Neural networks show distinct geometric patterns during training

Representations Have Trajectories: Characterizing Temporal Representation Dynamics in Neural Networks

Machine Learning

Summary

Deep neural networks change the shape of the way they represent information as they learn, following a few common paths. Sometimes the representations become very simple and then recover, sometimes they stay complex, and sometimes they collapse and don’t get better. The researchers studied these patterns over time during training to understand how the network’s internal picture evolves and affects learning. These findings help us see deep learning as a dynamic process rather than a static result.

What this means in practice

  • For machine learning engineers: Identify and address representation collapse issues during training to improve model performance and stability.
  • For ai system designers: Design architectures with awareness of geometric training trajectories to optimize learning outcomes and avoid pathological collapse.

Authors

Shivaprasad B. Gowda, Sparsh Mittal

Abstract

Recent work has shown that deep neural networks exhibit an implicit low-rank simplicity bias. In this paper, we investigate the questions that naturally arise from this observation. What are the temporal consequences of low-rank simplicity bias during optimization? How does representation geometry evolve throughout optimization rather than only at convergence? Do different neural architectures exhibit distinct geometric optimization behaviors? To answer these questions, we conduct a large-scale empirical study tracking representational geometry throughout training. We find that neural networks follow three recurring geometric trajectories: persistent high-rank, collapse-and-recovery, and persistent low-rank. We further identify Destructive Representation Collapse (DRC) as a pathological regime in which representations collapse to rank-one and fail to recover, leading to a sustained loss of representational capacity. While collapse-and-recovery is consistently associated with successful optimization in controlled settings, experiments on real-world models show that the relationship between geometric trajectories and optimization outcomes is configuration dependent. Overall, our findings suggest that representation geometry evolves dynamically during training and provide a new perspective on how deep neural networks learn.