Updates, talks, threads. Text supports markdown and LaTeX; embeds pull from X & Bluesky; links point back to longer posts.
Each layer of a plain MLP integrates out short-distance degrees of freedom in input space. The mutual information $I(X; T_\ell)$ at layer $\ell$ follows a renormalisation-group trajectory — the same monotonicity (Data Processing Inequality) as block-spin RG.
Writing it up; preprint by July.
New preprint up: "Depth as Successive Coarse-Graining in Plain MLPs." The information-bottleneck profile across layers is, formally, an RG flow. Code + figures linked.
View on x.com →Slides + recording available.
OpenSpent the morning re-reading Cajal's monograph. The dendritic field of a Purkinje cell is a beautifully asymmetric receptive volume — and the spines are the first published evidence that learning needs surface area.
View on x.com →Re-analysing the 2010 Journal of Neurophysiology dataset with modern source-localisation. Same result holds: the spindle is several spindles thinking at once.