Recent advances in physical reservoir computing: A review

Tanaka, Gouhei; Yamane, Toshiyuki; Héroux, J. B.; Nakane, Ryosho; Kanazawa, Naoki; Takeda, Seiji; Numata, Hidetoshi; Nakano, Daiju; Hirose, Akira

doi:10.1016/j.neunet.2019.03.005

Cited by 1,218 publications

(877 citation statements)

References 252 publications

Supporting

Mentioning

873

Contrasting

Order By: Relevance

“…Our modeling framework is poised to address a broad spectrum of applications in machine learning of natural and artificial signals. With recent advances in reservoir computing (56) and its physical implementations (57), our approach offers an alternative to using external arbitrary time-varying signals to control the dynamics of a recurrent network. Our model may also be extended to neuromorphic hardware, where it may benefit chaotic networks employed in robotic motor control (58).…”

Section: Discussionmentioning

confidence: 99%

Learning long temporal sequences in spiking networks by multiplexing neural oscillations

Vincent-Lamarre

Calderini

Thivierge

2019

Preprint

View full text Add to dashboard Cite

Many cognitive and behavioral tasks -such as interval timing, spatial navigation, motor control and speech -require the execution of preciselytimed sequences of neural activation that cannot be fully explained by a succession of external stimuli. We use a reservoir computing framework to explain how such neural sequences can be generated and employed in temporal tasks. We propose a general solution for recurrent neural networks to autonomously produce rich patterns of activity by providing a multi-periodic oscillatory signal as input. We show that the model accurately learns a variety of tasks, including speech generation, motor control and spatial navigation. Further, the model performs temporal rescaling of natural spoken words and exhibits sequential neural activity commonly found in experimental data involving temporal processing. In the context of spatial navigation, the model learns and replays compressed sequences of place cells and captures features of neural activity such as the emergence of ripples and theta phase precession. Together, our findings suggest that combining oscillatory neuronal inputs with different frequencies provides a key mechanism to generate precisely timed sequences of activity in recurrent circuits of the brain. 1 2 3 4 5 6 7 8 9 10 Reservoir computing | Neural oscillations | Temporal processing | Balanced networks 1. Introduction 1 Virtually every aspect of sensory, cognitive and motor processing in biological organisms involves operations unfolding in time 2(1). In the brain, neuronal circuits must represent time on a variety of scales, from milliseconds to minutes and longer circadian 3 rhythms (2). Despite increasingly sophisticated models of brain activity, time representation remains a challenging problem in 4 computational modelling (3, 4). 5 Recurrent neural networks offer a promising avenue to detect and produce precisely timed sequences of activity (5). However, 6 it is challenging to train these networks due to their complexity (6), particularly when operating in a chaotic regime associated 7 with biological neural networks (7, 8). 8 One avenue to address this issue is with the use of reservoir computing (RC) (9, 10). Under this framework, a recurrent 9 network (the reservoir) projects onto a read-out layer whose synaptic weights are adjusted to produce a desired response. 10 However, while RC can capture some behavioral and cognitive processes (11)(12)(13), it often relies on biologically implausible 11 mechanisms (14). Further, current RC implementations offer little insight to understand how the brain generates activity that 12 does not follow a strict rhythmic pattern (1, 5). That is because RC models are either restricted to learning periodic functions, 13 or require an aperiodic input to generate an aperiodic output, thus leaving the neural origins of aperiodic activity unresolved 14 (5). A solution to this problem is to train the recurrent connections of the reservoir to stabilize innate patterns of activity (12), 15 but this approach is more com...

show abstract

Section: Discussionmentioning

confidence: 99%

Learning long temporal sequences in spiking networks by multiplexing neural oscillations

Vincent-Lamarre

Calderini

Thivierge

2019

Preprint

View full text Add to dashboard Cite

show abstract

“…Then, can DynMat be applied to recurrent networks? It could be possible to use DynMat as a reading layer for Echo or Liquid state machines, which have been successfully applied to multiple real-world problems (Jaeger, 2007;Tanaka et al, 2018). Liquid (or Echo) state machines (LSM) generate dynamically changing patterns which are read out by linear layer to perform computations (Jaeger, 2007;Maass, Natschl, & Markram, 2003).…”

