Behavior and Pooling Requirements - numenta/htmresearch GitHub Wiki

This is a rough list of requirements for behavior and pooling to help guide us in our brainstorming. The items are fairly vague, open for debate, and in no particular order.

• Unique learned inputs should produce distinct and stable representations
• Novel input should appear stable
• Novel input should be a generalized (union) of previously seen similar things
• Anomaly (surprisal?) should indicate a mis-prediction: when one or more predictions were made, but input didn't match any of them (if no prediction was made, then there was no anomaly)
• More training should form more unique representations and more precise predictions
• Similar things should produce overlapping representations somewhere
• Should be able to invariantly recognize something
• Should be able to invariantly act on something