Can Model Weights Serve as the Next Training Data for AI?

Motivation

Publicly available high‑quality training data are approaching saturation, causing the marginal returns of scaling laws to decline. Researchers therefore seek alternatives to raw data, such as synthetic data generation, reallocation of inference compute, and reuse of already‑trained models.

Weight‑Space Learning

The paradigm, introduced by Damian Borth’s group in 2020 and formalized in a 2021 NeurIPS paper, treats the parameters of a completed neural network as an independent data modality. By encoding weight tensors into a low‑dimensional latent space, a learner can (i) predict downstream performance from the latent representation and (ii) generate new weight tensors by sampling that space.

Key Publications and Experiments

2020 – “Investigation of the Weight Space” observed that weights shift slightly each training iteration, indicating a lack of stable reference points for individual layers.[1-1]

2021 – “Self‑Supervised Representation Learning” trained an auto‑encoder on a library of miniature networks. The encoder compressed each network’s weights into a latent vector; a linear regression head on the latent vector predicted the network’s test accuracy. This demonstrated that weight tensors contain learnable structure.[1-2]

The approach was inspired by Google Research’s “Predicting Neural Network Accuracy from Weights”. Borth’s team replaced hand‑crafted features with an end‑to‑end encoder that jointly compresses multiple weight sets and reconstructs them, confirming a shared latent structure across models.[1-3]

Experiments used 600 networks for training and 300 distinct networks for testing. The resulting R² score between predicted and true accuracies exceeded that of prior hand‑crafted baselines.[1-2]

Community Development

Since 2023 the field has converged on common benchmarks and terminology. The first Weight‑Space Learning workshop was held in 2024, and ICML introduced a dedicated session on weight symmetry. Related work includes Haggai Maron’s 2023 paper “Equivariant Architectures for Learning in Deep Weight Spaces” and Anthropic’s 2021 “Mathematical Framework for Transformer Circuits” and 2022 “Toy Models of Superposition”, which also treat neural‑network artifacts as research objects.[1-4][1-5][1-6]

Challenges Identified in 2022

Generation quality – sampled latent vectors often decode to unstable or low‑performing models.

Geometric properties – the curvature and topology of the weight manifold remain poorly understood.

Training data – assembling a diverse, high‑quality collection of weight tensors for the learner is non‑trivial.

Outlook

Weight‑space learning establishes a “model‑as‑data” pipeline that could reshape deep‑learning production paradigms, provided the above obstacles are addressed.

Code example

① 2020 年，Borth 团队在 「Investigation of the Weight Space」 论文中，提出神经网络的权重在训练过程中每一轮迭代都会发生微小偏移，导致局部层面缺乏稳定的参照点。[1-1]
3、在 2021 年，Borth 团队发表论文「Self-Supervised Representation Learning」 验证了「把神经网络权重作为可学习模态」的可行性。通过使用自编码器在微型网络上实现稳定压缩，潜向量可通过一个线性回归头预测模型准确率。[1-2]
① 该工作受到 Google Research 团队「Predicting Neural Network Accuracy from Weights」工作的启发。Borth 团队采用端到端的表示学习替代手工特征。他们设计了一个自编码器架构，将多个神经网络的权重序列编码至低维潜空间后再重构，以此验证权重之间是否存在可被学习的共享结构。[1-3]