PERSON DIRECTORY
Randall Balestriero
Randall Balestriero appears in 2 indexed conversations across Machine Learning Street Talk. This directory brings every appearance, source, TL;DR, digest, and transcript into one searchable feed.
Prof. Randall Balestriero - LLMs without pretraining and SSL
A randomly initialized 7-billion-parameter transformer matched a pretrained LoRA model on some narrow classification tasks using roughly 20,000 labeled samples, with stable training and no hyperparameter optimization.Pretraining remains necessary for generation, open-ended reasoning, and distribution shifts, making task drift the key deployment variable while multitask objectives could offer a cheaper middle ground.
Neural Networks Are Elastic Origami! [Prof. Randall Balestriero]
Deep networks’ “elastic origami” geometry exposes capacity, bias, and extrapolation levers hidden by accuracy curves.On CIFAR and broader vision settings, ordinary training eventually produced adversarial robustness without attacks, suggesting smaller models trained longer could deliver “smart FLOPs,” though batch normalization and weight decay may block the endpoint.