Sequential Decision Making

From KIP Wiki
Revision as of 14:23, 19 April 2012 by Schrader (talk | contribs) (Created page with "=== Overview === The model aims at mimicking and understanding the dynamics of different cortical areas during a discrimination task. Such a task can be a vibrotactile discri...")
(diff) ↠Older revision | Latest revision (diff) | Newer revision → (diff)
⧼kip-jumptonavigation⧽⧼kip-jumptosearch⧽

Overview

The model aims at mimicking and understanding the dynamics of different cortical areas during a discrimination task. Such a task can be a vibrotactile discrimination task [Romo and Salinas 2003, Nat Rev Neurosci], but also other experiments are conceivable. An important aspect of the model is an instance that serves as a short-term memory, which is a prerequisite for discrimination. For instance, enhanced activity in VPC (Ventral Premotor Cortex) Neurons point to such a memory implementation, which can be reached with facilitating synapses. The original model is able to account for the activity of the so-called "partially differential neurons" in VPC.

Developing institutions

UPF

Developers

Current developers are G. Deco, E. Rolls, R. Romo and E. Hugues.

Results of the ESS workshop, Oct 2011

The model was presented at the Using the ESS + Neuromorphic hardware workshop in Dresden, Oct 4th, 2011 by Etienne Hugues. The slides of the presentation can be found here

  • NMDA dynamics cannot be directly mapped to hardware. It has been discussed how the dynamics can be simulated with sufficient precision using the existing parameters.
  • A first version of the model has been sucessfuly implemented in PyNN.
  • The original model requires a massive input rate, which is currently beyond hardware limits -> downscale necessary.
  • Statistics of the membrane potential (mean membrane potential and its standard deviation) are considered important for the model, so a "good" noise is critical.
  • This can be alleviated a bit by increasing the number Poisson generators, so lower individual rates can be afforded.
  • A delay distribution required. The hardware will support a series of different delays (ESS currently not) which may be further broadened by setting several time-constants, or combining DenMems with different membrane time constants.
  • Facilitation time constants are required in the order of 20ms, which may exceed hardware limits. A possible compensation by interspersing synfire-chain-like memories has been discussed.
  • Facilitation "memory" may have a longer effect if the input rate is further enhanced.