illustrating an example probability distribution for a change in position of a domain wall of the resistive memory device of [Fig.l];

[0025] [Fig.4] is a flowchart illustrating the operations of a method for generating a random value according to an exemplary embodiment of the present disclosure;

[0026] [Fig.5] shows graphs illustrating an example of a probability distribution for a change in position of a domain wall in the case where programming operations of opposite polarities are applied to the resistive memory device;

[0027] [Fig.6] shows graphs illustrating another example of probability distribution for a change in position of a domain wall in the case where programming operations of opposite polarities are applied to the resistive memory device;

[0028] [Fig.7] shows graphs illustrating another example of probability distribution for a change in position of a domain wall in the case where programming operations of opposite polarities, and an additional programming operation, are applied to the resistive memory device;

[0029] [Fig.8] shows graphs illustrating another example of probability distribution for a change in position of a domain wall in the case where programming operations of opposite polarities and different durations are applied to the resistive memory device;

[0030] [Fig.9] is a graph illustrating an example of an SGLD (Stochastic Gradient Langevin Dynamics) algorithm involving random sampling;

[0031] [Fig. 10] schematically illustrates a device capable of machine learning according to an exemplary embodiment of the present disclosure;

[0032] [Fig. 11 A] is a flowchart illustrating the operations of a method of implementing implementing an SGLD model update operation according to an exemplary embodiment of the present description;

[0033] [Fig. 1 IB] represents graphs illustrating an operation of updating the SGLD model according to an exemplary embodiment of the present description;

[0034] [Fig. 12] is a graph representing examples of Langevin gradient and Gaussian samples according to an exemplary embodiment of the present disclosure;

[0035] [Fig. 13] schematically illustrates an array of domain wall resistive memory devices according to an exemplary embodiment of the present disclosure;

[0036] [Fig. 14] is a graph representing probability distributions of discrete levels of a domain wall device;

[0037] [Fig. 15] is a flowchart illustrating operations of an MCMC-based machine learning algorithm according to a specific example in which it is implemented using a domain wall device;

[0038] [Fig. 16] is a flowchart illustrating a method of generating a prediction according to an exemplary embodiment of the present disclosure;

[0039] [Fig. 17] illustrates an example of input and output data of the distribution;

[0040] [Fig. 18] is a graph representing results as a function of the accuracy of a DW device;

[0041] [Fig. 19] is a graph illustrating the effect of table size; and

[0042] [Fig.20] is a graph illustrating the performance of input detection and output of the distribution of a model. Description of the embodiments

[0043] The same elements have been designated by the same references in the different figures. In particular, the structural and / or functional elements common to the different embodiments may have the same references and may have identical structural, dimensional and material properties.

[0044] Unless otherwise specified, when referring to two elements connected between them, it means directly connected without intermediate elements other than conductors, and when referring to two elements connected (in English "coupled") between them, this means that these two elements can be connected or linked through one or more other elements.

[0045] In the following description, when reference is made to absolute position qualifiers, such as the terms "front", "back", "top", "bottom", "left", "right", etc., or relative position qualifiers, such as the terms "above", "below", "upper", "lower", etc., or to orientation qualifiers, such as the terms "horizontal", "vertical", etc., reference is made unless otherwise specified to the orientation of the figures.

[0046] Unless otherwise specified, the expressions "approximately", "approximately", "substantially" and "in the order of" mean within 10%, preferably within 5%.

[0047] [Fig.l] schematically illustrates a random value generation device 100 according to an exemplary embodiment of the present description. The device 100 comprises for example a resistive memory device 102 and a control circuit 104 (CTRL CIRCUIT).

[0048] The resistive memory device 102 is for example of the type that can be programmed incrementally. This means in particular that a programming operation can be used to modify the conductance from an initial conductance value, and that the resulting conductance is a function of the magnitude and polarity of a given parameter of the programming operation and also of the initial conductance value. Examples of resistive memory devices having this property include the following: 1) domain wall (DW) magnetic devices, described for example in the publications: Kumar, Durgesh, et al. "Domain wall memory: Physics, materials, and devices". Physics Reports 958 (2022): 1-35; and Sbiaa, R., and S. N. Piramanayagam. "Multi-level domain wall memory in constricted magnetic nanowires". Applied Physics A 114 (2014): 1347-1351; and 2) phase change memory as described in the publication: La Barbera, Selina, et al. "Narrow Heater Bottom Electrode-Based Phase Change Memory as a Bidirectional Artificial Synapse", Adv. Electron. Mater. 2018, 4, 1800223.

[0049] [Fig.l] illustrates an example in which the device 102 is a DW device, which comprises for example two stacked magnetic layers 106 and 108. A DW device is a non-volatile memory which can be multi-level. The device comprises three electrodes 110, 112, 114, each being for example connected to the control circuit 104.

[0050] Electrode 110 is in contact with the top "pinched" layer 106, which includes only the high magnetic spin, represented by upward pointing arrows in layer 106. For example, electrode 110 is in contact with an upper surface of layer 106.

[0051] The electrodes 112, 114 are in contact with the "free" lower layer 108 and, for example, with the lateral ends of the lower layer 108. The lower layer 108 comprises a high magnetic spin portion, represented by arrows pointing upwards in the layer 108, and a low magnetic spin portion, represented by arrows pointing downwards in the layer 108. A domain wall (DW) corresponds to the interface between the high and low spin portions. In the example of [Fig. 1], the high magnetic spin portion extends from the electrode 112 to the domain wall DW, and the low magnetic spin portion extends from the electrode 114 to the domain wall DW.

[0052] By passing, using the control circuit 104, at least one electric current pulse between the electrodes 112 and 114, in one direction or the other, the position of the domain wall DW between the high magnetic spin and low magnetic spin portions can be moved linearly towards the electrode 112, or towards the electrode 114, symmetrically. In particular, the polarity of this at least one current pulse determines the direction in which the domain wall DW moves. For example, during a write operation, the control circuit 104 is configured to connect the electrode 114 to ground, to leave the electrode 110 floating or at high impedance, and to apply at least one positive or negative write current pulse to the electrode 112.

