I suggested something like this to the Kempner Institute at Harvard when it was created, but they liked their GPUs too much!

[footnote from this paper: gershmanlab.com/pubs/NeuroAI...

‘It’s not that much work’

And the work of Pillow on the actual latents guiding population activity suggests it doesn’t look like a DDM—at least not as trivially.

A major example of theory experiment (and neuroscience as a field) failure to really test and build on its theories. One inactivation study showed very slight effects on memory guided saccades (1999). But the DDM/LLR model was only tested in 2016+ and showed little effect on decisions.

Depends on what you mean by anti-neuron. Areas have neurons tuned to opposite features—and some clearly have cross interactions like V1. But evidence for the exact theoretical proposal is lacking because it hasn’t been tested. Also, same evidence is (more?) compatible with Bayesian implementations.

I thought that as this was originally proposed it was a theoretical simplification, primarily because they were doing single unit recordings. The original MT papers definitely accepted that the idea was pools of neurons and pools of antineutrons (with regard to tuning).