This workgroup is for those who would like to work with spiking networks which learn attractor states - how can you resist?!

Introduction

Neuromorphic chips have been largely devoted to duplicating in silicon the operation of sensory systems, and sometimes to implementing simple, general purpose computational elements supposedly at work in a variety of neural circuits. Typically the studied architectures are either essentially feedforward, as in perceptrons or filters, or they include simple feedback mechanisms, as in winner-take-all or central pattern generators.

In this workgroup we focus on recurrent neural networks with massive feedback, exhibiting attractor behavior. Networks consist of interacting populations of neurons sparsely connected among themselves. We concentrate on collective emerging behaviors at the population level, preferring the mean population firing rate as the informative observable.

We will work on a multi-chip setup developed in the eternal city, Rome, the great point attractor of the empire, to which, 2000 years ago, all roads led. The set up supports a reconfigurable network of IF neurons with Hebbian bistable plastic synapses. The goal is to interface an attractor network with a silicon retina to learn to recognize visual stimuli in real-time.

Relevant Information

Attractor networks

Attractor models have been developed and improved to account for a wide array of experimental evidence related to working memory. They can store and retrieve prescribed patterns of collective activation as ”memories”. Upon the arrival of a “known” external stimulus the network dynamics relax to the closest fixed-point attractor (stored representation) and the resulting firing activity is self-sustained even after the end of the external stimuli. An attractor network acts as an ”associative memory”, retrieving a prototypical memorized representation for a whole class of stimuli which define the ”basin of attraction”.

It is increasingly becoming clear that the dynamic scheme has a wider scope. Models based on bistable or multistable networks have been proposed as theoretical underpinnings for understanding perceptual decision mechanisms and processes of information integration (see Wang08, Amit89), as well as multi-stable perception and binocular rivalry (see Gigante09 and Marti08).

It therefore appears that attractor networks could be considered as general-purpose processing elements, worth the effort of studying them in silicon, with a view to creating complex neuromorphic systems.

Learning

The amount of positive feedback in the attractor network should be adequate to guarantee memory storage and retrieval. Self-excitation depends on synaptic efficacy which can be adjusted through a training process. The neuromorphic chip we use is endowed with Hebbian plastic bistable synapses whose dynamics depends on both the mean firing rate of the pre-synaptic neuron and on the instantaneous value of the post-synaptic neuronal depolarization. For more details on the model see Brader_etal07; for its implementation see Giulioni_etal09 and Mitra_etal.

The neuromorphic setup

The neuromorphic chip we use is described in Giulioni_etal08, it is endowed with 128 neurons and 16k reconfigurable synapses. A first implementation of an attractor network on this chip is reported in Camilleri_etal10. For communication we use the PCI-AER board described in Dante_etal05. The silicon retina we are planning to use has been developed in Zurich by T. Delbruck, please refer to his website for a description of the sensor.

Projects

The starting point is building a stable setup including the silicon retina and the multi-chip platform - a controllable environment for reproducible experiments.

The final goal is to learn to recognize a class of visual stimuli. We can play with different network topologies and different stimuli, using images on a screen or real-world objects.

The same setup could be used for experiments on competitive attractor-based networks for decision-making tasks.

Prerequisites

Hardware-Software requirements

We will bring the entire hardware-software setup needed.

Theoretical Requirements

Basic knowledge about spiking neural networks is suggested. Basic programming skills are suggested (Matlab, C).

Where to catch us

You can find us in the Disco.

An almost-real-time log

Meeting 1: overview and ideas

Discussions

put discussions here

References

book: D.J. Amit, "Modeling Brain Function", Cambridge University Press, 1989.

articles: see attachments