In-Context Learning Demystified?
Published:
📖 TL;DR: a transformer block implicitly uses the input context to modify its MLP weights.
Blog posts
2025
Energy-based Transformers
Published:
📖 TL;DR: Energy-based Transformers (EBTs) learn a scalar energy function parameterised by a transformer. Empirically, EBTs show promising scaling and reasoning properties on both language and vision tasks.
🔥 Scaling Predictive Coding to 100+ Layer Networks
Published:
📖 TL;DR: We introduce \(\mu\)PC, a reparameterisation of predictive coding networks that enables stable training of 100+ layer ResNets with zero-shot hyperparameter transfer.
♾️ Infinite Widths (& Depths) Part III: The Maximal Update Parameterisation (\(\mu\)P)
Published:
This is the third and last post of a short series on the infinite-width limits of deep neural networks (DNNs). In Part I, we showed that the output of a random network becomes Gaussian distributed in the infinite-width limit. Part II went beyond initialisation and showed that infinitely wide nets trained with GD are basically kernel methods.
♾️ Infinite Widths Part II: The Neural Tangent Kernel
Published:
This is the second post of a short series on the infinite-width limits of deep neural networks (DNNs). Previously, we reviewed the correspondence between neural networks and Gaussian Processes (NNGP), showing that, as the number neurons in the hidden layers grows to infinity, the output of a random network becomes Gaussian distributed.
2024
♾️ Infinite Widths Part I: Neural Networks as Gaussian Processes
Published:
This is the first post of a short series on the infinite-width limits of deep neural networks (DNNs). We start by reviewing the correspondence between neural networks and Gaussian Processes (GPs).
KANs Made Simple
Published:
🤔 Confused about the recent KAN: Kolmogorov-Arnold Networks? I was too, so here’s a minimal explanation that makes it easy to see the difference between KANs and multi-layer perceptrons (MLPs).
⛰️ The Energy Landscape of Predictive Coding Networks
Published:
📖 TL;DR: Predictive coding makes the loss landscape of feedforward neural networks more benign and robust to vanishing gradients.
💥 Thermodynamic Natural Gradient Descent
Published:
I recently came across this paper Thermodynamic Natural Gradient Descent by Normal Computing. I found it very interesting, so below is my brief take on it.
2023
🧠 Predictive Coding as a 2nd-Order Method
Published:
📖 TL;DR: Predictive coding implicitly performs a 2nd-order weight update via 1st-order (gradient) updates on neurons that in some cases allow it to converge faster than backpropagation with standard stochastic gradient descent.