Page 50 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

5. Adaptive Resonance Theory 37

and act upon a desired object in space. Cortical Where stream spatial and motor
representations can locate objects and trigger actions towards them, but cannot
recognize them. By interacting together, the What and Where streams can recognize
valued objects and direct appropriate goal-oriented actions towards them.

5. ADAPTIVE RESONANCE THEORY

Abundant psychological and neurobiological data have confirmed all of the founda-
tional predictions concerning how perceptual/cognitive processes in the What
stream use excitatory matching and match-based learning to create self-stabilizing
categorical representations of objects and events, notably, recognition categories
that can be learned quickly without experiencing catastrophic forgetting during sub-
sequent learning. In other words, this learning process solves the stability-plasticity
dilemma. They thereby enable increasing expertise, and an ever-expanding sense of
self, to emerge throughout life. See Refs. [1,12,13] for reviews.
Excitatory matching by object attention is embodied by the ART Matching Rule
(Fig. 2.4). This type of attentional circuit enables us to prime our expectations to
anticipate objects and events before they occur, and to focus attention upon expected
objects and events when they do occur. Good enough matches between expected and
actual events trigger resonant states that can support learning of new recognition cat-
egories and refinement of old ones, while also triggering conscious recognition of
the critical feature patterns that are attended and enable recognition to occur. Excit-
atory matching also controls reset of the attentional focus when bottom-up inputs
significantly mismatch currently active top-down expectations. Cycles of resonance
and reset underlie much of the brain’s perceptual and cognitive dynamics (Fig. 2.5).

FIGURE 2.4
ART matching rule. Bottom-up inputs can activate their target featural cells, other things
being equal. A top-down expectation, by itself, can only modulate, prime, or sensitize cells
in its excitatory on-center (green [light gray in print version] pathways with hemicircular
adaptive synapses) because of the wider off-surround (red [dark gray in print version]
pathways) that tends to balance the top-down excitation (“one-against-one”) within the
on-center, while causing driving inhibition in the off-surround. When bottom-up inputs
and a top-down expectation are both active, only cells where bottom-up excitation and the
top-down excitatory prime converge in the on-center can fire (“two-against-one”), while
other featural cells are inhibited.