[0053] By applying, using the control circuit 104, a voltage between one of the electrodes 112, 114 and the upper electrode 110, it is possible to read the current flowing through the DW device, this current depending on the amplitude of the tunnel magnetoresistance (TMR) between the two layers 106, 108. The TMR itself depends on the ratio between the high spin in the upper fixed layer 106 and the high spin in the lower free layer 108. Therefore, by moving the position of the DW domain wall, the conductance of the device can be changed linearly. This allows the device to store data in a non-volatile manner. For example, to perform a read operation, the control circuit 104 is configured to connect the upper electrode 110 to ground, to leave the electrode 114 floating or at high impedance, and to apply a read pulse to the electrode 112.The conductance of the device between the electrode 112 and the upper electrode 110 is then, for example, proportional to the reading current.

[0054] For example, assuming that the domain wall DW is in an initial position represented by a dotted line in FIG. 1, applying a current pulse between the electrodes 112, 114 causes, for example, a displacement AX of the domain wall, and this displacement may be positive or negative depending on the polarity of the current. For example, a positive current pulse applied through the layer 108 from the electrode 112 to the electrode 114 causes the wall to move of DW domain to electrode 114, and a negative current pulse applied through layer 108 from electrode 112 to electrode 114, which is equivalent to a positive current pulse applied from electrode 114 to electrode 112, causes the DW domain wall to move toward electrode 112.

[0055] [Fig.2] is a graph illustrating an example of variations in the conductance (g) of a resistive memory device of the device of [Fig.l] as a function of time t, and in response to current pulses applied to the device.

[0056] For example, the conductance g of the device 102 is initially at a level gl. At a time t1, a negative current pulse p1 is applied to the electrode 112, causing the DW domain wall to move towards the electrode 112, and thus the conductance to decrease to a level g2. At a time t2, a negative current pulse p2, whose amplitude and width are similar to those of the pulse p1, is applied to the electrode 112, which causes the DW domain wall to move again towards the electrode 112, and thus the conductance to decrease to a level g3. At a time t3, a positive current pulse p3 is applied to the electrode 112, causing the DW domain wall to move towards the electrode 114, and the conductance to increase.For example, pulse p3 has a similar amplitude to pulses pl and p2, but twice as wide, and so the conductance increases to a level close to the initial conductance level gl.

[0057] However, the change in position of the DW domain wall in response to a current pulse is subject to various sources of variability. It is therefore more accurate to describe the change in position, and hence the change in conductance, using a probability distribution, as will be described with reference to [Fig.3],

[0058] Figure 3 is a graph illustrating an exemplary probability distribution p(AX) for a change in position AX of the wall of the domain DW of the resistive memory device 102 of Figure 1 in response to a positive current pulse of a given pulse width and amplitude. The distribution may, for example, extend between a minimum displacement A Xmin and a maximum displacement A Xmax and have a single peak, at a displacement A Xpeak.

[0059] It has been found that the shape of the probability distribution is influenced by the width and amplitude of the programming current pulse. For example, for a given intensity of the programming current, the greater the width of the pulse, the greater the width of the probability distribution, in other words, the distance between A Xmin and A Xmax.

[0060] It was also found that the probability distribution is symmetrical with respect to the polarity of the current pulse. In other words, the same distribution probability will be obtained for a negative impulse than for a positive impulse, the difference being that the displacement of the domain wall will be done by a negative rather than a positive shift.

[0061] [Fig. 4] is a flowchart illustrating the operations of a method 400 for generating a random value, centered on a current value, according to an exemplary embodiment of the present description. The method of [Fig. 4] is for example implemented by the control circuit 104 connected to the resistive memory device 102.

[0062] In an operation 401, a first programming operation of the resistive memory device 102 is performed based on first programming parameters to create a conductance variation of a first polarity. The term "polarity" is used in reference to the conductance variation to mean an increase or decrease in conductance. For example, in the case of the domain wall device, the programming parameters define the pulse width, amplitude, and polarity of the current pulse, and in some cases the number of current pulses, applied between the electrodes 112 and 114 of the device, and the conductance variation is a result of the position variation of the domain wall. In other resistive memory technologies, the programming parameters may define other variables, such as the amplitude of the voltage.

[0063] In an operation 402, a second programming operation of the resistive memory device 102 is performed based on second programming parameters to create a conductance change of a second polarity opposite to the first polarity. In other words, if the conductance change of the first polarity is an increase in conductance (corresponding to a decrease in resistance), the conductance change of the second polarity is a decrease in conductance (corresponding to an increase in resistance), and vice versa. In some embodiments, the first and second programming parameters are equal and opposite, such that the probability distributions are also equal and opposite.For example, in the case of a DW device, the first and second programming operations are both based on current pulses of the same width and amplitude, but one of the pulses is positive and the other is negative. Furthermore, as described in more detail below, in some embodiments, there may be an offset between the first and second programming parameters to achieve a desired distortion between the conductance variations.

[0064] In an operation 403, the programmed conductance of the resistive memory device 102 is read to generate a random value. In some embodiments, the read random value is converted from an analog value, such as a

[0065]

[0066]

[0067]

[0068]

