Temporal filtering of dopamine

[emiwar]

Plot with different levels of temporal filtering versus task time constant.

[emiwar]

Maybe one can construct a biochemcial pathway that "deconvolves" the dopamine signal?

[emiwar]

Check cross-frequency between dopamine and eligability when dopamine has temporal dynamics. Are high frequencies lost?

[emiwar]

New idea: scrap the ADAM fig (current fig 3) and instead use the biochemical-pathway schematic to introduce deconv filters. Make a new figure 3 with the story:

• [ ] Low-pass filtering dopamine impairs performance (x-axis: ratio task-timescale dopamine-time scale, y-axis: squared loss).
• [ ] Spine schematic figure on how a high-pass filtered can be constructed
• [ ] Same as first panel, but with added high-pass filter in the synapse. This of course rescues the learning improvement (it has too, since in the ideal (noise-free) case it's mathematically equivalent to not adding any temporal smoothing).
• [ ] Plot with systematically mismatching time constant of the synaptic filter (what's the tolerance?)

Then write in the main text and argue filters can in principle be created for various dopamine kernels (correct Green function, noise, anisotropy, etc) and that the realistic inverse filter is beyond the scope of this paper...

Bonus: nice experimental prediction that there should some node in the biochemical pathway that detects changes in dopamine, or at least some high-frequency filtered version of it.

[emiwar]

One more thought: we know dopamine axons has D2 receptors. Maybe the release sites themselves acts as high-pass filters before the low-pass filtering of the diffusion. Would this work?