WO2019097494A1 - Synaptic weight transfer between conductance pairs with polarity inversion for reducing fixed device asymmetries - Google Patents
Synaptic weight transfer between conductance pairs with polarity inversion for reducing fixed device asymmetries Download PDFInfo
- Publication number
- WO2019097494A1 WO2019097494A1 PCT/IB2018/059102 IB2018059102W WO2019097494A1 WO 2019097494 A1 WO2019097494 A1 WO 2019097494A1 IB 2018059102 W IB2018059102 W IB 2018059102W WO 2019097494 A1 WO2019097494 A1 WO 2019097494A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- conductance
- ann
- pair
- significance
- neurons
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
- G06N3/065—Analogue means
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0499—Feedforward networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
Definitions
- the present invention relates generally to the field of Artificial Neural Networks (ANNs). More specifically, the present invention is related to a system and method for constructing synaptic weights for artificial neural networks from signed analog conductance-pairs of varying significance, in which the polarity of lower- significance conductance-pairs is periodically inverted.
- ANNs Artificial Neural Networks
- ANNs Artificial Neural Networks
- ANNs are distributed computing systems, which consist of a number of neurons interconnected through connection points called synapses. Each synapse encodes the strength of the connection between the output of one neuron and the input of another. The output of each neuron is determined by the aggregate input received from other neurons that are connected to it, and thus by the outputs of these "upstream” connected neurons and the strength of the connections as determined by the synaptic weights.
- the ANN is trained to solve a specific problem (e.g., pattern recognition) by adjusting the weights of the synapses such that a particular class of inputs produce a desired output.
- the weight adjustment procedure is known as "learning.”
- learning There are many algorithms in the ANN literature for performing learning that are suitable for various tasks such as image recognition, speech recognition, language processing, etc. Ideally, these algorithms lead to a pattern of synaptic weights that, during the learning process, converges toward an optimal solution of the given problem.
- CMOS complementary metal-oxide-semiconductor
- RRAM resistive random access memory
- PCM phase-change memory
- the conductances of the NVM elements are programmed by sending them pulses that can either increase or decrease the conductance according to a learning rule.
- One common learning rule investigated is backpropagation (see Rumelhart et.al., "Learning Internal Representations by Error Propagation,” Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Vol 1 , 1986), which is used extensively in deep learning networks that are currently being implemented on graphical processing units (GPU's) for image recognition, learning to play video games, etc.
- Any real NVM element has a non-ideal response. It is nonlinear and has a limit to the maximum conductance it can achieve.
- the conductance change to a pulse designed to increase conductance is different from that to a pulse designed to decrease conductance, i.e., the response is asymmetric.
- some devices will be inoperable, either stuck in a high conductance state or stuck in a low conductance state. Our work has shown that many of these defects cause very little decrease in ANN
- the backpropagation learning rule is used to update the NVM conductances after each (or after some small number of inputs, called a minibatch).
- Some weights in the network tend to evolve steadily toward some stable value, while others tend to dither up and down, sometimes increasing, other times decreasing.
- the NVM response is nonlinear or asymmetric, the response to a pulse intended to decrease the weight value will be stronger than one intended to increase the weights. This tends to push many of these weights towards zero, making the backpropagation learning rule ineffective and decreasing network performance.
- Embodiments of the present invention are an improvement over both prior art and prior invention systems and methods.