[0069] voltage or a current level, into a digital value by an analog-to-digital converter (not shown) which, for example, is part of the control circuit 104. Advantageously, this random value will be a function of a probability distribution having the form of a normal distribution, in other words a Gaussian distribution, with a mean value and a standard deviation which can be controlled independently using the first and second programming parameters, as will be described in more detail with reference to Figures 5 to 8. Figures 5 to 8 illustrate probability distribution graphs resulting from the example of programming a domain wall memory device with a variable width current pulse. It will be obvious to those skilled in the art that similar effects can be achieved using other types of resistive memory devices capable of being incrementally programmed, such as the PCM devices mentioned above. Figure 5 shows graphs illustrating an example of a probability distribution for a change in position of a domain wall in the case where programming operations of opposite polarities are applied to the resistive memory device. Furthermore, the programming parameters are substantially equal and opposite and, in particular, the widths and amplitudes of the current pulses are substantially the same.In particular, Figure 5 represents a graph 502 of the probability density associated with the first programming operation based on a positive current pulse, which causes a positive displacement ( AX + ) of the domain wall, a graph 504 of the probability density associated with the second programming operation based on a negative current pulse, which causes a negative displacement ( AX - ) of the domain wall, the negative pulse having the same pulse width and the same amplitude as the positive current pulse, and a graph 506 representing the resulting probability distribution of the two programming operations, which corresponds to the normal summation law. In particular, this probability distribution p( AX) is centered around zero, which implies that the average displacement is the initial position of the domain wall. For example, the position of the sampled domain wall x can be expressed as: P(X) ~ / OR [Math.l] x + And [Math.l] x. are the means of the normal distributions of the positive and negative displacement of the domain wall, respectively, [Math.l] a(T) is a function giving the standard deviation of the normal depending on the pulse width T, and [Math.l] x0 is the original position of the domain wall.

[0070] By varying the pulse widths T used for the positive and negative pulse pairs, it is possible to determine the standard deviation of the probability distribution used to generate the random number. It is important to note that this standard deviation can be defined independently of the mean of the distribution.

[0071] Furthermore, for small values ​​of T, it is likely that the ratio of the device conductance range divided by the minimum standard deviation achievable with the shortest pulse width will be relatively high. This allows, for example, tighter distributions with a lower standard deviation, which is useful for some applications. For example, according to the SGLD (Stochastic Gradient Langevin Dynamics) algorithm described below, the optimal ratio is generally relatively low.

[0072] Figure 6 shows graphs illustrating another example of a probability distribution for a change in position of a domain wall in the case where programming operations of opposite polarities are applied to the resistive memory device. Graphs 602, 604 and 606 of Figure 6 are similar to graphs 502, 504 and 506 of Figure 5, except that the positive and negative current pulses have a larger pulse width in the example of Figure 6, and therefore the resulting probability distribution p(AX) is also wider, with a mean always equal to zero.

[0073] In some cases, it may be desired that the random value be selected from a probability distribution having a mean shift relative to the initial value. Solutions for generating such a shift are described in relation to Figures 7 and 8.

[0074] Figure 7 shows graphs illustrating another example of a probability distribution for a change in position of a domain wall in the case where programming operations of opposite polarities, and an additional programming operation, are applied to the resistive memory device. In particular, graphs 702 and 704 illustrate the probability distributions for a positive current pulse and a negative current pulse respectively, similar to graphs 502 and 504 of Figure 5. Furthermore, after performing the first and second programming operations, a third programming operation is performed, represented by graph 706, and involving the application of a new current pulse that shifts the resulting probability distribution by an amount delta 5. In the example of Figure 7, the third programming operation is to apply a new negative current pulse. As shown in graph 708, the resulting probability distribution is the sum of the probability distributions of the three graphs 702, 704 and 706 and therefore the offset 5 (OFFSET 5) shifts the probability distribution in the negative direction.

[0075] Figure 8 shows graphs illustrating another example of a probability distribution for a change in position of a domain wall in the case where programming operations of opposite polarities and different durations are applied to the resistive memory device. In the example of Figure 8, only two programming operations are applied, but the pulse widths during the first and second programming operations are different. Figure 8 illustrates in particular an example of a current pulse 802 (PUSH) applied during the first programming operation, and an example of a current pulse 804 (PULL) applied during the second programming operation. Current pulses 802 and 804 have, for example, the same amplitude. Current pulse 802, for example, has a pulse width of e and current pulse 804, for example, has a pulse width of e + L.The part of the pulse width é which is common to both pulses 802 and 804 leads for example to a symmetrical distribution, centered on the current value (for example a normal probability distribution p(£)) represented by graph 806. The additional part L of pulse 804, leads for example to a probability distribution p(L) represented by graph 808, which modifies the overall probability distribution in a similar way to what is illustrated in graph 708 of [Fig.7].

[0076] The possibility of defining the standard deviation of the probability distribution independently of the mean of the distribution makes the solution described here particularly advantageous in Markov chain Monte Carlo machine learning methods. Indeed, the probability distribution is the same regardless of the current position of the domain wall. Examples of application of the described solution to MCMC methods will now be described in more detail with reference to Figures 9 to 15.

[0077] [Fig.9] is a graph illustrating an example of the SGLD algorithm, which involves random sampling. In particular, this algorithm is a type of Markov Chain Monte Carlo (MCMC) sampling approach widely applied in machine learning. It is a probabilistic alternative to approaches involving fully gradient-based maximum likelihood model learning.

[0078] According to gradient-based learning, the algorithm is based on a loss function and the objective is to determine an optimal model that minimizes the loss. One of the main drawbacks of loss-based approaches is that, in the case of low-precision model parameters or parameter updates, the algorithm may not converge to the optimal model. This is often the case when learning using a resistive memory-based model.

[0079] In contrast, according to MCMC learning, the objective is to estimate the probability density of a posterior distribution. In particular, as represented by Figure 9, MCMC learning is based on determining an approximation of the posterior distribution 902 of the probability density (PROBABILITY DENSITY) of the model S. To do this, from an initial model (INITIAL MODEL), a chain of random samples is generated using localized random jumps guided by the Langevin gradient. [Fig.9] illustrates an example of a first random sampling operation in which a calculated weight gradient (WEIGHT GRADIENT) is added to the initial model, and the weight sample (WEIGHT SAMPLE) corresponds to a random sampling around the new model (NEW MODEL).This process is repeated, for example based on an update equation of the SGLD algorithm, until an approximation of the posterior distribution902 has been generated. In some cases, the approximation is further facilitated by rejection steps according to the Metropolis-adjusted Langevin algorithm (MALA).

[0080] An advantage of the approach represented by [Fig.9] is that the resulting a posteriori approximation can be used to make accurate predictions, especially with a small amount of noisy training data, in addition to providing well-calibrated uncertainty estimates related to these predictions.

