MesoNet: unresolved experiment matrix

too many possible combinations rn. make the encoder × architecture × update-rule grid before another round of parameter spaghetti

questions collected in the January review: efficient multiscale STDP; adapting unsupervised clustering to spiking activity; dataset encoding; and whether pure STDP is sufficient for the intended hierarchy.

proposed experiments from the source: alternative encoders on CIFAR-100, hybrid STDP/surrogate-gradient learning, and comparisons with conventional deep-learning architectures.

comparison structure: separate encoder, architecture and update-rule changes. For each combination, retain activity/weight diagnostics, representation/readout metrics and compute cost. This organization is a reference scaffold, not a record that all combinations were run.

Experiment constants

Planned model changes

• review question: which earlier assumptions should be tested first to rule out an unproductive direction? In particular, distinguish a computationally feasible rule from one that actually learns higher-level representations.

• References saved in this journal section

• https://ieeexplore.ieee.org/document/8094937/
• https://pubmed.ncbi.nlm.nih.gov/24244113/
• https://pmc.ncbi.nlm.nih.gov/articles/PMC2756108/
• https://dl.acm.org/doi/10.5555/2969033.2969154
• https://doi.org/10.1038/s41467-017-01827-3

Experiment design and hardware