A preprint of the following paper is now online available:
REAL-TIME COMPUTING WITHOUT STABLE STATES:
A NEW FRAMEWORK FOR NEURAL COMPUTATION BASED ON PERTURBATIONS
by
Wolfgang Maass, Thomas Natschlger, and Henry Markram.
(Graz Univ. of Technology, Austria, and Weizmann Institute, Israel)
ABSTRACT:
This paper has the goal to establish a theoretical framework for
computations on spike trains that can also be applied to biologically
realistic models for recurrent circuits of spiking neurons. This new
theoretical framework, the liquid state machine, differs strongly from
the computational models that have emerged from computer science and
artificial neural networks: it is not based on transitions between
stable internal states or attractors, but rather exploits the natural
transient dynamics of recurrent neural circuits as a potentially
powerful analog memory device. It directs attention to the
investigation of trajectories of transient internal states in very
high dimensional dynamical systems, thereby providing a complement to
the analysis of attractors in low dimensional dynamical systems that
have so far been used as primary sources of inspiration for
understanding the dynamics of neural computation.
Like the Turing machine this model allows for basically unlimited
computational power under idealized conditions, but for real-time
computing on time-varying inputs with fading memory (rather than for
offline-computing on static discrete inputs like the Turing machine).
Based on this new framework we have for the first time been able to
carry out complex real-time computations on spike trains with
biologically realistic computer models of neural microcircuits.
This approach also suggests a radically new approach towards
neuromorphic engineering: Look directly for efficient hardware
implementations of adaptive liquid state machines in order to build
devices for real-time processing of sensory inputs that capture
aspects of the organisation of neural computation.
Learning issues in the context of this model (especially biologically
plausible algorithms for unsupervised learning and applications of
reinforcement learning) are topics of current research.
------------------------------------------------------------------------------------
This paper is online available (PDF, 243 KB) as # 130 from
http://www.igi.tugraz.at/maass/publications.html