[0081] [Fig. 10] schematically illustrates a device 1000 capable of Bayesian machine learning according to an exemplary embodiment of the present disclosure. For example, the device 1000 comprises a resistive memory array 1002 (RESISTIVE MEMORY ARRAY) 1002, which is an array of resistive memory devices 102 as described herein. The array 1002 is for example connected to a processing device 1004 (PROCESSING DEVICE), which is configured to control read and write operations on the memory devices. resistive of the matrix 1002. The processing device 1004 is also connected, for example, to another memory 1006 (MEMORY), implemented by one or more non-volatile memory devices, such as a hard disk and / or a Flash memory, and / or by one or more volatile memory devices, such as a RAM (Random Access Memory). For example, the processing device 1004 comprises one or more processors configured to execute instructions stored in the memory 1006 and perform calculations such as calculating the partial derivative of a parameter and converting it to a pulse width, which can be used to program a device. In addition or alternatively, the functions of the processing device 1004 could at least partially be implemented by hardware, for example by one or more ASICs (Application Specific Integrated Circuits)..

[0082] The memory 1006 is also configured, for example, to store a training data set 1008 (TRAINING DATASET) used for training a Bayesian neural network model. In some cases, the memory 1006 may also store a Bayesian neural network model 1010 (BAYESIAN NN MODEL) resulting from the training operation, for example in the case where the resistive memory array 1002 is used to generate the samples for the model, and then an ADC is used to store these samples in the memory 1010. The Bayesian NN model, for example, also stores parameters related to the Bayesian NN, such as activation vectors calculated by applying activation functions to the internally produced currents of the network or the activation mean and standard deviations used to perform batch normalization.

[0083] In some embodiments, the processing device 1004 is also connected to a control interface 1012 (CTRL INTERFACE). For example, after training, the device 1000 is configured to perform inference using the trained Bayesian NN model 1010. The control interface 1012 is, for example, connected to one or more sensors (not shown) and / or one or more actuators (also not shown). The one or more sensors include, for example, one or more image sensors, depth sensors, heat sensors, microphones, or any other type of sensor. For example, the one or more sensors 320 include an image sensor with a linear or two-dimensional array of pixels.The image sensor is, for example, a visible light image sensor, an infrared image sensor, an ultrasonic image sensor, or an image depth sensor, such as a LIDAR (Light Detection And Ranging) image sensor. In this case, the input data samples captured by the sensors and provided to the processing device 1004 are images, and the processing device 1004 is . configured to perform image processing on the images to determine one or more actions to be applied via the actuators. The one or more actuators comprise, for example, a robotic system, such as a robotic arm trained to pull weeds or pick ripe fruit from a tree, an automatic steering or braking system in a vehicle, or an electronic actuator, which is, for example, configured to control the operation of one or more circuits, such as waking up a circuit from sleep mode, entering a circuit from sleep mode, generating text by a circuit, performing a data encoding or decoding operation, etc. In another example, the one or more sensors are indoor and / or outdoor temperature sensors of a heating and / or cooling system of a building, comprising, for example, a heat pump as the primary energy source.In this case, the one or more actuators are, for example, activation circuits that activate heating or cooling systems.

[0084] Although [Fig. 10] illustrates an example in which training and inference are implemented by a single device, in other embodiments, these functions could be implemented by separate devices, with one device being used for training to generate the NN model 1010, and the other device receiving the trained NN model 1010, and performing inference based on that model.

[0085] [Fig. 11 A] is a flowchart illustrating operations of a method 1100 of implementing an SGLD model update operation according to an exemplary embodiment of the present description. The method 1100 is for example implemented by the device 1000 of [Fig. 10].

[0086] According to the embodiments described here, an operation for updating the SGLD model is for example implemented on the basis of the following equation: [OO87] =

[0088] where [Math.2] is the updated model, [Math.2] is the current model, [Math.2] is a Langevin gradient of the model, calculated, for example, from a batch of data with respect to a set of labels. The term [Math.2] is usually the sum of a likelihood term and the model's prior distribution. For a classification problem, the likelihood might be the cross-entropy over a mini-batch of labeled data, and the prior might be defined as a Gaussian distribution. The Langevin gradient is also commonly multiplied by a factor when used in a mini-batch setting, e.g., equal to N / n, where N is the number of data points in a training set and n is the mini-batch size. In practice, these quantities are, for example, calculated in the logarithmic domain, which turns multiplications into sums and avoids numerical underflow problems. The symbol [Math.2] is a learning rate, for example between le-5 and le-3, and [Math.2] is a unit Gaussian sample. In some cases, the learning rate, and thus the standard deviation of Gaussian samples, is reduced exponentially. Alternatively, the standard deviation of the Gaussian sample e is kept at a fixed value or is commanded to move from a lower value to a higher value following a cosine function.

[0089] In an operation 1101, the Langevin gradient ÂA ( ) is for example calculated as the derivative of the following equation:

[0090] aa (^) = rlogp (y fa) - logX^)

[0091] where the hyperparameter [Math.3] allows to scale the likelihood contribution in order to compensate for the noise of stochastic gradients, [Math.3] is the likelihood of the model [Math.3] on the data of a mini-batch [Math.3] of the training dataset, which for example includes mini-batches containing d data points, and [Math.3] logX^) is the prior on the model parameters.

[0092] After operation 1101, a parameter update operation is performed on each parameter of the model based on the Langevin gradient calculated during operation 1101. For example, this operation is performed using three programming steps 1103, 1104 and 1105 performed on each parameter.

[0093] In programming step 1103, a resistive memory device 102 of the array 1002 storing the parameter is programmed based on first programming parameters to create a conductance variation of a first amplitude that is proportional to the Langevin gradient, and of a first polarity that is a function of the sign of the Langevin gradient. For example, in the case of the domain wall device, the amplitude is defined by the pulse width, and the polarity of the current pulse applied between the electrodes 112 and 114 of the device defines the sign.

