search

breznak / neural.benchmark

Comparisons of HTM to other ML algorithms on well known datasets and synthetic anomaly benchmarks
GNU General Public License v2.0
8 stars 1 forks source link

Enhanced HTM model learns high-anomaly states #57

Open breznak opened 8 years ago

breznak commented 8 years ago

@breznak on remembering "rare but important" (aka anomalous) inputs

Exactly right again, the snake scenario is handled by our "animal instincts", in hippocampus (not cortex, HTM). See Thinking fast and Slow, Thaleb. In my opinion the mechanisms for "rare but important" (aka anomalies!) events are:

  • chemical - "adrenaline boost", hippocampus, snake.
  • emotional - from evolution, boost to learning in cases it helps survive - pain from touching a hot iron.
  • attention - will-controlled, we notice and remember something curious, the famous "red-dress woman from the Matrix", psychological aspect; already implemented in HTM in a way that anomaly is implicitly detected. Currently there is not feedback loop with the anomaly value; or HTM could be extended to somehow boost synapses in regard to very high anomaly value!

The others are out of scope for HTM, as they belong more to Agent behavior, but the Attention would be very interesting to implement.