- the present invention provides a method to mitigate undesired contributions to weight induced by fixed device asymmetry in an artificial neural network (ANN), a synaptic weight of each synapse in the ANN being represented by a weighted current flow from multiple conductance-pairs, where a plurality of training examples are serially input to the ANN while observing its output, the method comprising: during a first transfer interval representing a first training cycle, transferring a portion of the synaptic weight from a conductance- pair of lower significance, (g - (f hared ), to a conductance-pair of higher significance, (G + - G ); and during a second transfer interval after the first transfer interval and representing a second training cycle, transferring a portion of the synaptic weight from the conductance-pair of lower significance that is inverted in polarity, -( g - (f hared ), to a conductance-pair of higher significance, (G + - G ), wherein fixed device
- the present invention provides a method implemented in an artificial neural network (ANN), the ANN comprising a plurality of neurons arranged in layers with the outputs of one layer connected to the inputs of a plurality of the neurons of a subsequent layer, where neurons are connected to each other via a plurality of synapses, each of the synapses having a synaptic weight encoding a connection strength between two connected neurons, the magnitude of the synaptic weight of each synapse being represented by a weighted current flow from multiple conductance-pairs, where a plurality of training examples are serially input to the ANN while observing its output, where a backpropagation algorithm updates the synaptic weight in response to the difference between the output from a given layer and a desired output from the given layer, the method comprising: (a) periodically transferring a portion of the synaptic weight from a conductance-pair of lower significance to a conductance-pair of higher significance, such that the total synaptic weight remains substantially
- the present invention provides a method implemented in an artificial neural network (ANN), the ANN comprising a plurality of neurons arranged in layers with the outputs of one layer connected to the inputs of a plurality of the neurons of a subsequent layer, where neurons are connected to each other via a plurality of synapses, each of the synapses having a synaptic weight encoding a connection strength between two connected neurons, the magnitude of the weight of each of the synapses represented by a weighted current flow from multiple conductance pairs, each of the multiple conductance pairs representing a joint contribution and having a higher-significance conductance-pair and a lower-significance conductance-pair, where a plurality of training examples are serially input to the ANN while observing its output, where a backpropagation algorithm updates the synaptic weight in response to a difference between the output from a given layer and a desired output from the given layer, the method comprising: (a) pausing training and measuring conductances across
- the present invention provides a method implemented in an artificial neural network (ANN), the ANN comprising a plurality of neurons arranged in layers with the outputs of one layer connected to the inputs of a plurality of the neurons of a subsequent layer, where neurons are connected to each other via a plurality of synapses, each of the synapses having a synaptic weight encoding a connection strength between two connected neurons, the magnitude of the weight of each of the synapses represented by a weight current flow from multiple conductance pairs, each of the multiple conductance pairs representing a joint contribution and having a higher-significance conductance-pair and a lower-significance conductance-pair, where a plurality of training examples are serially input to the ANN while observing its output, where a backpropagation algorithm updates the synaptic weight in response to a difference between the output from a given layer and a desired output from the given layer, the method comprising: (a) pausing training and measuring conductances across analog memory elements
- a method implemented in an artificial neural network comprising a plurality of neurons arranged in layers with the outputs of one layer connected to the inputs of a plurality of the neurons of a subsequent layer, where neurons are connected to each other via a plurality of synapses, each of the synapses having a synaptic weight encoding a connection strength between two connected neurons, the magnitude of the weight of each of the synapses represented by a weighted current flow from multiple conductance pairs, each of the multiple conductance pairs representing a joint contribution and having a higher-significance conductance-pair and a lower-significance conductance-pair, where a plurality of training examples are serially input to the ANN while observing its output, where a backpropagation algorithm updates the synaptic weight in response to a difference between the output from a given layer and a desired output from the given layer, the method comprising: (a) pausing training and measuring conductances across analog
- a system implemented in an artificial neural network comprising a plurality of neurons arranged in layers with the outputs of one layer connected to the inputs of a plurality of the neurons of a subsequent layer, where neurons are connected to each other via a plurality of synapses, each of the synapses having a synaptic weight encoding a connection strength between two connected neurons, the magnitude of the synaptic weight of each synapse being represented by a weighted current flow from multiple conductance-pairs, where a plurality of training examples are serially input to the ANN while observing its output, where a backpropagation algorithm updates the synaptic weight in response to the difference between the output from a given layer and a desired output from the given layer, the system operable to perform a method comprising: periodically transferring a portion of the synaptic weight from a conductance-pair of lower significance to a conductance-pair of higher significance, such that the total synapt
- a system implemented in an artificial neural network comprising a plurality of neurons arranged in layers with the outputs of one layer connected to the inputs of a plurality of the neurons of a subsequent layer, where neurons are connected to each other via a plurality of synapses, each of the synapses having a synaptic weight encoding a connection strength between two connected neurons, the magnitude of the weight of each of the synapses represented by a weighted current flow from multiple conductance pairs, each of the multiple conductance pairs representing a joint contribution and having a higher-significance conductance-pair and a lower-significance conductance-pair, where a plurality of training examples are serially input to the ANN while observing its output, where a backpropagation algorithm updates the synaptic weight in response to a difference between the output from a given layer and a desired output from the given layer, the system operable to perform a method comprising: (a) pausing training
- a system implemented in an artificial neural network comprising a plurality of neurons arranged in layers with the outputs of one layer connected to the inputs of a plurality of the neurons of a subsequent layer, where neurons are connected to each other via a plurality of synapses, each of the synapses having a synaptic weight encoding a connection strength between two connected neurons, the magnitude of the weight of each of the synapses represented by a weighted current flow from multiple conductance pairs, each of the multiple conductance pairs representing a joint contribution and having a higher-significance conductance-pair and a lower-significance conductance-pair, where a plurality of training examples are serially input to the ANN while observing its output, where a backpropagation algorithm updates the synaptic weight in response to a difference between the output from a given layer and a desired output from the given layer, the system operable to perform a method comprising: (a) pausing training
- FIG. 1 illustrates a prior art array with one conductance-pair per synapse.
