- Úvod
- Ústav
- Lidé
- Výzkum
- Aplikace
- Semináře a akce
- Knihovna
- Doktorské studium
- Kariéra
Společný seminář katedry teoretické informatiky a matematické logiky MFF UK a oddělění umělé inteligence Ústavu informatiky AV ČR
Čtvrtek 14:00 v posluchárně S8, Malostranské náměstí 2.
Letní semestr našeho semináře začneme ve čtvrtek 27. února.
Victor’s talk will show how Winner-takes-all training can effectively address ambiguous machine learning tasks, and how this approach connects to a quantization objective.
Jelle will in his talk investigate the connection between confidence and uncertainty and shine a spotlight on uncertainty calibration and ways to improve it during training, as well as the implications of uncalibrated uncertainty for practical applications.
Marcel will in his talk challenge the assumption that that noise in stochastic gradient descent is uncorrelated, showing that epoch-based training introduces anti-correlations over time that reduce weight variance in flat directions, with significant implications for neural network training.
Andreas will present his work on large language models’ ability to generalize to complex proofs and make human-like mistakes on arithmetic word problems, which uses a symbolic world-model framework that formalizes reasoning in terms of proof trees.
Antoni will show that Image autoregressive models, which recently outperformed diffusion models on image generation, are significantly more prone to privacy attacks
Thomas will talk about the role of incentives in machine learning through reinforcement learning and mechanism design principles, highlighting works on contextual bandits with manipulated contexts and reinforcement learning from human feedback with strategic preferences