Reservoir computing using self-sustained oscillations in a locally connected neural network DOI Creative Commons
Yuji Kawai, Jihoon Park, Minoru Asada

и другие.

Scientific Reports, Год журнала: 2023, Номер 13(1)

Опубликована: Сен. 19, 2023

Understanding how the structural organization of neural networks influences their computational capabilities is great interest to both machine learning and neuroscience communities. In our previous work, we introduced a novel system, called reservoir basal dynamics (reBASICS), which features modular architecture (small-sized random networks) capable reducing chaoticity activity producing stable self-sustained limit cycle activities. The integration these cycles achieved by linear summation weights, arbitrary time series are learned modulating weights. Despite its excellent performance, interpreting structure isolated small as brain network has posed significant challenge. Here, investigate local connectivity, well-known characteristic networks, contributes system generates based on empirical experiments. Moreover, present performance locally connected reBASICS in two tasks: motor timing task Lorenz series. Although was inferior that reBASICS, could learn tens seconds while constant units ten milliseconds. This work indicates locality connectivity may contribute generation oscillations long-term series, well economy wiring cost.

Язык: Английский

An organized view of reservoir computing: a perspective on theory and technology development DOI Creative Commons
Gisya Abdi, Tomasz Mazur, Konrad Szaciłowski

и другие.

Japanese Journal of Applied Physics, Год журнала: 2024, Номер 63(5), С. 050803 - 050803

Опубликована: Апрель 1, 2024

Abstract Reservoir computing is an unconventional paradigm that uses system complexity and dynamics as a computational medium. Currently, it the leading in fields of materia computing. This review briefly outlines theory behind term ‘reservoir computing,’ presents basis for evaluation reservoirs, cultural reference reservoir haiku. The summary highlights recent advances physical points out importance drive, usually neglected implementations However, drive signals may further simplify training reservoirs’ readout layer training, thus contributing to improved performance computer performance.

Язык: Английский

Процитировано

3
Classical and Quantum Physical Reservoir Computing for Onboard Artificial Intelligence Systems: A Perspective DOI Creative Commons

A. H. Abbas,

Hend Abdel-Ghani,

Ivan S. Maksymov

и другие.

Dynamics, Год журнала: 2024, Номер 4(3), С. 643 - 670

Опубликована: Авг. 12, 2024

Artificial intelligence (AI) systems of autonomous such as drones, robots and self-driving cars may consume up to 50% the total power available onboard, thereby limiting vehicle’s range functions considerably reducing distance vehicle can travel on a single charge. Next-generation onboard AI need an even higher since they collect process larger amounts data in real time. This problem cannot be solved using traditional computing devices become more power-consuming. In this review article, we discuss perspectives development neuromorphic computers that mimic operation biological brain nonlinear–dynamical properties natural physical environments surrounding vehicles. Previous research also demonstrated quantum processors (QNPs) conduct computations with efficiency standard computer while consuming less than 1% battery power. Since QNPs are semi-classical technology, their technical simplicity low cost compared make them ideally suited for applications systems. Providing perspective future progress unconventional reservoir surveying outcomes 200 interdisciplinary works, article will interest broad readership, including both students experts fields physics, engineering, technologies computing.

Язык: Английский

Процитировано

3
Dynamical measures of developing neuroelectric fields in emerging consciousness DOI Creative Commons
William J. Bosl, Jenny Capua-Shenkar

Current Opinion in Behavioral Sciences, Год журнала: 2025, Номер 61, С. 101480 - 101480

Опубликована: Янв. 10, 2025

Язык: Английский

Процитировано

0
Asymmetrically connected reservoir networks learn better DOI Creative Commons
Shailendra K. Rathor,

Martin Ziegler,

Jörg Schumacher

и другие.

Physical review. E, Год журнала: 2025, Номер 111(1)

Опубликована: Янв. 24, 2025

We show that connectivity within the high-dimensional recurrent layer of a reservoir network is crucial for its performance. To this end, we systematically investigate impact on performance, i.e., examine symmetry and structure in relation to computational power. Reservoirs with random asymmetric connections are found perform better an exemplary Mackey-Glass time series than all structured reservoirs, including biologically inspired connectivities, such as small-world topologies. This result quantified by information processing capacity different topologies which becomes highest randomly connected networks. Published American Physical Society 2025

Язык: Английский

Процитировано

0
Reduced connection strength leads to enhancement of working memory capacity in cognitive training DOI Creative Commons

Guiyang Lv,

Tianyong Xu,

Junfa Li

и другие.