- FIG. 2 depicts, in accordance with one embodiment, an array having more than one conductance-pair per synapse, with one set of pairs having its read-current amplified with respect to the other by a factor of F.
- FIG. 3 depicts, in accordance with one embodiment, an array having more than one conductance-pair per synapse, with one set of pairs having its read-current amplified with respect to the other by a factor of F, in which the polarity of the lower-significance pairs has been inverted.
- FIG. 4 depicts, in accordance with one embodiment, simulation results showing benefits of the technique of periodically inverting the polarity of a lower-significance conductance pair.
- FIG. 5 shows, in accordance with one embodiment, a situation immediately after transfer of synapses in the same column as neuron M-i, which is sharing to the left and has just had the polarity of its g value inverted (g subtracts from the net weight), while neuron M m continues to have its g value add to the net weight, and shares to a g shared device located to its right.
- FIG. 6 shows, in accordance with one embodiment, a situation after transfer of column M m , with its g value now also inverted, and the shared bias read current from the (f hared device now being delivered from the shared device to the left which was recently reset to its initial value (at the same instant it gained its first shared column).
- the invention described here builds upon an earlier disclosure of a technique for mitigating the effects of nonlinearity and asymmetry in conductance response by distributing the synaptic weight across multiple pairs of conductances, each of varying significance.
- Each pair contributes in the usual way towards synaptic weight through the difference between an excitatory conductance, G+, and an inhibitory conductance, G-.
- the contribution of each pair varies in the amplification factor implemented at the end of the bit lines where these pairs are summed.
- Example-by-example programming can be implemented in many ways - in one manifestation, the least-significant conductance pair is updated during network training.
- one or more of the less-significant conductance pairs could be comprised of volatile analog memory elements, such as a capacitor tied to the gate of a read transistor. In this case, adding or subtracting charge to the capacitor changes the gate voltage of the read transistor and thus its effective conductance.
- one of the conductances in each conductance pair could be shared across many different synapses, with all individual synaptic programming taking place by adjusting the non-shared conductance.
- FIG. 2 shows the previously disclosed array (i.e., disclosed in USPTO Serial No. 15/660,945) having more than one conductance-pair per synapse, with one set of pairs having its read-current amplified with respect to the other by a factor of F.
- each synaptic weight is comprised of two pairs of conductances.
- the first pair contributes to the summed current of its downstream neuron at the bottom of the array with a gain factor of F with respect to the second pair.
- this technique increases the total dynamic range of the overall synapse, and allows a fast but volatile analog memory element exhibiting more linear and symmetric conductance change to be used as the second or lower-significance conductance pair for ANN example-by- example programming, with the accumulated synaptic weight change occasionally transferred carefully into the first conductance pair.
- the weight update of the lower-significance pair be as symmetric as possible, so that the network can request many weight increases, and almost but not quite as many weight decreases, and have these opposite-sign contributions cancel correctly.
- CMOS-based volatile conductance a transistor with its gate tied to a capacitor. Charge can be added and subtracted to the capacitor in a quite linear way, using a pair of transistors.
- CMOS fabrication variabilities will cause each transistor on a wafer - including both members of this pair of charge/discharge transistors - to vary in threshold voltage, and thus in the amount of charge they add or subtract to their capacitor. This can re-introduce a strong fixed asymmetry between the size of a conductance increase and the size of a conductance decrease.
- PCM Phase-Change Memory
- NVM Non-Volatile Memory
- This current weight value would be recorded in the peripheral circuitry (for the whole column), after which both g and g shared would be reset to their initial conditions, thus changing the weight stored on the array.
- the variability and nonlinear response of the nonvolatile higher-significance conductance-pair (G + - G ), implemented for instance with PCM) can be compensated by closed-loop weight tuning to return the weight stored on the array back to the desired target value (by comparison to the value preserved temporarily in the peripheral circuitry).
