Then, to do what pharmacology cannot, we shuffled learned synapse locations after training, keeping every weight and the nonlinearity intact. Excitatory shuffling drops performance to LDA. Inhibitory shuffling drops it to chance. 9/

As for how placement pays off, inhibition tracks excitation tightly for non-preferred classes, so shunting holds the branch subthreshold. For the preferred class the tracking is looser, excitation escapes, crosses threshold, and is amplified. 8/

Clusters emerge on their own, on distal branches, for excitatory and inhibitory synapses. Remove the NMDAR-like nonlinearity and excitatory clustering falls, matching the APV experiments, but inhibitory clustering rises. 7/

It solves the task and approaches QDA performance with depth. Collapse it to one compartment and performance falls to LDA. Removing any one biological component drops it toward chance. 6/

Dendrinet is a feedforward network of neurons with binary-branching dendritic trees, sparse conductance-based excitatory and inhibitory synapses, shunting inhibition, and an NMDAR-proxy nonlinearity. Nothing is clustered by hand. Synapses relocate during training. 5/

So no single channel is informative. LDA gets ~35%. QDA gets 100%. Any above-chance performance has to be second-order, which is exactly what neighboring synapses on a shared nonlinear branch are positioned to do. 4/

We needed a task where placement could matter. In Permuted Covariance Classification, 200 input channels follow a block correlation structure, and the four classes differ only in which channels co-vary. Means and variances are matched across classes. 3/

ME: Daily notes, reading journals, and dictated thoughts... these are all voice preserved and never touched by AI. That's the part that's actually mine. AI: Linking nodes, surfacing open questions, and updating project records are all the AI's job. The AI provides the glue.