search

UCI-CARL / CARLsim6

CARLsim is an efficient, easy-to-use, GPU-accelerated software framework for simulating large-scale spiking neural network (SNN) models with a high degree of biological detail.
MIT License
41 stars 16 forks source link

[CSTP] Ensuring all IM refractory period steps complete #23

Open nmsutton opened 1 year ago

nmsutton commented 1 year ago

In the connection-specific STP branch (feat/ca3net) the refractory period (RP) included for Izhikevich model neurons can in some cases not process all integration calculation steps in an erroneous way. Specifically, when v_next > vpeak, only the first of the Runge-Kutta 4 (RK4) integration calculation steps are computed and not the rest of them. A note is that the RP code has only currently been implemented in RK4, not other integration methods.

Update 12/30/22: code in kernel_findFiring() has been updated to correctly support RP of any time value. A constant named RP_MAX_LST has been created to support that. However, the value of RP_MAX_LST being 20M is perhaps arbitrary and possibly should be set to a much higher value. Also, looking into the issues related to the kernel_findFiring() code may reveal the same problem exists for LIF-based RP. Another Github issue is likely to be created about this related to LIF.

nmsutton commented 1 year ago

@jkopsick has created code that may fix this issue in the branch ca3net_imfix. Note: this branch also includes code from the proposed bug fix for issue 18. His code update also enables a custom refractory period time value in milliseconds to be set. Either @jkopsick or I will create a push request with this code.

SigmaX commented 1 year ago

How hard would it be to create a unit test that specifically reveals this bug, then passes with the fix?

I see a lot of new tests on that branch, but they are large and complex. Wondering if smaller tests can help cover bugs like this one.

nmsutton commented 1 year ago

Thanks @SigmaX , it should be relatively straightforward to create such a unit test. I think it could be a small simulation to test that the spike timings occur as expected from the IM model. I don't know who authored each unit test (e.g., Jeffrey, Lars, or I) but perhaps moving forward some smaller tests can be made. Jeffrey and I are still testing the code and could work on such a unit test once we are closer to ready to submit a push request.