{
  "id": 10749784,
  "title": "Neural networks as decision trees: an analytical framework for learning and neural selectivity",
  "url": "https://urgent.news/2026/09/29/neural-networks-as-decision-trees-an-analytical-framework-for",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-29T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.09.23.753718v1?rss=1"
  },
  "original_language": "en",
  "account": "Neural networks utilize piecewise-linear functions to process information, but how their decision-making geometry develops in response to learning tasks is not fully understood. This study introduces a framework that connects a network's learning process, the structure of its activation regions, and the selectivity of individual neurons. At a point where the network's learning process is stable, the nonlinear network breaks down into numerous local linear regression models, each operating within a specific region of activation. Any discrepancies from the ideal least-squares solution are bound by shared weights across these regions and virtually disappear when the network's error is minimal, resulting in a near-perfect approximation of a piecewise least-squares decomposition of the overall task. This geometric structure can be visualized as a decision tree, which the researchers confirm through numerical analysis by training decision trees to predict the network's activation patterns based on the input. The study also establishes a relationship between the active regions of neurons and the specific statistics of the task they process, leading to the classification of neural selectivity into separate groups. Empirical evidence from simulated networks and two real neural datasets supports these findings, demonstrating that the proposed predictions accurately mirror the observed patterns of selectivity. Lastly, the research reveals that neural baseline activity plays a crucial role in controlling the diversity of activation patterns, organizing networks along a spectrum ranging from simple, generalizing representations to intricate, expressive representations. These findings provide a comprehensive framework for understanding how nonlinear networks break down complex tasks into manageable local computations and how these computations can be observed through neural activity.",
  "summary": "Nonlinear neural networks develop structured internal representations, yet how their geometry is determined by the tasks being learned remains poorly understood. Here, we develop an analytical framework for piecewise-linear feedforward and recurrent networks that links learning, activation-region structure, and neural selectivity. We show that, at any stationary point of gradient-aligned…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}