next up previous
Next: Using Disinhibition to Up: A Grounded System Previous: A Grounded System

Learning Separate Representations for Visible and Invisible Implies a Grounded Learning System

Hebbian-type learning rules are attractive because of their simplicity. Although we considered standard variants of Hebbian rules in our quest for a visual system that can learn to represent depth information, we rejected them because they were subject to a severe theoretical problem. If a Class II neuron, which is activated by lateral excitation alone, excites other neurons and causes them to become active without bottom-up excitation, then a Hebbian-type learning rule would cause its lateral excitatory connections to the other neurons to strengthen, and it would cause the bottom-up inputs to the other neurons to weaken. The other neurons would thereby tend to be converted to Class II neurons. Thus, if the network contains Class II neurons, then Hebbian-type rules have no way to prevent all neurons within the same network stage from becoming Class II neurons. If that happened, then the network stage would be divorced from actual visual input flowing from prior stages (Figure 3b). Neuron activations within the stage would propagate in an uncontrolled, ``hallucinatory'' manner.

  
Figure 3: Hallucination problem. (a) The learning rules must be chosen so that bottom-up image information can flow to every layer, e.g., via connections indicated by vertical arrows. (b) Otherwise, some layers might learn strong lateral connections at the expense of bottom-up connections. Such layers would become disconnected from actual visual input signals.

Clearly, that would be a failure. We sought a learning system that would prevent hallucinatory networks from developing. The rules that we chose were thus required to be grounded -- i.e., to keep some neurons supplied with actual bottom-up input.


next up previous
Next: Using Disinhibition to Up: A Grounded System Previous: A Grounded System