Memory without Feedback in a Neural Network

By Mark S. Goldman. Neuron, 2009

338 citations recorded by the source corpus.

Summary Memory storage on short time scales is thought to be maintained by neuronal activity that persists after the remembered stimulus is removed. Although previous work suggested that positive feedback is necessary to maintain persistent activity, here it is demonstrated how neuronal responses can instead be maintained by a purely feedforward mechanism in which activity is passed sequentially through a chain of network states. This feedforward form of memory storage is shown to occur both in architecturally feedforward networks and in recurrent networks that nevertheless function in a feedforward manner. The networks can be tuned to be perfect integrators of their inputs, or to reproduce the time-varying firing patterns observed during some working memory tasks but not easily reproduced by feedback-based attractor models. This work illustrates a new mechanism for maintaining short-term memory in which both feedforward and feedback processes interact to govern network behavior.

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

View Paper: https://www.semanticscholar.org/paper/2ab80fd2f97e908b43cd8c010b71dbeda69aef74

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

ROBUST TIMING AND MOTOR PATTERNS BY TAMING CHAOS IN RECURRENT NEURAL NETWORKS (2012, neutral): “In linear networks the structure of recurrent networks, as analyzed though a number of techniques, including eigen and Schur decompositions, provide valuable methods to understand and describe the dynamics of such systems (e.g., Goldman, 2009).”
A model of interval timing by neural integration (2011, methodology): “We propose one such mechanism in the supplemental material (available at www.jneurosci.org) based on computing temporal derivatives in a neurally plausible way, but other approaches also exist (e.g., Koulakov et al., 2002; Goldman et al., 2003; Goldman, 2009).”
Spiking activity propagation in neuronal networks: reconciling different perspectives on neural coding (2010, background): “Thus, even in a recurrent network, the flow of spiking activity on a local scale ('locality' being defined by connection delays, firing rates and refractoriness), can be treated as an FF”
From the Phenomenology to the Mechanisms of Consciousness: Integrated Information Theory 3.0 (2014, neutral): “To achieve a memory of x past time steps in the feed-forward system, the relevant elements were unfolded over time: the state of each element is passed on through a chain of x nodes, one node for each of the x time steps [54,55].”
A dedicated population for reward coding in the hippocampus (2018, neutral): “If reward-associated cell axons did reach the ventral striatum, their sequential activation could conceivably (Goldman, 2009) underlie the ramping spike rate that precedes reward (Atallah et al., 2014) and might even contribute to reward prediction error signals of dopaminergic neurons (Schultz,…”
The Neural Basis of Timing: Distributed Mechanisms for Diverse Functions (2018, background): “Computational models have proposed that sparse population clocks (in which each neuron is active only once during a trajectory) are produced by synfire chains or functionally feedforward patterns of activity (Goldman, 2009; Liu and Buonomano, 2009).”
Recurrent Network Models of Sequence Generation and Memory. (2016, background): “Highly structured circuits of this type have a long history [Kleinfeld & Sompolinsky, 1989; Goldman, 2009], for example, as synfire chain models [Hertz & Prugel-Bennett,1996; Levy et al, 2001; Hermann, Hertz & PrugelBennet, 1995; Fiete et al, 2010], in which excitation flows unidirectionally from…”
Neuronal circuits underlying persistent representations despite time varying activity. (2012, background): “After stimulus presentation at time zero, the network has sustained complex activity for over 30 neural time constants (Figure 2D), as would be expected from complex eigenvalues with real values close (but not equal to) unity, yet stimulus representation remains perfectly stable (in other FEVER netw”