Finite State Machines in VLSI Spiking Neurons

Introduction

Complex behavior requires the completion of subtasks executed in a sequential order. To support such behavior, a neural system must have the ability to encode and maintain a previously presented cue in a form of short-term memory.

In the field of neuroscience, this ability relies on working memory, whose function is to hold an event in mind in order to perform the next steps of a complex task. To support high-order cognitive tasks, the system must also be able to use the stored information and guide future behavior, in a state-dependent fashion.

Finite State Machines (FSM) are ideal candidates which fulfill the requirements described above. Based on previous work (Rutishauser Douglas, Neural Computation 2009), we devised a method which configures a multi-neuron VLSI system composed of soft Winner-Take-All networks to implement any FSM. In this workgroup, the participants can either implement their desired FSM, or study and extend the underlying network.

Requirements

• To program the state machine, we use the open source  JFLAP software. You can either install it or use the provided applet (use the "Finite Automaton" or the "Mealy Machine").

Note: if you wish to learn more about the implementation details, join the tutorial on Tuesday, 15:30 in the disco

Project: Controlling the behaviour of a robot grasping a cup or a book

Bibliography

• State-dependent computation using coupled recurrent networks, Rutishauser U., Douglas, R., Neural Computation 2009