logbook.json

{
  "schema_version": "1.0",
  "title": "Reproduction: HiPPO Zoo — Explicit Memory Mechanisms for Interpretable State Space Models",
  "emoji": "🧮",
  "space_id": "snaykey/repro-qot-localization",
  "paper": {
    "arxiv_id": "2602.21340",
    "openreview_id": "nB0TrIRAs1"
  },
  "tags": [
    "icml2026-repro",
    "paper-nB0TrIRAs1"
  ],
  "updated_at": "2026-07-30T00:00:00+00:00",
  "root": {
    "slug": "index",
    "title": "Reproduction: HiPPO Zoo — Explicit Memory Mechanisms for Interpretable State Space Models",
    "file": "pages/index.md",
    "children": [
      {
        "slug": "executive-summary",
        "title": "Executive summary",
        "file": "pages/executive-summary/page.md",
        "children": []
      },
      {
        "slug": "claim-1-five-extensions",
        "title": "HiPPO Zoo introduces five explicit polynomial-based extensions to classical HiPPO memory that individually replicate capabilities usually attributed to opaque learned SSMs like Mamba: Volterra, Salience, Associative Memory, Multiscale, and Forecasting HiPPO (Section 3).",
        "file": "pages/claim-1-five-extensions/page.md",
        "children": []
      },
      {
        "slug": "claim-2-salience-hippo",
        "title": "Salience HiPPO implements adaptive, input-dependent memory allocation via a scalar measure-deformation mechanism (Equation 2) with a time-warping interpretation (Equation 3), demonstrated on a selective copying task with explicit history-measure visualizations (Section 3.2, Figure 2).",
        "file": "pages/claim-2-salience-hippo/page.md",
        "children": []
      },
      {
        "slug": "claim-3-volterra-hippo",
        "title": "Volterra HiPPO decomposes nonlinear dynamics into interpretable kernels parameterized in an orthogonal polynomial basis, validated by learning a quadratic Volterra system on the Wray-Green benchmark (Section 3.1, Figure 1B).",
        "file": "pages/claim-3-volterra-hippo/page.md",
        "children": []
      },
      {
        "slug": "claim-4-assoc-memory-hippo",
        "title": "Associative Memory HiPPO separates temporal encoding from content-addressable storage on a continuous address space, demonstrated via key-value associative recall with explicit visualization (Section 3.3, Figure 3).",
        "file": "pages/claim-4-assoc-memory-hippo/page.md",
        "children": []
      },
      {
        "slug": "claim-5-multiscale-hippo",
        "title": "Multiscale HiPPO represents scale-dependent states as polynomial expansions (Equation 5) that support stable reconstruction across three orders of magnitude of timescale (Section 3.4, Figure 4).",
        "file": "pages/claim-5-multiscale-hippo/page.md",
        "children": []
      },
      {
        "slug": "claim-6-streaming-online",
        "title": "In the streaming/online training setups, the Salience model uses truncated backpropagation through time with fixed HiPPO parameters and only the neural readout learned online, while the Associative model uses chunk-wise gradient computation with state detachment between chunks (Section C.2-C.3).",
        "file": "pages/claim-6-streaming-online/page.md",
        "children": []
      },
      {
        "slug": "conclusion",
        "title": "Conclusion",
        "file": "pages/conclusion/page.md",
        "children": []
      }
    ]
  },
  "revision": 1
}