The paper “Process-Tensor Tomography of SGD: Measuring Non-Markovian Memory via Back-Flow of Distinguishability” was published in the Proceedings of Machine Learning Research for AISTATS 2026.
The work models neural training as a multi-time process and introduces a practical diagnostic for observable non-Markovian memory in stochastic gradient descent. For MediaLab, it is a strong machine-learning research highlight linking optimisation dynamics, measurement and explainable behaviour of training processes.
- Authors: Vasileios Sevetlidis, George Pavlidis
- Venue: The 29th International Conference on Artificial Intelligence and Statistics (AISTATS 2026)
- Conference dates: 2-5 May 2026
- Open record: PMLR record
You may also like
-
TMLR 2026: prescriptive SVD-inspired attention via spectral energy retention
-
Towards Heritage World Models: from digital twins to predictive heritage systems
-
ICASSP 2026: representation-diverse self-supervision for bioacoustic learning
-
ICPRAM 2026: two-stage angular alignment for positive-unlabeled learning
-
IEEE Access 2026: interpretable Vision Transformers via SVDA