A Computational Model of Neuronal Spiking and Cognitive Tasks

Anderson, WS

doi: 10.1227/01.neu.0000432668.70016.a6
Science Times

Computational modeling of neural systems, whether in studies of small collections of modeled neurons to understand brain rhythms and seizures, or in larger more abstract systems-based calculations used to study cognition or sensory processing, has risen in prominence as a means of summarizing known experimental data and for hypothesis testing. Eliasmith et al (Eliasmith C, Stewart TC, Choo X, et al. A large-scale model of the functioning brain. Science 2012;338:1202-1205). now describe a spiking neuron model (composed of 2.5 million neurons) which is able to perform certain types of sensory processing and exhibit behavioral responses. Using this model (referred to as “Spaun”) the authors explore the responses of the model to visually presented sequences, including a selection of 8 cognitive tasks. In the case of 1 of these tasks specifically (the rapid variable creation task), Spaun may actually be the first neural network model to accurately solve it independently.

The Spaun model consists of 2.5 million connected neurons that are divided into discrete brain regions, in which the neurons for a particular type of brain region draw their electrophysiologic responses in an architecturally and location specific manner. Regions of motor cortex are represented (M1, SMA, premotor cortex), along with opercular and frontal lobe regions (DLPFC, VLPFC, OFC), the dorsal and ventral visual streams (V1, V2, V4, IT, AIT and the posterior parietal cortex), as well as subcortical structures including thalamus, STN, GPi & GPe, the striatum, substantia nigra, and the ventral tegmental area. As designed, the model is able to process visual inputs, utilize a working memory, perform action selection functions through its subcortical network, and perform computer modeled drawing tasks based on its responses to visual input. Functioning of the model has been optimized under several task conditions (numbered A0-A7), and these consist of: A0 – Drawing copies of images, A1 – Image recognition, A2 – The 3 armed bandit task (reward/choice task), A3 – A functioning serial working memory and recall, A4 – Counting, A5 – Answering questions regarding lists, A6 – Pattern completion tasks, and A7 – Fluid reasoning, ie, a separate pattern completion task adopted from the Raven’ Progressive Matrices test.

In describing the modeling results, the authors use a few exemplar tasks, taking the reader through the model's functioning. For instance, during the A3 working memory task, Spaun is required to store a sequence of symbolic images into a visually based firing pattern, and then use the subcortical network (including the striatum to produce decoded possible actions, and the GPi for action selection) to reproduce the stored memory images. Raster plots of the network activity are shown in the Figure. The authors demonstrate similar results with all eight tasks explored with Spaun. Novel aspects of this modeling effort include the ability to solve different classes of tasks with a coordinated set of neural substrates in a fashion that can be used to form testable hypotheses. The solution for the rapid variable creation task (A6, above) implemented by Spaun, for instance, is novel and may very well differ in a measurable fashion from the human analog.

In summary, Eliasmith et al have developed a computational spiking neuron model of the brain capable of supporting a series of cognitive tasks incorporating sensory inputs and motor outputs. The power of this effort, and in similar brain modeling studies, lies in the ability of the model to summarize known electrophysiology and cognitive behavior and then compare this behavior with experiment. Similar models incorporating pathologies of connectivity or abnormal neuronal firing would be very useful for designing functional neurostimulation therapies or provide clues for therapeutic targets.

Copyright © by the Congress of Neurological Surgeons