this might be a bit semantic about what we get to call backprop...
-
-
Replying to @achristensen56 @tyrell_turing and
If back prop is just about using past rates with time dependency then a decaying elegibility trace is also back prop..
2 replies 0 retweets 1 like -
Replying to @JonAMichaels @tyrell_turing and
IMO we get to call something 'approx backprop', if it approximates the updates that would happen if using actual backprop.
1 reply 0 retweets 2 likes -
Replying to @achristensen56 @tyrell_turing and
That's probably the most sane way of looking at it. We can test that.
1 reply 0 retweets 1 like -
Replying to @JonAMichaels @tyrell_turing and
as for where to look, and what to look for, e.g. motor prosthetics might be a great place to start, you supply the target fxn.
1 reply 0 retweets 3 likes -
Replying to @achristensen56 @tyrell_turing and
I have such an experiment running so we'll see... neurons are fed into an RNN. I can test if the FR changes match ideal weight updates.
2 replies 0 retweets 7 likes -
Replying to @JonAMichaels @achristensen56 and
Yes sounds awesome
1 reply 0 retweets 4 likes -
Replying to @KordingLab @JonAMichaels and1 reply 0 retweets 1 like
-
Replying to @AdamMarblestone @KordingLab and
But... be mindful that only a *sub-population* of neurons may be subject to backprop-style updates...
2 replies 0 retweets 2 likes -
Replying to @AdamMarblestone @KordingLab and
...and the "weights" in question may be at a higher level than neuron-neuron connections, e.g., could be between assemblies/attractors...
2 replies 0 retweets 3 likes
Much depends on relating back to the anatomy, and on ideally seeing the errors themselves, e.g., http://biorxiv.org/content/early/2016/12/08/092551 …
-
-
Replying to @AdamMarblestone @KordingLab and
I like that https://arxiv.org/abs/1610.00161 and Walter Senn's work try to relate this back to specific anatomy and neuron types...
0 replies 0 retweets 5 likesThanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.