跳到正文
返回论文列表
cs.CE提交于 已译

可解释权重空间编程构建透明 Transformer 与大脑编码模型

Interpretable weight-space programming builds transparent transformers and brain encoding models

Yash Mehta · Chandan Singh · Weijia Xu · Richard J. Antonello · Adeen Flinker · Nima Mesgarani · et al.

中文摘要

神经网络是强大的预测器,但其不透明性限制了其在以解释而非单纯预测为目标的科学中的实用性。在神经科学中,深度编码模型能准确预测自然刺激诱发的大脑活动,却难以揭示哪些刺激特征驱动了这些预测。为此,提出可解释权重空间编程(iWSP):由 LLM 智能体在代码中显式编程权重与特征,迭代构建可解释模型,无需基于梯度的训练(反向传播)。由于没有反向传播调参,智能体必须借助先验知识并基于先前试验进行推理,以假设、手工构造并精炼人类可解释的计算,模拟人类科学家的工作流程。LLM 审计确认了手工模型的解释性。首先在算术与情感分类等标准基准上构建 Transformer 验证 iWSP,发现 LLM 智能体可直接构建紧凑且准确的网络。随后将 iWSP 应用于视觉、语言与听觉的神经编码模型:在视觉皮层,iWSP 的预测优于现有可解释基线,揭示早期/高级视觉区域的不同计算,并能泛化到未见数据;在语言 fMRI 中,iWSP 发现数据中的混杂因素;在听觉 ECoG 中,系统发现可解释的声学特征以预测神经活动。这些结果确立了 iWSP 作为对齐高预测性能与人类可解释科学发现的通用范式。

关键要点

  1. 01问题:深度编码模型预测准确但难以解释哪些刺激特征驱动预测,缺乏透明度。
  2. 02方法:LLM 智能体在代码中显式编程权重与特征,无需反向传播,迭代构建可解释模型。
  3. 03结果:在算术与情感分类基准上直接构建紧凑准确的 Transformer,LLM 审计确认可解释性。
  4. 04结果:视觉皮层预测优于可解释基线并泛化到未见数据,语言 fMRI 发现数据混杂,听觉 ECoG 发现可解释声学特征。

解读

尚无解读。

原始英文摘要

arXiv:2611.18114v1 Announce Type: cross Abstract: Neural networks are powerful predictors, yet their opacity limits their usefulness in science, where the goal is explanation rather than mere prediction. In neuroscience, for example, deep encoding models accurately predict brain activity in response to natural stimuli, but struggle to reveal which stimulus features drive those predictions. Here we introduce interpretable weight-space programming (iWSP), a method in which an LLM agent iteratively builds interpretable models by explicitly programming their weights and features in code, without gradient-based training (backpropagation). Without backpropagation to tune parameters, the agent must draw on its prior knowledge and reason over previous trials to hypothesize, hand-craft, and refine human-interpretable computations mirroring the workflow of a human scientist. LLM audits confirm the interpretability of the hand-crafted models. We first validate iWSP by tasking it with constructing transformers for canonical benchmarks such as arithmetic and sentiment classification, finding that LLM agents can successfully craft compact, accurate networks directly. Next, we apply iWSP to build neural encoding models across vision, language, and audition. In the visual cortex, iWSP outpredicts existing interpretable baselines, uncovers distinct computations across early/higher visual areas, and generalizes to unseen data. In language fMRI, iWSP surfaces a confound in the data and in auditory ECoG, it systematically discovers interpretable acoustic features that predict neural activity. Together, these results establish iWSP as a general paradigm for aligning high predictive performance with human-interpretable scientific discovery.

同方向论文 · cs.CE

查看全部 →

反向发现 · 这篇论文相关的工具与提示词

去交叉发现 →

关联工具 · 1

Hacker News · 首页78

AgentRun 开源发布:为 Agent 加上自改进与持久执行

Parcha Labs 开源 AgentRun,在 Claude Code、Codex、OpenCode、pi 等 Agent harness 之上加入自改进、持久执行和控制面,采用 Apache-2.0 许可。

推荐理由:原文给出自改进、持久执行与控制面的具体机制和成本案例,读者可据此判断它如何接入现有 Agent 工作流。