Getting/Interpreting neural weights


Hi All,

This is Larry. I did the coin-flipping task at this summer’s brain camp
(which I loved). I’m still working on my project and have confirmed that my
model consistently underestimates alternations at a Probability of
Alternation around 0.6. (This is the same area where people have a bias and
will underestimate it as well.)

Now, I want to figure out why. My current theory is that the learning
process is not stable/does not converge under this substantial uncertainty.
To research this, I’ve created a probe to look at the neural weights. I
have a two dimensional input ensemble (200 neurons) connected via PES
learning to a one dimensional prediction (100 neurons). When I use a probe
on the connection to get the weights, it is 3 dimensions:

  • dimension 0 is the time during the learning. The range for this index
    varies depending upon the ‘sample_every’ parameter.

  • dimension 1 only has a range of 1 (value = 0)

  • dimension 2 has a range of 200 (the size of my input ensemble)

I expected the weights to have 3 dimensions (time, 200 for the input
ensemble, 100 for the output ensemble). I seem to be mis-understanding
something. Are not each neuron in the input connected to each neuron of the output? What are the definitions of the dimensions? Any thoughts or advice?




Oops! Sorry. I found an answer to my question on the forum. The dimensions of the weight probe results are: time X dimension of output X neurons of input.

While I’m here… Can anyone steer me toward work on the instability in weight values during simulated learning relating to real world phenomena?