Janis AIAD
  • about
  • blog (current)
  • research
  • cv
  • submenus
    books
    teaching and exercises
  • Hierarchical three-layer learning Three-Layer Learning as a DMFT–RMT–BBP Closure Problem

    A master reduction from exact Hermite geometry to a conditional dynamic BBP law, with cavity, replica, spectral AMP, Kac–Rice, and a time-resolved Hessian diagnostic.

    11 min read   ·   July 26, 2026

    2026   ·   hierarchical-learning   dmft   bbp   random-matrix   muon   three-layer   hermite   cavity   amp   kac-rice   ·   blog

  • Hierarchical three-layer learning Three Layers, One Hierarchy: From Exact Hermite Dynamics to DMFT and BBP

    A deterministic theory map for hierarchical three-layer learning, Muon, dynamical mean-field theory, and block-Wishart spectral transitions.

    14 min read   ·   July 22, 2026

    2026   ·   muon   hierarchical-learning   hermite-dmft   bbp   random-matrix   three-layer   ·   blog

  • Hierarchical three-layer learning Does Muon Help a Hierarchical Three-Layer Model?

    A twenty-pair confirmation of endpoint risk and sector clocks, plus sample-size and fresh-Hessian diagnostics.

    19 min read   ·   July 22, 2026

    2026   ·   muon   hierarchical-learning   empirical   scaling-laws   bbp   three-layer   ·   blog

  • Spectral transitions in multi-index models When Both Layers Learn: Where Hessian Outliers Come From

    A non-technical guide to joint feature-amplitude learning, width, Hessian branches, and dynamic BBP transitions.

    13 min read   ·   July 14, 2026

    2026   ·   phase-retrieval   bbp   hessian   random-matrix   multi-index   ·   blog

  • Spectral transitions in multi-index models Beyond Quadratic Phase Retrieval: Spectral Learning with General Gaussian Links

    Why the finite-state and spectral picture extends from quadratic phase retrieval to smooth multi-index models.

    9 min read   ·   July 14, 2026

    2026   ·   hermite-dmft   multi-index   random-matrix   bbp   phase-retrieval   ·   blog

  • Transformers for In-Context PDE Solving and Inverse Problems transformers for inverse problems on PDEs, Part 6: How Attention Learns a Preconditioner

    Which training curves are exact, which are local diagnostics, and how query-key-value matrices become a learned solver step.

    7 min read   ·   July 06, 2026

    2026   ·   transformers   optimization   preconditioning   ·   research

  • Muon and controlled training dynamics Muon for Phase Retrieval III: What Is Proved and What Remains Open

    A reader-friendly map of the exact results, imported theorems, numerical evidence, and open probabilistic bridge.

    9 min read   ·   July 06, 2026

    2026   ·   muon   proof   random-matrix   learning-curves   spectral-transitions   ·   blog

  • Transformers for In-Context PDE Solving and Inverse Problems transformers for inverse problems on PDEs, Part 5: Separating Encoder, Decoder, and Generalization Error

    A readable error budget for task inference, solver depth, training tasks, and held-out generalization.

    8 min read   ·   June 23, 2026

    2026   ·   transformers   optimization   generalization   ·   research

  • Muon and controlled training dynamics Muon for Phase Retrieval II: Choosing the Power During Training

    A non-technical guide to adapting Muon's spectral power as strong and weak hidden directions are learned.

    10 min read   ·   June 14, 2026

    2026   ·   muon   phase-retrieval   optimal-control   spectral-optimization   ·   blog

  • Transformers for In-Context PDE Solving and Inverse Problems transformers for inverse problems on PDEs, Part 4: What Can Be Proved About the Solver?

    The finite-dimensional encoder and decoder certificates, and why a full training theory needs replica order parameters.

    9 min read   ·   June 10, 2026

    2026   ·   transformers   proofs   generalization   ·   research

  • Transformers for In-Context PDE Solving and Inverse Problems transformers for inverse problems on PDEs, Part 3: Turning PDE Solutions Into Transformer Tokens

    A plain-language pipeline: functions become vectors, prompts become weak equations, and decoder layers become solver steps.

    8 min read   ·   May 27, 2026

    2026   ·   transformers   pde   methodology   ·   research

  • Spectral transitions in multi-index models When Does the Hessian Reveal a Hidden Direction?

    A non-technical introduction to dynamic BBP transitions and energy-resolved Hessian spectra in multi-index phase retrieval.

    9 min read   ·   May 23, 2026

    2026   ·   phase-retrieval   bbp   hessian   random-matrix   kac-rice   ·   blog

  • <
  • 1
  • 2
  • >
© Copyright 2026 Janis HERAN AIAD. If you are nightly, put the dark mode on by clicking on the moon icon in the top right corner.