NeuroImage, Год журнала: 2025, Номер 308, С. 121055 - 121055

Опубликована: Янв. 30, 2025

Язык: Английский

Процитировано

0
A simulated memristor architecture of neural networks of human memory DOI Creative Commons

Tihomir Taskov,

Juliana Dushanova

Brain Organoid and Systems Neuroscience Journal, Год журнала: 2025, Номер unknown

Опубликована: Фев. 1, 2025

Язык: Английский

Процитировано

0
The Drosophila Connectome as a Computational Reservoir for Time-Series Prediction DOI Creative Commons
Leone Costi, Alexander Hadjiivanov, Dominik Dold

и другие.

Biomimetics, Год журнала: 2025, Номер 10(5), С. 341 - 341

Опубликована: Май 21, 2025

In this work, we explore the possibility of using topology and weight distribution connectome a Drosophila, or fruit fly, as reservoir for multivariate chaotic time-series prediction. Based on information taken from recently released full connectome, create connectivity matrix an Echo State Network. Then, use only most connected neurons implement two possible selection criteria, either preserving breaking relative proportion different neuron classes which are also included in documented to obtain computationally convenient reservoir. We then investigate performance such architectures compare them state-of-the-art reservoirs. The results show that connectome-based architecture is significantly more resilient overfitting compared standard implementation, particularly cases already prone overfitting. To further isolate role synaptic weights, hybrid reservoirs with but random weights topologies study, demonstrating both factors play increased resilience. Finally, perform experiment where entire used Despite much higher number trained parameters, remains has lower normalized error, under 2%, at regularisation, all other regularisation.

Язык: Английский

Процитировано

0
Circuit design in biology and machine learning. I. Random networks and dimensional reduction DOI Creative Commons
Steven A. Frank

Evolution, Год журнала: 2025, Номер unknown

Опубликована: Май 29, 2025

Abstract A biological circuit is a neural or biochemical cascade, taking inputs and producing outputs. How have circuits learned to solve environmental challenges over the history of life? The answer certainly follows Dobzhansky’s famous quote that “nothing in biology makes sense except light evolution.” But leaves out mechanistic basis by which natural selection’s trial-and-error learning happens, exactly what we understand. does process designs actually work? much insight can gain about form function studying processes made those circuits? Because life’s must often same problems as faced machine learning, such tracking, homeostatic control, dimensional reduction, classification, begin considering how computational problems. We then ask: do provide design differ from computers particular it uses problems? This article steps through two classic models set foundation for analyzing broad questions circuits. One surprising power randomly connected networks. Another central role internal environment embedded within circuits, illustrated model reduction trend prediction. Overall, many analogs, suggesting hypotheses biology’s are designed.

Язык: Английский

Процитировано

0
Biologically plausible models of cognitive flexibility: merging recurrent neural networks with full-brain dynamics DOI Creative Commons

Maya van Holk,

Jorge F. Mejías

Current Opinion in Behavioral Sciences, Год журнала: 2024, Номер 56, С. 101351 - 101351

Опубликована: Фев. 6, 2024

Cognitive flexibility, a cornerstone of human cognition, enables us to adapt shifting environmental demands. This brain function has been widely explored using computational modeling, although oftentimes these models focus on the operational dimension cognitive flexibility and do not retain sufficient level neurobiological detail lead electrophysiological or neuroimaging insights. In this review, we explore recent advances future directions neurobiologically plausible flexibility. We first cover progress in recurrent neural network trained perform flexible tasks, followed by discussion how whole-brain large-scale have approached distributed nature functions. Ultimately, propose here hybrid framework which both modeling philosophies converge, advocating for balanced approach that merges power with realistic spatiotemporal dynamics activity, early examples direction.

Язык: Английский

Процитировано

2
Physical Reservoir Computing Enabled by Solitary Waves and Biologically Inspired Nonlinear Transformation of Input Data DOI Creative Commons
Ivan S. Maksymov

Dynamics, Год журнала: 2024, Номер 4(1), С. 119 - 134

Опубликована: Фев. 8, 2024

Reservoir computing (RC) systems can efficiently forecast chaotic time series using the nonlinear dynamical properties of an artificial neural network random connections. The versatility RC has motivated further research on both hardware counterparts traditional algorithms and more-efficient RC-like schemes. Inspired by processes in a living biological brain solitary waves excited surface flowing liquid film, this paper, we experimentally validated physical system that substitutes effect randomness underpins operation algorithm for transformation input data. Carrying out all operations microcontroller with minimal computational power, demonstrate so-designed serves as technically simple counterpart to ‘next-generation’ improvement algorithm.

Язык: Английский

Процитировано

2