Biomedical Engineering Reference
In-Depth Information
interaction with the environment, in effect, on-the-job training. To be sure, quite
a bit of structure has to be imposed for this development to happen. How to
come up with the appropriate structures is an interesting problem. One strategy
is to borrow from what is known of structure in brains; an alternative is to
evolve structure through evolutionary methods. Another problem is how to regu-
late excitabilities so that development can occur in an adaptive fashion. As we
have seen, for anything sensible to be learned by Hebbian mechanisms, units
must be biased so that they are driven to fire when learning is appropriate and
inhibited when it is not. Brains seem to employ diffuse neuronal projections and
humoral controls for this purpose and these could well be emulated in neuro-
morphic systems. Among the technical problems that need to be solved for neu-
romorphic systems to gain independence in small-scale applications is how to
store and modify connections locally or "on chip" so that learning can occur
without the intervention of external, digital computers. Notwithstanding these
problems and unknowns, the hardware system described achieves, to greater or
lesser extents, the following desirable features of a neuromorphic system:
&
Parallelism of processing elements for fast processing of large
amounts of information
&
Distributed processing whereby many units simultaneously rep-
resent a sensory stimulus or execute a movement
&
Robustness in the face of variation and faults in individual
components, tolerance to damage and noise, and performance
degrading gradually, not catastrophically
&
Real-time processing with wide time-scale dynamics that can
be matched to events in the real world, and to the dynamics of
limbs and other body systems
&
Flexible, rapidly reprogrammable architecture and connections;
capable of ontogenetic and evolutionary adaptation
&
Learning based on short-, medium-, and long-term changes in
synaptic efficacy, unit excitability, and persistent patterns of
activity
&
Efficient, noise-immune signaling by impulses with rich tem-
poral coding possibilities
&
Powerful analog computation distributed throughout a dendritic
tree and soma
&
Compact size and low power consumption, essential for mobil-
ity and portability
