-
Notifications
You must be signed in to change notification settings - Fork 109
Extraneous Predictions
September 21, 2014 - Chetan Surpur
Even after sufficient training, the temporal memory can make extraneous predictions during a sensorimotor sequence. This happens when patterns are shared between worlds.
The reason this happens comes down to the fact that the input on the distal dendrites does not contain information about which world is currently being explored. Remember, the input on the distal dendrites consists of: A) the first-order (columnar) representation of the sensory pattern; and B) the encoded motor command. So, for a given sensory pattern and motor command, the temporal memory will always make a certain prediction, regardless of which world this sensory pattern and motor command was generated from. In other words, if a sensory pattern appears in multiple worlds (and motor commands are universal), that sensory pattern combined with a motor command will cause a prediction independent of the world.
Let's take the case of two worlds: ABCD and DCBA. Given a particular sensory pattern (C) and motor command (+1), world ABCD will generate (D) and world DCBA will generate (B). However, after training the temporal memory on many transitions in these two worlds, given sensory pattern (C) and motor command (+1) will have to predict D and B. Even if it just saw (B) and (+1) generate (C) (so it's obvious it's in world ABCD), that information isn't propagated to the next prediction; the only information that is used to make a prediction is (C) and (+1). Which is not enough information to distinguish between the two worlds, since (C) and (+1) is a valid transition seen in both worlds.
It's possible that this is perfectly acceptable, and the sensorimotor temporal memory is only responsible for representing first-order transitions. The pooling layer can still form a single representation of all transitions in a given world, from the correctly predicted cells in the temporal memory layer.
If we want the temporal memory layer to make fewer and more accurate predictions, we can use feedback from the pooling layer, which has information on which world we're currently in, to prime the temporal memory layer towards the best predictions. This would work by having the pooling layer feedback depolarize all the cells that represent valid transitions in this world (these are cells that the pooling layer has pooled over for this world). These cells would then win against other cells that could become predicted due to a given sensorimotor transition.
An open question with this approach is: how would it affect the pooling process, and the rate of learning? Because then, pooling would be dependent on which cells were correctly predicted, and which cells were correctly predicted would be dependent on pooling.
September 22, 2014 - Subutai Ahmad
We discussed adding lateral connections in Layer 4 to disambiguate. Now that we have winner cell hysteresis, this may work.