[above] At the bottom, a standard 3-layer neural net; these neural nets provide a toy model of neural activity during object recognition tasks. On the left is depicted the net's progress through the space of possible weight values during training (several dimensions have been suppressed). On the right we can see how the two prototypes on which the network was trained partition the space of hidden unit activation.
(Typo in original diagram: one point should be labeled "Prototype B")
[below] Such a trained neural network can perform pattern completion on a partial input (say, when only part of a nearby animal is in view) and categorize the partially perceived object with respect to one of the learned prototypes (say, as a rat).
From Churchland and Sejnowski, The Computational Brain, 1992.
No comments:
Post a Comment