A computational model for long-term memory storage in networks of neurons

This research perspective was sponsored by Sparrho. Click here to read more about Sparrho’s new initiative to reward scientists for communicating their research to a wider audience.

Research perspective written by Dylan Festa
Computational & Biological Learning Lab, Department of Engineering
University of Cambridge, Cambridge, UK

How can I explain the importance of this research to the general public?

How do we remember things? Scientific evidence shows that memory consists in modifications in the connections between the cells that compose the brain (the neurons) that influence their activity, with distinct memories corresponding to distinct patterns of brain (neural) activity. Modelling this process at the level of neurons proves a difficult task, especially if trying to take into account the constraints and features present in biological cells. A limitation often encountered, for example, is that modelled neurons display only an “active” and “inactive” state while recalling a memory, whereas real neurons feature “memory states” at any level of activity, over an analog scale.

This work introduces a new technique based on control theory, a numerical optimisation procedure capable of storing memory states in neural network models. The resulting model surpasses several of the previous limitations: in addition to analog memory states, the neurons can assume roles more similar to what is experimentally observed, and their connectivity and dynamics appear more realistic, while still being able to incorporate a large number of distinct memory patterns.

This study advances our understanding of how memory works by producing more realistic and flexible models; moreover its methods could be extended to the general class of problems that require the creation of memory states in the form of stable points of the dynamical activity of some modelled system.

Why is this important for researchers in fields other than computational neuroscience?

This study introduces a new technique that can be effectively used to build more realistic computational models of auto-associative memory in the brain. It proposes a new approach to the construction of dynamically stable models of neural population dynamics that allows for some realism in how the single neurons and their activity patterns are defined. Researchers who model interconnected neurons might find this kind of optimisation a valid alternative to manual fine-tuning of the numerical parameters. In a broader perspective, this technique can be extended to the general problem of embedding locally robust fixed-point attractors in a high-dimensional dynamical system by tuning the internal parameters of the dynamics under arbitrary constraints.

Why is this important for researchers in the same field?

This study introduces and validates a new control-theoretic technique aimed at building functioning attractor networks that satisfy a set of relevant physiological constraints. The networks are rate-based models featuring distinct excitatory (E) and inhibitory (I) populations, directly optimised to force sets of arbitrary analog patterns to become stable fixed points of their dynamics. The neurons operate in a E-I balanced regime, and the recalled memory patterns are robust to corruptions of the memory cue as well as to ongoing noise. Overall this novel class of networks successfully overcomes the typical limitations of the previous models in a unified way, and constitutes a step forward in understanding the neural substrate of memory.

Original article

Analog Memories in a Balanced Rate-Based Network of E-I Neurons
Dylan Festa, Guillaume Hennequin, Mate Lengyel
Advances in Neural Information Processing Systems 27 (NIPS 2014)

Acknowledgements

This work was supported by the Wellcome Trust, the European Union Seventh Framework Programme, and the Swiss National Science Foundation. The original text was published in the Electronic Proceedings of the Neural Information Processing Systems Conference.

Originally published at blog.sparrho.com.