- any remaining weight-tuning error can be accommodated by programming the g conductance until the total weight again matches the value that was originally recorded in the peripheral circuitry. At this point, the current weight values that were temporarily retained in the peripheral circuitry are discarded.
- FIG. 3 depicts an array having more than one conductance-pair per synapse, with one set of pairs having its read-current amplified with respect to the other by a factor of F, in which the polarity of the lower-significance pairs has been inverted.
- FIG. 4 shows that this polarity inversion technique (the solid and dotted line graphs on top that extend to 20 epochs) markedly improves performance under simulation conditions that include the expected variabilities of CMOS transistors as modeled by SPICE circuit simulations. Without polarity inversion, the fixed asymmetry between charge addition and subtraction due to the unmatched CMOS devices degrades training accuracy noticeably. Yet by inverting the polarity upon each weight-transfer operation, training accuracy is high, matching the expectations of pure software-based training despite the presence of highly imperfect devices.
- the polarity bit must be stored at the ends of the columns of the array, so that the downstream neuron knows how to control which currents are combined on forward inference, which devices are activated for reverse read, and which of the two transistors is activated to add/subtract charge to the capacitor.
- the transfer operation must be performed on a column-by-column basis, with this polarity bit at the end of the column (either south-side or north-side) being inverted to record the current polarity inversion for that column going forward.
- finesse finesse
- One approach is to use the polarity bit to signal whether the column-neuron should compare the incoming current from g against the shared device to the left (lower column number) or to the right (higher column number).
- the idea is that at the same instant that a column chooses to invert its polarity in how it interacts with its dedicated column of g devices, it ALSO starts sharing with a ( f hared device that has ALREADY been recently reset to its initial condition. This implies that when moving from left-to-right in addressing columns for transfer, the inversion of the polarity bit should change from having previously shared to the right to now sharing to the left.
- FIG. 5 shows the situation immediately after transfer of synapses in the same column as neuron Mi, which is sharing to the left and has just had the polarity of its g value inverted (g subtracts from the net weight). Meanwhile, neuron M m continues to have its g value add to the net weight, and shares to a (f hared device located to its right.
- the present invention provides a method to mitigate undesired contributions to weight induced by fixed device asymmetry in an artificial neural network (ANN), a synaptic weight of each synapse in the ANN being represented by a weighted current flow from multiple conductance-pairs, where a plurality of training examples are serially input to the ANN while observing its output, the method comprising: during a first transfer interval representing a first training cycle, transferring a portion of the synaptic weight from a conductance- pair of lower significance, (g - (f hared ), to a conductance-pair of higher significance, (G + - G ); during a second transfer interval after the first transfer interval and representing a second training cycle, transferring a portion of the synaptic weight from the conductance-pair of lower significance that is inverted in polarity, -( g - (f
- the present invention provides a method implemented in an artificial neural network (ANN), the ANN comprising a plurality of neurons arranged in layers with the outputs of one layer connected to the inputs of a plurality of the neurons of a subsequent layer, where neurons are connected to each other via a plurality of synapses, each of the synapses having a synaptic weight encoding a connection strength between two connected neurons, the magnitude of the synaptic weight of each synapse being represented by a weighted current flow from multiple conductance-pairs, where a plurality of training examples are serially input to the ANN while observing its output, where a backpropagation algorithm updates the synaptic weight in response to the difference between the output from a given layer and a desired output from the given layer, the method comprising: (a) periodically transferring a portion of the synaptic weight from a conductance-pair of lower significance to a conductance-pair of higher significance such that the total synaptic weight remains substantially unchanged