[0094] In programming step 1104, after programming step 1103, a second programming operation of the resistive memory device 102 is performed based on second programming parameters to create a second conductance variation of a second polarity and a second amplitude. For example, the second polarity may be the same as or opposite to the first polarity.

[0095] In programming step 1105, after programming step 1104, a third programming operation of the resistive memory device 102 is performed based on third programming parameters to create a conductance variation of a third polarity opposite to the second polarity. In other words, if the conductance variation of the second polarity is an increase in conductance, the conductance variation of the third polarity is a decrease in conductance, and vice versa. The second and third programming parameters are for example equal and opposite, so the probability distributions will also be equal and opposite. For example, in the case of a DW device, the second and third programming operations are both based on current pulses of the same width and amplitude, but one of the pulses is positive and the other negative.

[0096]

[0097]

[0098] The second and third programming operations 1104 and 1105 are for example used to apply the sampling noise of the operation of update of the SGLD model. For example, in the case of programming a wall domain memory device with a variable width current pulse, the amplitude of the pulse widths of the second and third programming operations is for example proportional to • In some modes embodiment, it would also be possible to reduce the pulse widths of the second and third programming operations of the noise part of the pulse width of the first programming operation 1103. In other words, assuming that the total noise part corresponds to a pulse width E, if the amplitude of the Langevin gradient part results in noise corresponding to a pulse width of 1, the pulse widths applied during the second and third programming operations will generate a noise contribution equal to (El). This is illustrated in more detail in [Fig.l IB]. [Fig. 11B] represents graphs illustrating an operation of updating the SGLD model according to an exemplary embodiment of the present description. The principle of SGLD is illustrated by the left-hand side graph of Figure 1 IB showing the probability density p(g) of a model g, where the samples (crosses) are generated by moving in the direction of a Langevin gradient (arrow L) and then perturbing this update with a sample from a unitary noise distribution (distribution fe). The right-hand side graphs of Figure 1 IB illustrate how this update can be achieved by a pair of push-pull current pulses in antiphase ("push", "pull"), which provide a sample from the distribution £ as well as the Langevin gradient V ^^( 1¾ ) • In particular, in the example of Figure 1 IB, the "push" current pulse implements the second programming operation 1104, and the "pull" pulse implements the first and third programming operations 1103, 1105 in the form of a single "pull" current pulse. For example, the pulse duration TpUSh of the push current pulse is equal to: [00991 ^ = ^-^(5(7^(9,)1)

[0100] where [Math.4] <p is a function that converts the noise introduced by applying the Langevin gradient into a pulse width, and this part [Math.4] I ) represents the quantity 1 by which the pulse widths can be shortened. The pulse duration [Math.4] TpuU of the traction current pulse is for example equal to [Math.4] T pull = T push +TL Or [Math.4] Tl is proportional to the Lange gradient [Math.4]

[0101] More generally, whether or not the programming operations are performed by current pulses or by other means, one of the programming operations has, for example, an amplitude proportional to a programming amplitude equal to:

[0102] programming = ^2^ - (p ( j | V pÂA ( Qt ) | )

[0103] where [Math.5] <p is now a function that converts the noise introduced by applying the Langevin gradient into a corresponding programming amplitude.

[0104] Referring again to [Fig. 11 A], in some embodiments, a random value is generated by reading the programmed conductance of the resistive memory device 102 following the programming operation 1105. Further, this random value is, for example, converted from an analog value, such as a voltage or current level, to a digital value by an analog-to-digital converter (not shown) that is, for example, part of the control circuit 104.

[0105] In alternative embodiments, it would be possible to perform the first programming operation 1103 after the second and third programming operations 1104, 1105, or after the second programming operation 1104 and before the third programming operation 1105 and / or to combine the first programming operation with the second or third programming operation.

[0106] An advantage of the method of Figure 1 IA for performing SGLD is that any lack of precision available in programming a resistive memory device, such as a DW device, is effectively drowned out in the processing of random sampling or even absorbed in the noise generated by the two programming operations, which are equal and opposite. For example, assuming the learning factor is less than 1, which is usually the case, the square root operation ensures that Gaussian noise contributes more to the update on average than the Langevin gradient. This phenomenon is illustrated in [Fig. 12].

[0107] Figure 12 is a graph showing examples of the Langevin gradient term VyÂ^( (¾) and Gaussian samples according to an exemplary embodiment of the present disclosure. In particular, [Fig. 12] illustrates examples based on a typical SGLD update step, and the weight update histogram for all synapses in a layer is shown. It can be observed that, in most cases, the noise from the Gaussian sampling dominates the deterministic updates, which is typically about 95% of the cases. For the remaining cases, typically about 5% of the cases, although the Langevin update is larger than the Gaussian noise update, because the gradient is actually so large, it is easy to get the update step in the right direction despite the variability associated with the device programming.

[0108] [Fig. 13] schematically illustrates an array 1300 of domain wall resistive memory devices 102 according to an exemplary embodiment of the present disclosure. This DW memory array 1300 is for example part of the resistive memory array 1002 of [Fig. 10].

[0109] The DW memory array 1300 comprises M columns (M COLUMNS), M being equal to 4 in the example of [Fig. 13], and N rows (N ROWS), N being equal to 3 in the example of [Fig. 13]. In some embodiments, each row of the DW memory array 1300 acts as a sample of the parameters of a neuron in a Bayesian neural network. A fully connected neural network layer may be composed of a number of such Bayesian neurons in a layer, and then of several such layers. Each Bayesian neuron may be stored and sampled using a separate array or several Bayesian neurons may share the same physical array in which the different neurons use different row index ranges to store their samples. Alternatively, multiple matrices such as matrix 1300 can be used as multiple samples of a full layer in the neural network model.

[0110] Each device 102 is part of a crosspoint cell in the matrix 1300 which also includes a transistor 1302 and a transistor 1304.

