Alessandro Achille Principal Applied Scientist at AWS

aachille at amazon.com

I am a Principal Applied Scientist at AWS Agentic AI, where I am science lead for the Fundamental Research team. I study information in machine learning and AI agents: when it is acquired, where it is stored and whether it can be unlearned, and how much time it saves an agent solving a new task. This work spans critical learning periods and the geometry of tasks in deep networks, some of the first machine-unlearning methods for deep networks, and, most recently, a theory of reasoning agents in which information is measured in time. It has contributed to AWS products including Rekognition Custom Labels, Titan Image Generator, and Amazon Quick.

I received my Ph.D. in Computer Science from UCLA, advised by Stefano Soatto, and before that an M.Sc. in pure mathematics (model theory and algebraic topology) at the Scuola Normale Superiore in Pisa.

I am currently co-teaching UCLA CS269 on the control of AI agents.

Research

Selected papers, grouped by topic. The full list is on Google Scholar.

Information Dynamics of Learning

Deep networks learn representations of their data that are invariant to nuisances and disentangled, without being trained to do so. To explain why, I introduced an information-theoretic framework for the learning dynamics of deep networks, and showed that invariance and disentanglement emerge from minimizing the information stored in the weights. The framework also shows that learning is irreversible: when information is acquired matters and has permanent effects. As a result, deep networks have critical learning periods analogous to biological organisms: a change in data distribution early in training can permanently impair them. The same view gives tasks a geometry, which we turned into model-selection and fine-tuning methods used in Amazon Rekognition Custom Labels.

  • A. Achille, S. Soatto
    Emergence of Invariance and Disentanglement in Deep Representations
    Journal of Machine Learning Research (JMLR), 2018
    [ paper, arXiv ]
  • A. Achille, M. Rovere, S. Soatto
    Critical Learning Periods in Deep Networks
    International Conference on Learning Representations (ICLR), 2019
    [ paper, arXiv ]
  • A. Achille, M. Lam, R. Tewari, A. Ravichandran, S. Maji, C. Fowlkes, S. Soatto, P. Perona
    Task2Vec: Task Embedding for Meta-Learning
    International Conference on Computer Vision (ICCV), 2019
    [ arXiv, code ]
  • M. Kleinman, A. Achille, S. Soatto
    Critical Learning Periods Emerge Even in Deep Linear Networks
    International Conference on Learning Representations (ICLR), 2024 · Spotlight
    [ paper, arXiv ]

Machine Unlearning

Responsible use of data includes being able to remove the effects of specific training data, such as private or protected content, from a model after it has been trained. For large models this cannot practically be done by retraining from scratch, and, because learning is irreversible, it is not trivial to do otherwise. We introduced some of the first methods to make deep networks forget specific training data, and designed models in which data stays removable by construction or is protected at generation time. In a PNAS paper, we surveyed the resulting methods and choices for AI model disgorgement. Measuring similarity between images is a key question for copyright. We developed an interpretable measure of conceptual similarity, and proved that no such measure can be canonical, which points to the broader question of semantic information.

  • A. Golatkar, A. Achille, S. Soatto
    Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networks
    Conference on Computer Vision and Pattern Recognition (CVPR), 2020
    [ arXiv, code ]
  • A. Golatkar, A. Achille, Y. X. Wang, A. Roth, M. Kearns, S. Soatto
    Mixed Differential Privacy in Computer Vision
    Conference on Computer Vision and Pattern Recognition (CVPR), 2022 · Oral
    [ paper, arXiv, blog ]
  • A. Golatkar, A. Achille, L. Zancato, Y. X. Wang, A. Swaminathan, S. Soatto
    CPR: Retrieval Augmented Generation for Copyright Protection
    Conference on Computer Vision and Pattern Recognition (CVPR), 2024
    [ paper ]
  • A. Achille, M. Kearns, C. Klingenberg, S. Soatto
    AI Model Disgorgement: Methods and Choices
    Proceedings of the National Academy of Sciences (PNAS), 2024
    [ paper, press ]
  • A. Achille, G. Ver Steeg, T. Y. Liu, M. Trager, C. Klingenberg, S. Soatto
    Interpretable Measures of Conceptual Similarity by Complexity-Constrained Descriptive Auto-Encoding
    Conference on Computer Vision and Pattern Recognition (CVPR), 2024
    [ paper, arXiv, blog ]

Information in AI Agents and Transductive Learning

In inductive learning, a model is trained on past data and then applied in the same way to every future input; the role of information is to reduce uncertainty. AI agents are instead transductive: they reason about each specific input and construct a procedure to solve it. For such agents, information saves time: in verifiable tasks, the speed-up from past experience is exponential in the algorithmic information it shares with the solution. My current work builds on this view of information as time, and its implications for control of agents and economic-aware execution. These ideas are implemented in the open-source Strands AI Functions library.

  • A. Achille, S. Soatto
    AI Agents as Universal Task Solvers
    Entropy, 2026
    [ paper ]
  • R. Zabounidis, A. Golatkar, M. Kleinman, A. Achille, W. Xia, S. Soatto
    Re-FORC: Adaptive Reward Prediction for Efficient Chain-of-Thought Reasoning
    International Conference on Machine Learning (ICML), 2026
    [ paper ]
  • L. Zancato, A. Seshadri, Y. Dukler, A. Golatkar, Y. Shen, B. Bowman, M. Trager, A. Achille, S. Soatto
    B'MOJO: Hybrid State Space Realizations of Foundation Models with Eidetic and Fading Memory
    Neural Information Processing Systems (NeurIPS), 2024
    [ paper ]
  • Strands AI Functions
    Open-source library for transductive coding, economics-aware execution, and agentic learning, 2026
    [ code ]

Earlier work in mathematics

  • A. Achille, A. Berarducci
    A Vietoris-Smale Mapping Theorem for the Homotopy of Hyperdefinable Sets
    Selecta Mathematica, 2018
    [ paper, arXiv ]
  • M.Sc. thesis on definable groups in o-minimal and NIP theories
    Scuola Normale Superiore, 2015
    [ pdf ]

Teaching

Recent talks

  • UCLA Science Hub, 2026
  • MIT Brains, Minds and Machines Summer School, 2026
  • International Conference on Neuro-Symbolic Systems (NeuS), 2026
  • Symposium on Algorithmic Information Theory and Applications, 2026
  • DARPA expMath, 2026

Writing and media