- the present invention provides a method implemented in an artificial neural network (ANN), the ANN comprising a plurality of neurons arranged in layers with the outputs of one layer connected to the inputs of a plurality of the neurons of a subsequent layer, where neurons are connected to each other via a plurality of synapses, each of the synapses having a synaptic weight encoding a connection strength between two connected neurons, the magnitude of the weight of each of the synapses represented by a weighted current flow from multiple conductance pairs, each of the multiple conductance pairs representing a joint contribution and having a higher-significance conductance-pair and a lower-significance conductance-pair, where a plurality of training examples are serially input to the ANN while observing its output from the given layer, the method comprising: (a) pausing training and measuring conductances across analog memory elements in the ANN network, and computing an original synaptic weight for each synapse by arithmetic contributions from two or
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Neurology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Image Analysis (AREA)
- Semiconductor Memories (AREA)
Priority Applications (4)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| DE112018004992.2T DE112018004992B4 (de) | 2017-11-20 | 2018-11-19 | Übertragung synaptischer gewichte zwischen leitfähigkeitspaaren mitpolaritätsumkehr zum verringern fester einheitenasymmetrien |
| CN201880074824.XA CN111373414B (zh) | 2017-11-20 | 2018-11-19 | 用于减少固定设备不对称性的具有极性反转的电导对之间的突触权重传递 |
| JP2020526371A JP7139426B2 (ja) | 2017-11-20 | 2018-11-19 | 人工神経回路網用のシナプス重みを構築するための方法、システム、およびプログラム |
| GB2008959.5A GB2581310B (en) | 2017-11-20 | 2018-11-19 | Synaptic weight transfer between conductance pairs with polarity inversion for reducing fixed device asymmetries |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US15/818,521 US11074499B2 (en) | 2017-11-20 | 2017-11-20 | Synaptic weight transfer between conductance pairs with polarity inversion for reducing fixed device asymmetries |
| US15/818,521 | 2017-11-20 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2019097494A1 true WO2019097494A1 (en) | 2019-05-23 |
Family
ID=66533047
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/IB2018/059102 Ceased WO2019097494A1 (en) | 2017-11-20 | 2018-11-19 | Synaptic weight transfer between conductance pairs with polarity inversion for reducing fixed device asymmetries |
Country Status (6)
| Country | Link |
|---|---|
| US (2) | US11074499B2 (cg-RX-API-DMAC7.html) |
| JP (1) | JP7139426B2 (cg-RX-API-DMAC7.html) |
| CN (1) | CN111373414B (cg-RX-API-DMAC7.html) |
| DE (1) | DE112018004992B4 (cg-RX-API-DMAC7.html) |
| GB (1) | GB2581310B (cg-RX-API-DMAC7.html) |
| WO (1) | WO2019097494A1 (cg-RX-API-DMAC7.html) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11074499B2 (en) | 2017-11-20 | 2021-07-27 | International Business Machines Corporation | Synaptic weight transfer between conductance pairs with polarity inversion for reducing fixed device asymmetries |
| EP3701351A4 (en) | 2018-01-30 | 2021-01-27 | D5Ai Llc | SELF-ORGANIZING PARTIALLY ORGANIZED NETWORKS |
| US11694065B2 (en) | 2019-08-28 | 2023-07-04 | Micron Technology, Inc. | Spiking neural unit |
| US11183238B2 (en) | 2019-08-28 | 2021-11-23 | International Business Machines Corporation | Suppressing outlier drift coefficients while programming phase change memory synapses |
| US12052260B2 (en) | 2019-09-30 | 2024-07-30 | International Business Machines Corporation | Scalable and dynamic transfer learning mechanism |
| KR102782076B1 (ko) | 2019-12-13 | 2025-03-18 | 삼성전자주식회사 | 스위칭 소자와 저항 소자를 포함하는 3차원 뉴로모픽 장치 |
| US11475946B2 (en) | 2020-01-16 | 2022-10-18 | International Business Machines Corporation | Synapse weight update compensation |
| KR102873065B1 (ko) * | 2020-04-23 | 2025-10-16 | 삼성전자주식회사 | 뉴로모픽 컴퓨팅 장치 및 그것의 동작 방법 |
