Abstract
Neural network weights are typically viewed as the end product of training, while most deep learning research focuses on data, features, and architectures. However, recent advances show that the set of all possible weight values (weight space) itself contains rich structure: pretrained models form organized distributions, exhibit symmetries, and can be embedded, compared, or even generated. This emerging research direction, Weight Space Learning (WSL), treats neural weights as a meaningful domain for analysis and modeling. This survey provides the first unified taxonomy of WSL, categorizing existing methods into Weight Space Understanding (WSU), Weight Space Representation (WSR), and Weight Space Generation (WSG).
Introduction
In the traditional deep learning paradigm, weights are considered the static output of an optimization process. While research has historically prioritized input data and output predictions, the weights themselves are often treated as a black box. However, the proliferation of "model zoos"—repositories such as Hugging Face containing millions of pretrained models—has introduced a new challenge: the systematic organization, comparison, and reuse of this vast collective knowledge. Weight space is not a random collection of parameters; it possesses a deep mathematical structure governed by symmetries and optimization dynamics. This survey consolidates fragmented research into the unified discipline of Weight Space Learning (WSL).
Core Dimensions of Weight Space Learning
Weight Space Understanding (WSU)
WSU focuses on the underlying "physics" of weight space, specifically studying Invariance (different weight configurations producing identical functions) and Equivariance (predictable output changes resulting from weight transformations). Key techniques involve the use of quotient spaces to eliminate redundancies caused by symmetries, such as neuron permutations.
Weight Space Representation (WSR)
WSR involves learning latent embeddings of models for downstream analysis.
- Model-based approaches: These utilize specialized architectures like Graph Neural Networks (GNNs) or CNNs to process weight tensors directly as structured data.
- Model-free approaches (probing): These bypass raw weights to create a "behavioral signature" by observing model responses to specific input sets.
Weight Space Generation (WSG)
WSG moves from analysis to the synthesis of new models.
- Hypernetworks: Small auxiliary networks that output weights for a larger "target" network in a single forward pass.
- Generative Models: The application of Diffusion models or GANs to sample new, functional weight configurations from a learned distribution of pretrained models.
Datasets and Benchmarks
The primary data source for WSL is the Model Zoo. Key benchmarks include:
- MLP/CNN Zoos: Collections of thousands of small networks trained on standard datasets like MNIST or CIFAR-10.
- Implicit-Zoo: Focuses on Implicit Neural Representations (INRs) where weights represent signals such as images or 3D shapes.
- LoRA Zoos: Modern benchmarks storing Low-Rank Adaptation weights for large-scale models like Stable Diffusion or Large Language Models (LLMs).
Key Insights: Weights as a First-Class Modality
A central insight of WSL is that weights are "learners of data," encoding architectural inductive biases and optimization trajectories. A significant challenge is the Permutation Problem, where swapping neurons in a hidden layer preserves functionality but alters weight values. WSU is critical for "aligning" models to enable fair comparison. Furthermore, architecture-agnostic probing allows for the comparison of functionally similar models, such as CNNs and Transformers, within a shared latent space.
Practical Applications and Results
WSL enables several novel capabilities:
- Model Retrieval: Identifying the optimal pretrained model for a new task via embedding space search.
- Zero-Shot Prediction: Meta-models predicting test accuracy by analyzing weights without access to test data.
- Training Acceleration: Utilizing weight generators (e.g., GHN-3) to produce near-optimal initial weights, significantly reducing training time.
- Data-Free Reconstruction: Reconstructing training data from model weights, which has significant implications for privacy research.
Limitations and Future Directions
Current WSL methods face a Scaling Wall, as processing the billions of weights in modern LLMs is computationally prohibitive. Modular processing of layers or LoRA adapters is a proposed path forward. Additionally, real-world features like Batch Normalization break simple mathematical symmetries, complicating theoretical analysis. Future research aims to develop "Universal Weight Processors" and address the adversarial risks associated with generated models.
Summary updated