Generating Coherent Patterns of Activity from Chaotic Neural Networks

By David Sussillo, L. F. Abbott. 2009

1,136 citations recorded by the source corpus.

Neural circuits display complex activity patterns both spontaneously and when responding to a stimulus or generating a motor output. How are these two forms of activity related? We develop a procedure called FORCE learning for modifying synaptic strengths either external to or within a model neural network to change chaotic spontaneous activity into a wide variety of desired activity patterns. FORCE learning works even though the networks we train are spontaneously chaotic and we leave feedback loops intact and unclamped during learning. Using this approach, we construct networks that produce a wide variety of complex output patterns, input-output transformations that require memory, multiple outputs that can be switched by control inputs, and motor patterns matching human motion capture data. Our results reproduce data on pre-movement activity in motor and premotor cortex, and suggest that synaptic plasticity may be a more rapid and powerful modulator of network activity than generally appreciated.

Subfields: Artificial Intelligence, Cognitive Neuroscience, Electrical and Electronic Engineering

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Cognitive Neuroscience

In Electrical and Electronic Engineering

Citations in the corpus

Re-evaluating circuit mechanisms underlying pattern separation (2019, neutral): “, 2015), in analogy with reservoir computing (Yamazaki and Tanaka, 2007; Buonomano and Maass, 2009; Sussillo and Abbott, 2009).”
Toward an Integration of Deep Learning and Neuroscience (2016, positive): “In internal FORCE learning (Sussillo and Abbott, 2009), internally generated random ﬂuctuations inside a chaotic recurrent network are adjusted to provide feedback signals that drive weight changes internal to the network while the outputs are clamped to desired patterns.”
Physical reservoir computing—an introductory perspective (2020, positive): “…weights (although there is a slight difference concerning whether the output is injected into the nonlinear activation function before fed back to the reservoir), the FORCE learning scheme has been applied not only to the readout weights, but also to the internal weights of the reservoir [143–145].”
Organizing recurrent network dynamics by task-computation to enable continual learning (2020, positive): “Preconditioning methods [31–33] and FORCE learning [34] use within-batch correlational structure to adaptively change learning rates in order to speed up learning and convergence.”
Continuous Learning of Context-dependent Processing in Neural Networks (2018, positive): “This iterative method is essentially the Recursive Least Square (RLS) algorithm [24, 25] (see Methods), which has been used to train feedforward and recurrent neural networks to achieve fast convergence [26, 27], tame chaotic activities [28] and avoid interference between consecutively loaded patter”
Dimensionality reduction for large-scale neural recordings (2014, positive): “Their use of dimensionality reduction also provided a bridge between population recordings and network theory 49 , in which both the experimental data and the model indicated a similar dynamical process for information gating.”
Computation Through Neural Population Dynamics (2020, positive): “The fixed points and corresponding linearizations of an RNN’s dynamics can be identified via numerical optimizations (Golub & Sussillo 2018, Sussillo & Barak 2013).”
A Practical Guide to Applying Echo State Networks (2012, positive): “8) online learning algorithm to vigorously adapt W in the presence of the real feedbacks [46].”
Neural population dynamics during reaching (2012, neutral): “4) there exist non-rotational components as well, perhaps reflecting the nonlinear dynamics necessary for initiating/terminating movement, for stability 37 , and for feedback control 16,38 .”
Artificial Neural Networks for Neuroscientists: A Primer (2020, positive): “In particular, the FORCE algorithm [Sussillo and Abbott, 2009] rapidly modiﬁes the output connections of an RNN to match the target using a recursive least square algorithm.”
Training Excitatory-Inhibitory Recurrent Neural Networks for Cognitive Tasks: A Simple and Flexible Framework (2016, positive): “…it is convenient to initialize the recurrent weight matrix as W rec ¼ r W rec 0 , where W rec 0 is formed by setting a fraction p , 0 p (cid:4) 1, of elements to nonzero values drawn from a Gaussian distribution with mean 0 and variance ( pN ) − 1 , and the remaining fraction 1 − p to zero [31].”
Inferring single-trial neural population dynamics using sequential auto-encoders (2017, positive): “5, both of which produce chaotic dynamics at a relatively slow timescale compared to ⌧ (see [31] for more details).”
Artificial Neural Networks-Based Machine Learning for Wireless Networks: A Tutorial (2017, positive): “spread and exhibits a much faster convergence. Demonstrations of RLS for ESNs are presented in [96] and [103]. The backpropagation-decorrelation in [104] and the so-called FORCE learning algorithm in [105] are two other powerful methods for online training of single-layer output with feedback connec”