| US12156486B2 (en) | 2020-12-02 | 2024-11-26 | International Business Machines Corporation | Horizontal RRAM device and architecture for variability reduction |
| US11908458B2 (en) * | 2020-12-29 | 2024-02-20 | International Business Machines Corporation | Customization of recurrent neural network transducers for speech recognition |
| US12626111B2 (en) * | 2021-06-17 | 2026-05-12 | International Business Machines Corporation | Calibrating peripheral variability |
| KR102574156B1 (ko) | 2021-07-06 | 2023-09-01 | 고려대학교 산학협력단 | 역전파와 음수 가중치 설정이 가능한 뉴로모픽 회로 |
| KR102514652B1 (ko) * | 2021-11-19 | 2023-03-29 | 서울대학교산학협력단 | 뉴로모픽 소자를 위한 가중치 전사 장치 및 이를 이용한 가중치 전사 방법 |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102959565A (zh) * | 2010-07-07 | 2013-03-06 | 高通股份有限公司 | 用于生物启发型网络的通信和突触训练方法及硬件 |
| CN103430186A (zh) * | 2012-01-20 | 2013-12-04 | 松下电器产业株式会社 | 神经网络电路的学习方法 |
| US9715656B1 (en) * | 2016-09-12 | 2017-07-25 | International Business Machines Corporation | Killing asymmetric resistive processing units for neural network training |
Family Cites Families (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5101361A (en) * | 1989-09-29 | 1992-03-31 | The United States Of America As Represented By The Administrator Of The National Aeronautics And Space Administration | Analog hardware for delta-backpropagation neural networks |
| EP0768610B1 (en) | 1995-10-13 | 2000-08-09 | STMicroelectronics S.r.l. | Low-voltage, very-low-power neural network |
| CN101364594B (zh) * | 2007-08-09 | 2010-06-02 | 中国科学院半导体研究所 | 硅基单电子神经元量子电路 |
| US7814038B1 (en) | 2007-12-06 | 2010-10-12 | Dominic John Repici | Feedback-tolerant method and device producing weight-adjustment factors for pre-synaptic neurons in artificial neural networks |
| WO2010133925A1 (en) | 2009-05-20 | 2010-11-25 | Universite Paris Sud (Paris 11) | Teaching method for a neuronal nano- block |
| US9665822B2 (en) | 2010-06-30 | 2017-05-30 | International Business Machines Corporation | Canonical spiking neuron network for spatiotemporal associative memory |
| US8694452B2 (en) | 2010-07-07 | 2014-04-08 | Qualcomm Incorporated | Methods and systems for CMOS implementation of neuron synapse |
| US8676734B2 (en) | 2010-07-07 | 2014-03-18 | Qualcomm, Incorporated | Methods and systems for replaceable synaptic weight storage in neuro-processors |
| US8655813B2 (en) | 2010-12-30 | 2014-02-18 | International Business Machines Corporation | Synaptic weight normalized spiking neuronal networks |
| US8832010B2 (en) | 2012-01-04 | 2014-09-09 | International Business Machines Corporation | Electronic synapses from stochastic binary memory devices |
| US20160364643A1 (en) | 2012-03-08 | 2016-12-15 | Hrl Laboratories Llc | Scalable integrated circuit with synaptic electronics and cmos integrated memristors |
| US20150120631A1 (en) | 2012-05-10 | 2015-04-30 | Consejo Superior de Investagaciones Cientificas (CSIC) | Method and System for Converting Pulsed-Processing Neural Network with Instantaneous Integration Synapses into Dynamic Integration Synapses |
| GB201209174D0 (en) | 2012-05-25 | 2012-07-04 | Univ Bath | Neural network |
| US8811060B2 (en) | 2012-05-31 | 2014-08-19 | International Business Machines Corporation | Non-volatile memory crosspoint repair |
| US9424514B2 (en) | 2012-07-25 | 2016-08-23 | Board Of Trustees Of Michigan State University | Synapse maintenance in the developmental networks |
| US9218564B2 (en) | 2012-07-30 | 2015-12-22 | International Business Machines Corporation | Providing transposable access to a synapse array using a recursive array layout |
| US20150019468A1 (en) | 2013-07-09 | 2015-01-15 | Knowmtech, Llc | Thermodynamic computing |
| US9753959B2 (en) | 2013-10-16 | 2017-09-05 | University Of Tennessee Research Foundation | Method and apparatus for constructing a neuroscience-inspired artificial neural network with visualization of neural pathways |
| CN203661036U (zh) * | 2013-12-20 | 2014-06-18 | 浙江大学城市学院 | 一种基于浮栅技术的二值动态BiCMOS与门电路 |
| US9619749B2 (en) | 2014-03-06 | 2017-04-11 | Progress, Inc. | Neural network and method of neural network training |
| US9990580B2 (en) | 2015-03-13 | 2018-06-05 | International Business Machines Corporation | Neuromorphic synapses |
| US10460237B2 (en) * | 2015-11-30 | 2019-10-29 | International Business Machines Corporation | Neuron-centric local learning rate for artificial neural networks to increase performance, learning rate margin, and reduce power consumption |