[0111] An input voltage line Vin[0] to Vin[Ml] is for example associated with each column, and is for example connected via a control line corresponding to the electrode 112 of each of the devices 102 of its corresponding column. The electrode 114 of each of the devices 102 of a given column is for example connected, via the main conductive nodes of its corresponding transistor 1302, to a programming voltage Vprg[0] to Vprg[Ml] supplied on a corresponding column line by a pulse generation unit (PULSE GEN). The gates of the transistors 1302 of each line are for example connected to a corresponding programming line selection signal Vsp[0] to Vsp[Nl] supplied on a corresponding control line.The upper electrode 110 of each of the devices 102 of a given row is for example connected, via the main conductive nodes of its corresponding transistor 1304 and the main conductive nodes of a column output transistor 1306, to a corresponding row current output line providing a corresponding output current iout[0] to iout[Nl]. The gates of the transistors 1304 of a given column are for example connected to a corresponding column read selection signal Vsr[0] to Vsr[Ml] provided on a corresponding control line. The gates of the transistors 1306 are for example controlled by a corresponding N selection signal Vsn[0] to Vsn[Nl] provided on a corresponding control line.

[0112] The example of [Fig. 13] will be described assuming unsigned parameters, i.e., a single positive conductance per parameter, and a matrix configuration that does not support multiplication and accumulation. In the case of signed parameters, i.e., the subtraction of two positive conductances per parameter, two such matrices could be used and the currents produced by each could be subtracted, for example using a simple circuit based on two stacked current mirrors. The skilled person will understand how these principles can be extended to signed crosspoints, with a pair of devices and matrices capable of MAC.

[0113] In the case of the Bayesian neuron where a matrix stores the samples of its input weights, each column of the matrix 1300 stores for example all the samples of a parameter, N samples being for example generated per parameter. In read mode, the voltage Vin is for example applied to one of the columns. This voltage corresponds for example to a characteristic of input data. The signal of read selection Vsr is applied to the gates of the 1304 transistors in this column, which leads to the output of a current, iout, on the right side of the matrix and corresponding to the multiplication of the input data value by each of the samples of a parameter. By iterating over all columns, it is possible to access the stored samples of all N rows. It would also be possible to read a device from one row at a time - this is what would be done during the model training phase. This can for example be achieved by selecting only one of the rows using the N selection signal Vsn.

[0114] During the writing phase, i.e. the learning of the model by the SGLD algorithm, the pulse generation unit is for example connected to one of the N lines corresponding to the current sample by setting the appropriate programming selection signal Vsp. In this case, the input voltage Vin serves as a reference voltage or virtual ground. The pulse generation unit is for example configured to calculate, for each of the M parameters, the programming pulse sequence to be applied, where the width and polarity of the pulses determine for example to what extent the domain wall will be displaced. In some embodiments, the pulses are applied in parallel for all the M parameters of the network line being programmed. All other selection signals are grounded.

[0115] In some embodiments, an on-the-fly sample thinning procedure is possible. According to this procedure, instead of advancing one row at a time, many samples are produced in each row before moving to the next. Therefore, only the last sample of each row is stored in the table, the others having been "thinned" on-the-fly. When moving from one row to the next, a calibration step is for example performed in order to bring the devices in row n as close as possible to the corresponding devices in row n-1. This can be done, for example, by reading a device in row n-1 and applying a pulse to the device in row n to reduce the difference between their conductances, and repeating these steps until the conductance difference is within an acceptable tolerance.This may introduce a slight error between the last sample of row n-1 and the initial sample of row n. However, in practice, this perturbation has proven beneficial for the performance of the SGLD-trained model, as it facilitates better exploration of the model a posteriori, especially if it has many modes.

[0116] In some embodiments, for the DW-SGLD to operate continuously, instead of stopping when the process reaches the last row of the table, it is possible to return to the first row, which allows the sample that was previously stored in the first row. This solution is interesting from the point of view of continuous learning and changes in the distribution of domains.

[0117] Figure 14 is a graph representing the probability distributions (Probability, p( A$T, P)) of discrete domain-wall changes (AX (pm)) of a domain-wall device. In the example of [Fig. 14], a case is illustrated in which each domain-wall device 102 can be programmed to have one of 10 distinguishable discrete levels with positive and negative pulses of durations T = 10, 20, 30, 40 and 50 ns, and it is therefore possible to generate at least 14 non-overlapping Gaussians. In practice, it has been found that the domain wall device 102 can be programmed to have at least one of 256 distinguishable discrete levels, thereby generating at least 256 non-overlapping Gaussians - corresponding to an 8-bit memory element.

[0118] [Fig. 15] is a flowchart illustrating operations of an MCMC-based machine learning algorithm 1500 according to a specific example in which it is implemented using the device 1000 of [Fig. 10] and the domain wall matrix 1300 of [Fig. 13].

[0119] In an operation 1501, a variable n is for example set to 0, and a count value C is set to an initial value thinning_ratio, equal for example to 256 or another value. The thinning ratio is used to reduce memory requirements and improve the "mixing" of the chain. In particular, instead of storing all the samples, thinning is carried out by storing, for example, every 1000 samples.

[0120] In an operation 1502, for example, for each epoch e among one or more epochs, and for example, for each batch b among one or more batches, an operation 1503 is implemented. Operation 1503 consists of performing, for each data point d of the batch, a series of operations 1504 to 1511.

[0121] Operation 1504 involves, for example, for each column c of the set of columns of the resistive memory array, activating the selection signal N Vsn[n] and the read selection signal Vsr[c], setting Iout[n]=Vin[c]*G[n,c], where G[n,c] is the single device selected for reading that may have been previously written. The output current Iout[n,c] is, for example, read by a DAC and other peripheral circuits known to those skilled in the art of resistive memory array design.

[0122] Operation 1505 consists for example in calculating the Langevin gradient ÀA( 8t ) for G[n].

[0123] Operation 1506 consists, for example, of converting the gradients into pulse widths T(ÀA( ) and into pulse polarities sign(V^L).

[0124] Operation 1507 consists for example in deactivating the selection signal n Vsn[n] and the read selection signal Vsr[c],

[0125] Operation 1508 consists, for example, of activating the programming selection signal Vsp[n] and applying the calculated pulses to the M devices of the row of the matrix in parallel.

