I suppose it might be better to learn which synapses to connect this way:
http://svivek.com/teaching/machine-learning/fall2016/slides/07-winnow/winnow.pdf
http://svivek.com/teaching/machine-learning/fall2016/lectures/07-multiplicative-update.html
Moved from #htm-theory:papers to #htm-theory:tangential-theories because:
- linked paper does not relate to HTM
- proposal is non-biological
Being aware of it you can keep it in mind when looking at biological systems. Actually the idea was presented in a video on biological evolution by the same Christos H. Papadimitriou who wrote the Cortical Learning via Prediction paper.
https://youtu.be/KP0WFbdHhJM
If you add in hysteresis where patterns self-reinforce via positive feedback for a while (giving short-term memory) you get an interesting overall package of ideas to explore.
If the (short-term) positive feedback is is giving a good outcome there there is the possibility for its effects to be long-term potentiated by building synapses. You can end up with quite complex learning scenarios.
Paper (MWU and reinforcement learning):
http://www.pnas.org/content/110/49/19950.full.pdf
1 Like
In depth paper (evolution):
https://arxiv.org/abs/1511.01409