OCD: Learning to Overfit with Conditional Diffusion Models

TL;DR

This paper introduces a novel method called OCD that uses a conditional diffusion model to dynamically learn and overfit neural network weights conditioned on input samples, enabling versatile applications across multiple domains.

Contribution

The paper proposes a diffusion-based approach to learn input-conditioned network weights, allowing flexible overfitting and ensemble creation for diverse tasks.

Findings

Effective across image classification, 3D reconstruction, tabular data, speech separation, and NLP.

Ensemble of networks improves performance.

Diffusion model modifies a single layer conditioned on input, activations, and output.

Abstract

We present a dynamic model in which the weights are conditioned on an input sample x and are learned to match those that would be obtained by finetuning a base model on x and its label y. This mapping between an input sample and network weights is approximated by a denoising diffusion model. The diffusion model we employ focuses on modifying a single layer of the base model and is conditioned on the input, activations, and output of this layer. Since the diffusion model is stochastic in nature, multiple initializations generate different networks, forming an ensemble, which leads to further improvements. Our experiments demonstrate the wide applicability of the method for image classification, 3D reconstruction, tabular data, speech separation, and natural language processing. Our code is available at https://github.com/ShaharLutatiPersonal/OCD