Section: Potential Extension Of Dynmat For Recurrent Networkmentioning

confidence: 99%

DynMat, a network that can learn after learning

Lee

2019

Neural Networks

View full text Add to dashboard Cite

To survive in the dynamically-evolving world, we accumulate knowledge and improve our skills based on experience. In the process, gaining new knowledge does not disrupt our vigilance to external stimuli. In other words, our learning process is 'accumulative' and 'online' without interruption. However, despite the recent success, artificial neural networks (ANNs) must be trained offline and suffer catastrophic interference between old and new learning, indicating that ANNs' conventional learning algorithms may not be suitable for building intelligent agents comparable to our brain. In this study, we propose a novel neural network architecture (DynMat) consisting of dual learning systems inspired by the complementary learning system (CLS) theory suggesting that the brain relies on short-and long-term learning systems to learn continuously.Our empirical evaluations show that 1) DynMat can learn a new class without catastrophic interference and 2) it does not strictly require offline training.building continuous learning systems remains difficult. Moreover, catastrophic interference and a few-shot (i.e., rapid) learning have been studied separately.Then, how does the brain learn continuously? The complementary learning system (CLS) theory proposes that the brain utilizes short-and long-term learning systems for continuous (or continual/lifelong) learning (Norman & O'Reilly, 2003;O'Reilly, Bhattacharyya, Howard, & Ketz, 2014). The short-term learning system relies on fast memory to store new experience (or examples) using sparse and non-overlapping codes (representations); see also (Parisi et al., 2018).Due to its sparse representations, the interference between old and new items (i.e., knowledge) is minimized. The information stored in the short-term learning system can be replayed into a more effective learning system which utilizes dense and overlapping codes for information storage.

show abstract

“…An echo state network is a type of recurrent neural network (RNN), in a classical approach characterized by a randomly generated hidden layer with untrained connections, where only the output weights are subject to supervised training [21,22]. Such a hidden layer is called reservoir [23], and it is able to memorize the time series fed to the network. The reservoir should satisfy the echo state property, i.e., the state of the reservoir should be uniquely defined by the fading history of the input signal.…”

Section: Echo State Network Architecturementioning

confidence: 99%

A Multimodal Approach to the Quantification of Kinetic Tremor in Parkinson’s Disease

Szumilas

Lewenstein

Ślubowska

et al. 2019

Sensors

View full text Add to dashboard Cite

Parkinson's disease results in motor impairment that deteriorates patients' quality of life. One of the symptoms negatively interfering with daily activities is kinetic tremor which should be measured to monitor the outcome of therapy. A new instrumented method of quantification of the kinetic tremor is proposed, based on the analysis of circles drawn on a digitizing tablet by a patient. The aim of this approach is to obtain a tremor scoring equivalent to that performed by trained clinicians. Models are trained with the least absolute shrinkage and selection operator (LASSO) method to predict the tremor scores on the basis of the parameters computed from the patients' drawings. Signal parametrization is derived from both expert knowledge and the response of an artificial neural network to the raw data, thus the approach was named multimodal. The fitted models are eventually combined into model ensembles that provide aggregated scores of the kinetic tremor captured in the drawings. The method was verified with a set of clinical data acquired from 64 Parkinson's disease patients. Automated and objective quantification of the kinetic tremor with the presented approach yielded promising results, as the Pearson's correlations between the visual ratings of tremor and the model predictions ranged from 0.839 to 0.890 in the best-performing models.

show abstract

Recent advances in physical reservoir computing: A review

Cited by 1,218 publications

References 252 publications

Learning long temporal sequences in spiking networks by multiplexing neural oscillations

Learning long temporal sequences in spiking networks by multiplexing neural oscillations

DynMat, a network that can learn after learning

A Multimodal Approach to the Quantification of Kinetic Tremor in Parkinson’s Disease

Contact Info

Product

Resources

About