Trying to replicate the bar problem from Triesch 2007, "Synergies Between Intrinsic and Synaptic PlasticityMechanisms".
To preview, follow: https://htmlpreview.github.io/?https://github.com/tuanpham96/bar-ica-test/blob/master/html/replicate_barprob_triesch2007IPandSP.html
Main implementation in replicate_barprob_triesch2007IPandSP.m
.
Currently could not replicate bar detection. Either SP alone, or SP+IP (tried a few learning rates between 1E-4
to 1E-1
) would result in a roughly normal weight vector distribution, not a bimodal distribution like the paper. Also tried playing around with target rate mu = 0.01, 0.05, 0.1
, bar probability p_bar = 1/N, 1/2N
and initial activation parameters a_init, b_init
- did not matter.
Possible issues might arise from:
- Implementation of IP or Hebbian SP
- How inputs are generated, whether the way to normalize is correct
- How weights are initialized. Currently, drawn from a uniform distribution [0, 1], then normalized to unit length. Additionally, also tried a normal distribution, a log-normal distribution and a few beta distributions (didn't work).