[0126] Operation 1509 consists, for example, of applying the same pair of pulses of positive and negative polarity T(°) to each of the M devices.

[0127] Operation 1510 consists, for example, of decrementing the counting value C (C=C-1).

[0128] Operation 1511 consists, for example, of determining whether C is equal to 0 (C==0) and, if so, whether n is equal to the maximum number of rows MAX_R0WS, in other words, if it is the last row of the matrix, the variable n is set to zero, otherwise n is incremented and C is reset to thinning_ratio, and a calibration of row n is performed. For example, if one moves from row n to row n+1, a device in row n+1 will be read by activating the corresponding Vsr column. Based on the difference between the conductances of these two devices, a pulse width, or a number of fixed pulses, is calculated to move the device from row n+1 to that of row n. This pulse is, for example, applied by selecting the correct row using the Vsp signal, then applying the calculated pulse to the device.This operation is performed for example for each device until a maximum number of iterations has been reached or the conductance difference is within a tolerable error margin. Another solution is to read all the devices in a row and then program them in parallel.

[0129] [Fig. 16] is a flowchart illustrating a method 1600 of generating a prediction implemented using the device 1000 of [Fig. 10] and the domain wall matrix 1300 of [Fig. 13], according to an exemplary embodiment of the present disclosure.

[0130] In an operation 1601, all the selection signals n Vsn are for example activated and all the selection programming signals Vsp are for example deactivated.

[0131] In an operation 1602 following operation 1701, for each column c, the read selection signal Vsr[c] is for example activated, an output current Iout[:] is for example defined as Vin[c]*G[:c], and the output current Iout[:] is for example stored, for example after an analog-digital conversion. This value is for example the prediction distribution of a parameter.

[0132] In some embodiments, operation 1602 further comprises a step during which the values ​​of the output currents lout are added, for example using an accumulator. This sum is then, for example, processed by a activation in order to calculate the activation vectors sequentially for each neuron in a neural network and then calculate the outputs.

[0133] In an operation 1603, statistics are for example generated, such as distribution means (for predictions) and entropy.

[0134] In the case where the model is a single neuron, for example a logistic regression model, the statistics are for example calculated directly in iout. In other cases, one evaluates for example how the currents add up to form activation vectors and how these activation vectors propagate forward in the model, for example by generating voltages for other networks that store the parameters of these subsequent neurons and layers. For example, the entropy of the summed predictions of each sample is used to quantify the uncertainty of the model.

[0135] Figures 17-20 illustrate the results obtained using the method described herein. In particular, the described algorithm was implemented using a device model that limits the conductance range and models Langevin updates as Gaussian samples within a given range of bit precision, i.e., two 4-bit devices means that an update can land anywhere in a Gaussian distribution that is 1 / 32 of the conductance range (i.e., at three standard deviations, the value is 1 / 32 of the conductance range).The model is evaluated in terms of MNIST (Mixed National Institute of Standards and Technology) and KMNIST out-of-distribution detection on its ability to correctly classify unseen MNIST digits and to detect the presence of digits that do not correspond to the observed (KMNIST) classes in its training set - an example of data with and without distribution is shown in [Fig. 17].

[0136] [Fig. 18] shows the results (result - "SCORE / 1.0" vs. number of bits per device - "NO. BITS PER DEVICE") when the DW device precision increases from 3 bits to 11 bits, and quantifies the impact of Langevin update precision. The results show the impact on precision (curve 1802) and out-of-distribution detection (curve 1804) when the DW device precision increases from 11 bits to 3 bits. Above 3 bits, the impact on precision is negligible, while the lower precision actually seems to favor the out-of-distribution error detection capabilities. The 1806 region of the curve shows where not skipping gradient updates (i.e., not programming the device if the update is below a certain threshold) was important for sampler convergence.

[0137] [Fig. 19] shows the result ("SCORE / 1.0") as a function of the number of samples ("NO. SAMPLES"), and shows the effect of the matrix size (i.e. the number of samples stored - corresponding either to the number of rows, or to the number of matrices if one matrix per layer sample is used) for a ratio thinning ratio of 256 (curve 1902) and a thinning ratio of 50 (curve 1904). The diminishing benefit of a larger number of samples is clearly visible, as is the importance of large on-the-fly thinning ratios for the model's ability to detect out-of-distribution data. Accuracy is not shown here, but the effect of sample number did not have a significant impact beyond 8 samples.

[0138] [Fig.20] shows sensitivity and specificity curves representing the accuracy of detection inside and outside the distribution (IN DIST. ACC. / OUT DIST. ACC.), as the entropy threshold is increased, of the model with 128 samples and 4-bit precision Langevin updates compared to a deterministic model with floating-point precision. Curve 2002 represents a Bayesian EPI (epistemic uncertainty) neural network, while curve 2004 represents a Det (deterministic model) neural network. Here, a comparison is made for a 32-bit floating-point precision model, which demonstrates that the SGLD model, even at 4 bits, is significantly better at recognizing out-of-distribution data than a deterministic model trained by stochastic gradient descent.A point 2006 on the 2006 curve represents an AUC reading LAMA = 0.974, where AUC stands for "Area Under Curve," and a point 2108 on the 2002 curve represents an AUC reading det = 0.915. The epistemic uncertainty of each unseen data point is calculated and compared to a threshold value. This curve is generated by scanning values ​​for the threshold value, then by summing the area under the curve, it is possible to determine the AUC_ number.

[0139] An advantage of the embodiments described herein is that the solution based on resistive memory devices for the generation of random values ​​is easily scalable for applications in which a relatively large number of random numbers must be generated in parallel. Furthermore, it is possible to adjust the parameters of the probability distribution of the generated random numbers, such as the mean and the standard deviation of the probability distribution, and in particular, it is possible to decorrelate the adjustment of the standard deviation of the probability distribution from the mean of the probability distribution.

