Notes on projects, technical problems, and things I find interesting.
OpenAI's Lean formalisation of their AI proof shows that the code compiles, but it does not show that their natural language argument is correct. AI proofs are useful as guides, yet people must still read and verify them, especially when the text is long (and convoluted).
Read article >
An upcoming research project exploring if a phenomenon similar to human P300, where a surprising sequence of events causes a positive event-related potential measured in EEG around 300 ms after the fact, is translated inside LLMs.
Read article >
A retrospective on Yann LeCun's famous 2023 speech at MIT on Objective-Driven AI, and how AI could turn out to be dangerous even if we do not manage to make it reliable.
Read article >
Preparing for the upcoming NASA November Space Apps Hackaton
Read article >
How we tackled CUHK-X's Multimodal Action Recognition competition with synthetic generations and a multimodal pipeline for human action recognition. A description of our deep learning solution.
Read article >
An experiment where synthetic depth camera data is generated from a Blender scene, where we attempt to model the imperfect camera sensor in and replicate the disturbances which make depth data unique.
Read article >
In 1931 Gödel published his historical papers on completeness of formal languages with a painful construction of diagonalization. Nowadays, we have enough formal languages to fill the ocean. Can we replicate his efforts in Lean?
Read article >
How a small extension and free Edge Text-To-Speech is helping me learn Mandarin through spaced repetition, and some thoughts on Nielsen's famous piece on "Augmenting Long-term Memory".
Read article >
In polysomnography, we use EEG continuous recordings overnight to study the quality and physiology of sleep. Usually, a technician continuously monitors the recordings. Can we use transformers and more in general attention models and attention to predict sleep apnoea attacks frequency, and moreover to impute attacks to small windows during the night?
Read article >
A finite element model of the left ventricle that explores how myocardial fiber orientation, active contraction, and internal pressure interact to produce cardiac deformation and twist.
Read article >
In 1982, Hopfield published his historical paper "Neural networks and physical systems with emergent collective computational abilities" in which he showed how we can build associative, distributed memories through a self-organizing process. In this article, we analyze a C++ implementation of Hopfield networks, and study their dynamics.
Read article >
Our experience with the Italian Collegiate Programming Competition and a presentation of our @Politecnico di Milano team!
Read article >
Restricted Boltzmann machines, heavily popularized by Geoffrey Hinton, were an early model of distribution modelling inspired by the spin-glass model of statistical physics.
Read article >
A small domain-specific language for turning messy hydrological and weather data into clean datasets, with explicit rules for parsing, typing, preprocessing, normalization, and reproducible data preparation.
Read article >
A look at how RiverFlow aligns heterogeneous time series, builds sliding input-target windows, and creates reproducible training and test datasets for forecasting river levels from weather and recent history.
Read article >
A C++ implementation of graph-based automatic differentiation for neural networks, exploring computational graphs, forward evaluation, reverse-mode gradient propagation and the mechanics behind backpropagation.
Read article >
An exploration of reservoir computing in C++, showing how a fixed random recurrent network can encode temporal history and complex dynamics while only a simple readout needs to be trained.
Read article >
A C++ implementation of Kohonen self-organising maps: biologically inspired competitive neuron learning and neighbourhood updates to project high-dimensional data onto a topology-preserving two-dimensional representation.
Read article >
A computer vision pipeline for recovering a speed climber’s position and vertical velocity from uncalibrated video using hold detection, pose estimation, wall geometry, homographies, and Kalman filters
Read article >
An exploration of the historical Viola–Jones face detector: how small Haar-like features, integral images, AdaBoost and cascaded classifiers came together for efficient face detection before deep learning.
Read article >
An object-oriented guide for Fourier transforms in C++, exploring sampling, windowing, the DFT and FFT, spectral resolution, leakage, and the practical details behind frequency analysis.
Read article >
A lightweight approach to detecting email footers line by line, combining regex-based structural features with a small recurrent neural network to distinguish signatures from ordinary decoded IMAP message content.
Read article >