Role of synaptic variability in resistive memory-based spiking neural networks with unsupervised learning

Ly, D. R. B.; Grossi, Alessandro; Fenouillet-Béranger, C.; Nowak, E.; Querlioz, Damien; Vianello, Elisa

doi:10.1088/1361-6463/aad954

Cited by 23 publications

(18 citation statements)

References 52 publications

(71 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, no obvious advantage can be found over synaptic devices based on bilayer and trilayer structures. Due to a simple structure, oxide single‐layers have also been widely used for synaptic devices, including AlO x , FeO x , HfO x , PrCaMnO x , SrTiO 3 , TaO x , TiO x , WO x , ZnHfO x , ZrHfO x , KNbO 3 , BiFeO 3 , SiO x , and NiO x . The working mechanism of PrCaMnO x ‐based memristive devices is widely attributed to field‐driven oxygen migration and redox reaction at a metal/PrCaMnO x interface .…”

Section: Working Mechanisms Of Memristive Synapsesmentioning

confidence: 99%

Memristive Synapses for Brain‐Inspired Computing

Wang

Zhuge

2019

Adv Materials Technologies

View full text Add to dashboard Cite

be precisely regulated by the ionic flow, which is called synaptic plasticity. Generally, there are two types of synaptic plasticity: potentiation and depression. In the case of potentiation or depression, the synaptic weight increases or decreases upon repeated stimulation. Inspired by the human brain, novel non von Neumann computing architectures mainly composed of artificial synapses and artificial neurons have been proposed, which we can call "artificial neural networks," "brain-like computers," or "neuromorphic computers." [1,[3][4][5] However, the development of artificial brains that possess the efficiency of their biological counterparts in information processing remains a major challenge in computing. [6] One of the major challenges is the lack of suitable hardware to implement synapses, which are the main data processing elements in artificial brains. Generally, there are four essential requirements for an artificial synapse. [7,8] First, the analog weight should be nonvolatile in the absence of learning and weight updates should be bidirectional when the synapse is learning. Second, the synapse output is the product of the input signal and the synapse weight. Third, the weight updates vary with both the input signal and the stored weight value. Fourth, the synapse should be rather compact, and operate off a unidirectional information transfer with low power consumption.Mead and co-workers fabricated a four-terminal synaptic device based on a single floating gate silicon transistor in 1995. [8] It can simultaneously perform long term weight storage, compute the product of the input and floating gate value, and update the weight value according to a Hebbian or a backpropagation learning rule. In 1996, they developed a new floating-gate silicon transistor synapse with a threeterminal structure. [7] Hot-electron injection and electron tunneling make possible bidirectional weight updates. Given that these updates depend on both the stored weight value and the transistor terminal voltages, such synapse can implement a learning function. However, the integration density of transistor synapses is severely restricted to their three-or fourterminal structures. [6] The emergence of memristors endows artificial synapses with a new development opportunity. The memristor (short for memory resistor) is a two-terminal circuit element characterized by a relationship between the charge and the flux-linkage, which shows a distinctive "fingerprint" characterized by a pinched hysteresis loop confined to the first and the third quadrants of the Although the structure and function of the human brain are still far from being fully understood, brain-inspired computing architectures mainly consisting of artificial neurons and artificial synapses have been attracting more and more attentions due to their powerful computing capability and energy efficient operation. Synaptic plasticity is believed to be the origin of learning and memory. However, it is still a big challenge to realize artificial synapses with high reliability, go...

show abstract

Section: Working Mechanisms Of Memristive Synapsesmentioning

confidence: 99%

Memristive Synapses for Brain‐Inspired Computing

Wang

Zhuge

2019

Adv Materials Technologies

View full text Add to dashboard Cite

show abstract

“…GeO 2-x /HfON x [725] , MoO 3-x /Gd 2 O 3-x [726] , HfO 2-x /TiO 2-x /HfO 2-x /TiO 2-x [727] ) to reduce the RS variability and to implement multi-level storage functions. This synaptic plasticity plays a crucial role in the way the brain implements learning and memory [728] . In biological neurons, depending on its characteristic timescale, there is short-term synaptic plasticity (~10 -3 -10s) and long-term plasticity (~10 2 -10 4 s).…”

Section: Synaptorsmentioning

confidence: 99%

Functional Oxides for Photoneuromorphic Engineering: Toward a Solar Brain

Pérez‐Tomás

2019

Adv Materials Inter

View full text Add to dashboard Cite

New device concepts and new computing principles are needed to balance our ever-growing appetite for data and information with the realization of the goals of increased energy efficiency, reduction in CO 2 emissions and the circular economy. Neuromorphic or synaptic electronics is an emerging field of research aiming to overcome the current computer's Von-Neumann bottleneck by building artificial neuronal systems to mimic the extremely energy efficient biological synapses. The introduction of photovoltaic and/or photonic aspects into these neuromorphic architectures will produce self-powered adaptive electronics but may also open up new possibilities in artificial neuroscience, artificial neural communications, sensing and machine learning which would enable, in turn, a new era for computational systems owing to the possibility of attaining high bandwidths with much reduced power consumption. This perspective is focused on recent progress in the implementation of functional oxide thinfilms into photovoltaic and neuromorphic applications towards the envisioned goal of selfpowered photovoltaic neuromorphic systems or a solar brain.

show abstract

“…Nevertheless, despite their outstanding qualities, emerging memories are prone to device variation [4], [10], which can cause bit errors. In conventional applications, this is solved either by relying on error correcting codes [11], or by programming memory cells with high energy pulses that lead to more reliable programming [10]. In this work, based on the experimental measurements of RRAM cells and system level simulations, we investigate the impact of bit errors on in-memory BNNs.…”

Section: Introductionmentioning

confidence: 99%

Outstanding Bit Error Tolerance of Resistive RAM-Based Binarized Neural Networks

Hirtzlin

Bocquet

Klein

et al. 2019

2019 IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS)

View full text Add to dashboard Cite

Resistive random access memories (RRAM) are novel nonvolatile memory technologies, which can be embedded at the core of CMOS, and which could be ideal for the in-memory implementation of deep neural networks. A particularly exciting vision is using them for implementing Binarized Neural Networks (BNNs), a class of deep neural networks with a highly reduced memory footprint. The challenge of resistive memory, however, is that they are prone to device variation, which can lead to bit errors. In this work we show that BNNs can tolerate these bit errors to an outstanding level, through simulations of networks on the MNIST and CIFAR10 tasks. If a standard BNN is used, up to 10 −4 bit error rate can be tolerated with little impact on recognition performance on both MNIST and CIFAR10. We then show that by adapting the training procedure to the fact that the BNN will be operated on error-prone hardware, this tolerance can be extended to a bit error rate of 4 × 10 −2 . The requirements for RRAM are therefore a lot less stringent for BNNs than more traditional applications. We show, based on experimental measurements on a RRAM Hf O2 technology, that this result can allow reduce RRAM programming energy by a factor 30.

show abstract

Role of synaptic variability in resistive memory-based spiking neural networks with unsupervised learning

Cited by 23 publications

References 52 publications

Memristive Synapses for Brain‐Inspired Computing

Memristive Synapses for Brain‐Inspired Computing

Functional Oxides for Photoneuromorphic Engineering: Toward a Solar Brain

Outstanding Bit Error Tolerance of Resistive RAM-Based Binarized Neural Networks

Contact Info

Product

Resources

About