| US10891543B2 (en) | 2015-12-28 | 2021-01-12 | Samsung Electronics Co., Ltd. | LUT based synapse weight update scheme in STDP neuromorphic systems |
| US10614355B2 (en) | 2015-12-30 | 2020-04-07 | SK Hynix Inc. | Method for updating weights of synapses of a neuromorphic device |
| US11270194B2 (en) | 2017-07-26 | 2022-03-08 | International Business Machines Corporation | System and method for constructing synaptic weights for artificial neural networks from signed analog conductance-pairs of varying significance |
| US11074499B2 (en) | 2017-11-20 | 2021-07-27 | International Business Machines Corporation | Synaptic weight transfer between conductance pairs with polarity inversion for reducing fixed device asymmetries |
-
2017
- 2017-11-20 US US15/818,521 patent/US11074499B2/en active Active
-
2018
- 2018-11-19 JP JP2020526371A patent/JP7139426B2/ja active Active
- 2018-11-19 GB GB2008959.5A patent/GB2581310B/en active Active
- 2018-11-19 CN CN201880074824.XA patent/CN111373414B/zh active Active
- 2018-11-19 WO PCT/IB2018/059102 patent/WO2019097494A1/en not_active Ceased
- 2018-11-19 DE DE112018004992.2T patent/DE112018004992B4/de active Active
-
2021
- 2021-05-25 US US17/329,245 patent/US11915132B2/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102959565A (zh) * | 2010-07-07 | 2013-03-06 | 高通股份有限公司 | 用于生物启发型网络的通信和突触训练方法及硬件 |
| CN103430186A (zh) * | 2012-01-20 | 2013-12-04 | 松下电器产业株式会社 | 神经网络电路的学习方法 |
| US9715656B1 (en) * | 2016-09-12 | 2017-07-25 | International Business Machines Corporation | Killing asymmetric resistive processing units for neural network training |
Also Published As
| Publication number | Publication date |
|---|---|
| DE112018004992T5 (de) | 2020-06-10 |
| CN111373414B (zh) | 2023-08-25 |
| JP7139426B2 (ja) | 2022-09-20 |
| GB2581310A (en) | 2020-08-12 |
| GB2581310B (en) | 2020-12-09 |
| CN111373414A (zh) | 2020-07-03 |
| US20210287090A1 (en) | 2021-09-16 |
| US11074499B2 (en) | 2021-07-27 |
| GB202008959D0 (en) | 2020-07-29 |
| JP2021503649A (ja) | 2021-02-12 |
| DE112018004992B4 (de) | 2023-11-02 |
| US11915132B2 (en) | 2024-02-27 |
| US20190156194A1 (en) | 2019-05-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11915132B2 (en) | Synaptic weight transfer between conductance pairs with polarity inversion for reducing fixed device asymmetries | |
| US11270194B2 (en) | System and method for constructing synaptic weights for artificial neural networks from signed analog conductance-pairs of varying significance | |
| US11663457B2 (en) | Neural network circuits having non-volatile synapse arrays | |
| US11157810B2 (en) | Resistive processing unit architecture with separate weight update and inference circuitry | |
| US11842770B2 (en) | Circuit methodology for highly linear and symmetric resistive processing unit | |
| US11626884B2 (en) | System and methods for mixed-signal computing | |
| US11630993B2 (en) | Artificial neuron for neuromorphic chip with resistive synapses | |
| US9779355B1 (en) | Back propagation gates and storage capacitor for neural networks | |
| US10453527B1 (en) | In-cell differential read-out circuitry for reading signed weight values in resistive processing unit architecture | |
| Nandakumar et al. | Mixed-precision architecture based on computational memory for training deep neural networks | |
| Narayanan et al. | Toward on-chip acceleration of the backpropagation algorithm using nonvolatile memory | |
| US10460237B2 (en) | Neuron-centric local learning rate for artificial neural networks to increase performance, learning rate margin, and reduce power consumption | |
| Kim et al. | VCAM: Variation compensation through activation matching for analog binarized neural networks | |
| KR20240109086A (ko) | 뉴럴 네트워크 회로 및 이의 동작 방법 | |
| JPH03214373A (ja) | ニューロコンピュータ |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 18878262 Country of ref document: EP Kind code of ref document: A1 |
|
| ENP | Entry into the national phase |
Ref document number: 2020526371 Country of ref document: JP Kind code of ref document: A |
|
| ENP | Entry into the national phase |
Ref document number: 202008959 Country of ref document: GB Kind code of ref document: A Free format text: PCT FILING DATE = 20181119 |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 18878262 Country of ref document: EP Kind code of ref document: A1 |