Carlo Nicolini

La mente che conta fino a dieci

Mon, 01 Jun 2026 00:00:00 +0000

La cognizione è diffusa e costa poco. Un sé che vive nel tempo è raro, fragile, legato a un corpo. E una macchina che rimescola l’ordine dei propri pensieri senza accorgersene è lo specchio più strano che abbiamo mai costruito.

A controlled testbed for Repeat-Your-Self — SAT solving, message passing, and a predictive rho/phi theory

Fri, 29 May 2026 00:00:00 +0000

RYS Controlled SAT message passing

Deep Equilibrium Models as a model for RYS

Mon, 18 May 2026 00:00:00 +0000

In the last two posts I tried to put David Ng’s Repeat-Your-Self construction into a geometric language. In How skip connections define graphs in deep networks I started from the residual recursion

Fractal Connectivity Scaffold Networks

Sat, 16 May 2026 00:00:00 +0000

1. Question

Hippocampal scaffold networks

Fri, 15 May 2026 00:00:00 +0000

The tradeoff

Multiscale structured recurrence with fractal connectivity networks

Fri, 15 May 2026 00:00:00 +0000

1. Introduction

Similarity of neural networks representations

Wed, 29 Apr 2026 00:00:00 +0000

In yesterday’s post on How skip connections define graphs in deep networks I built an \(L\times L\) functional connectome of a Transformer by taking pairwise cosine similarities of layer activations. That edge had a clean closed form in the cumulative residual force \(\rho_{i,j}\) and the residual-vs-identity angle \(\phi_{i,j}\), and the same closed form predicted that David Ng’s Repeat Your Self (RYS) construction (Ng, 2026) amplifies the off-plateau deviation in the central reasoning block by a factor of four.

How skip connections define graphs in deep networks

Tue, 28 Apr 2026 00:00:00 +0000

The hidden graphs behind residual connections

Functional synthetic LLM connectome analysis

Tue, 28 Apr 2026 00:00:00 +0000

Connections to network neuroscience

PLP Inference Audited Against Murphy

Tue, 21 Apr 2026 00:00:00 +0000

Kevin Murphy’s treatment of inference on pp. 435-471 of Probabilistic Machine Learning: Advanced Topics gives us a clean gold standard for what exact and approximate inference algorithms are supposed to mean. Those pages cover the forwards-backwards algorithm for hidden Markov models, sum-product and max-product message passing on trees, factor-graph belief propagation, loopy belief propagation and its caveats, variable elimination, and the junction-tree viewpoint. They do not present these methods as vague metaphors. They present them as specific operators with specific mathematical guarantees.