[0140] Various embodiments and variations have been described. Those skilled in the art will understand that certain features of these various embodiments and variations could be combined, and other variations will occur to those skilled in the art. For example, domain wall updates could be performed based on pulse counts instead of pulse width. In this case, a fixed pulse is for example applied, with a certain polarity, a certain number of times and a very large number of array organizations and read and write strategies could be used.

[0141] Furthermore, thinning could be achieved in a different way than sampling and counting. For example, a metric combining the sample likelihood and the Euclidean distance between the internal representations of the model, i.e., at the penultimate layer of the model, could be used to reduce the total number of samples to be stored.

[0142] Furthermore, instead of using SGLD, one could use the Metropolis-adjusted Langevin algorithm (MALA), according to which an acceptance / rejection step is added after each sample. This acceptance is for example calculated based on the product of the ratios of the likelihoods and priorities of the proposed and previously accepted models. However, in MALA, sampling is done each time from the same starting point, while some resistive memory technologies, such as DW devices, do not offer a way to store this value, which will therefore be forgotten after the first generation of random numbers, and the more the model is sampled after rejection steps, the further it will move away from the starting point.Therefore, a mitigation strategy could be to limit the number of samples where, for example, the probability threshold is reduced to a smaller number as the number of rejections increases.

[0143] It should also be noted that computing the prior is a relatively expensive process since it requires reading each parameter at each mini-batch. It can also, in some cases, lead to a high rejection rate which can negatively impact the efficiency of the algorithm. In any case, the physical limits of conductance naturally dictate that a prior on the parameters have uniform truncation. In the case of such a uniform prior, since all parameters are equally likely, there is no need to compute the prior. Therefore, in some embodiments, it may be preferable to use a uniform prior (in fact no prior) on the model if this is not important for the task.

[0144] Furthermore, non-symmetrical pulse pairs could be applied to account for non-linear changes in conductance, which can occur in technologies such as PCM, and by providing asymmetry in this way, it is ensured that the noise distribution is symmetrical and centered on the current value.

[0145] Finally, the practical implementation of the embodiments and variants described is within the reach of those skilled in the art from the functional indications given above.

Claims

Claims

1. A device configured to generate a random value, the device comprising: a resistive memory device (102) capable of being incrementally programmed; and a control circuit (104) configured to: - perform a first programming operation of the resistive memory device with the first programming parameters to generate a first conductance variation of a first polarity in the resistive memory device; - perform a second programming operation of the resistive memory device with the second programming parameters to generate a second conductance variation of a second polarity in the resistive memory device, the second polarity being the opposite polarity to the first polarity; and - generate the random value by reading a conductance value of the resistive memory device after the first and second programming operations.

2. The device of claim 1, wherein the control circuit (104) is further configured to perform an analog-to-digital conversion on the read conductance value to generate a random digital value.

3. A device according to claim 1 or 2, wherein an amplitude of the first programming parameters is different from an amplitude of the second programming parameters.

4. The device of claim 3, wherein the resistive memory device is a domain wall device, and a pulse width, or number of pulses, applied to the domain wall device during the first programming operation is different from a pulse width, or number of pulses, applied to the domain wall device during the second programming operation.

5. The device of claim 3 or 4, wherein the control circuit (104) is configured to implement a Bayesian model update operation, wherein the magnitude of one of the first or second programming operations is based on the magnitude of a Langevin gradient, and the polarity of one of the first and second programming operations is based on the sign of the Langevin gradient.

6. A device according to claim 5, wherein the amplitude of said one of the first or second programming operations is proportional to a programming amplitude equal to: Programming amplitude = - ^ ( y | V ( 0t ) where [Math.5] is a Langevin gradient, [Math.5] is a learning rate, and [Math.5] <P est une fonction qui convertit le bruit introduit par l'application du gradient de Langevin en une amplitude de programmation.

7. The device of claim 1 or 2, wherein an amplitude of the first programming parameters is the same as an amplitude of the second programming parameters, the control circuit (104) being further configured to perform a third programming operation of the resistive memory device before the first programming operation or after the second programming operation.

8. The device of claim 7, wherein the control circuit (104) is configured to implement a Bayesian model update operation, wherein the magnitude of said third programming operation is based on the magnitude of a Langevin gradient, and the polarity of the third programming operation is based on the sign of the Langevin gradient.

9. A method of generating random values, comprising: - performing, by a control circuit (104), a first operation of programming a resistive memory device (102) with first programming parameters in order to generate a first conductance variation of a first polarity in the resistive memory device, the device resistive memory device being capable of being incrementally programmed; - performing, by the control circuit (104), a second programming operation of the resistive memory device (102) with the second programming parameters in order to generate a second conductance variation of a second polarity in the resistive memory device, the second polarity being the polarity opposite to the first polarity; and - generating the random value by reading, by the control circuit (104), a conductance value of the resistive memory device following the first and second programming operations.

10. The method of claim 9, further comprising performing an analog-to-digital conversion on the read conductance value to generate a random digital value.

11. The method of claim 9 or 10, wherein an amplitude of the first programming parameters is different from an amplitude of the second programming parameters.

12. The method of claim 11, wherein the resistive memory device is a domain wall device, and a pulse width, or number of pulses, applied to the domain wall device during the first programming operation is different from a pulse width, or number of pulses, applied to the domain wall device during the second programming operation.

13. A method of performing an update operation of a Bayesian model comprising the method of claim 11 or 12, wherein the magnitude of one of the first or second programming operations is based on the magnitude of a Langevin gradient, and the polarity of one of the first and second programming operations is based on the sign of the Langevin gradient.

14. The method of claim 9 or 10, wherein an amplitude of the first programming parameters is the same as an amplitude of the second programming parameters, the method further comprising performing, by the control circuit (104), a third programming operation of the device. resistive memory before the first programming operation or after the second programming operation.

15. A method of implementing an update operation of a Bayesian model comprising the method of claim 14, wherein the magnitude of said third programming operation is based on the magnitude of a Langevin gradient, and the polarity of the third programming operation is based on the sign of the Langevin gradient.

Images