How are linear representations learned? Exact solutions to the dynamics of abstraction

• URL: http://arxiv.org/abs/2607.08843v1
• Date: Thu, 09 Jul 2026 18:04:25 GMT
• Title: How are linear representations learned? Exact solutions to the dynamics of abstraction
• Abstract summary: In artificial and biological neural networks, concepts are often encoded as consistent linear directions in representation space.<n>In deep learning, this idea is known as the linear representation hypothesis and underpins many interpretability and control methods.<n>Here, we develop a framework to study the alignment of concept directions during training - a process we call "abstraction"
• Score: 20.467582967702672
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: In artificial and biological neural networks, concepts are often encoded as consistent linear directions in representation space. In deep learning, this idea is known as the linear representation hypothesis and underpins many interpretability and control methods based on linear probes, from concept detection to activation steering. Yet while prior work has studied whether such directions should exist $\textit{after}$ training, the dynamics of how they emerge $\textit{during}$ training remain poorly understood. Here, we develop a framework to study the alignment of concept directions during training - a process we call "abstraction". In a minimal linear network setting, we obtain exact solutions for the full trajectory of abstraction. These solutions reveal key analytic principles governing abstraction: (i) data and target geometry jointly determine abstraction at the end-of-learning, (ii) abstraction improves with network depth, and (iii) initialization scale controls the maximum abstraction reached during training. Extending our theory to nonlinear networks, we analyze how the choice of nonlinearity affects abstraction dynamics: erf networks approximate the linear theory, while abstraction in ReLU networks depends less on target geometry and more on input geometry. Across both, we prove a striking attenuation law: both nonlinearities weaken abstraction in activations relative to preactivations. We find evidence for this law in open models (DINOv3, Gemma 4) and apply our theory to improve linear probe generalization in LLMs. Together, our results provide a dynamical theory of abstraction with implications for interpretability and control.

Related papers

• A mathematical theory for understanding when abstract representations emerge in neural networks [4.415536082342714]
  We show that abstract representations of latent variables are guaranteed to appear in the last hidden layer of feedforward nonlinear networks.<n>These representations reflect the structure of the desired outputs or the semantics of the input stimuli.
  arXiv  Detail & Related papers  (2025-10-10T19:30:57Z)
• Deep Learning Through A Telescoping Lens: A Simple Model Provides Empirical Insights On Grokking, Gradient Boosting & Beyond [61.18736646013446]
  In pursuit of a deeper understanding of its surprising behaviors, we investigate the utility of a simple yet accurate model of a trained neural network. Across three case studies, we illustrate how it can be applied to derive new empirical insights on a diverse range of prominent phenomena.
  arXiv  Detail & Related papers  (2024-10-31T22:54:34Z)
• From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks [47.13391046553908]
  In artificial networks, the effectiveness of these models relies on their ability to build task specific representation.<n>Prior studies highlight that different initializations can place networks in either a lazy regime, where representations remain static, or a rich/feature learning regime, where representations evolve dynamically.<n>These solutions capture the evolution of representations and the Neural Kernel across the spectrum from the rich to the lazy regimes.
  arXiv  Detail & Related papers  (2024-09-22T23:19:04Z)
• On the Trade-off Between Efficiency and Precision of Neural Abstraction [62.046646433536104]
  Neural abstractions have been recently introduced as formal approximations of complex, nonlinear dynamical models. We employ formal inductive synthesis procedures to generate neural abstractions that result in dynamical models with these semantics.
  arXiv  Detail & Related papers  (2023-07-28T13:22:32Z)
• The Law of Parsimony in Gradient Descent for Learning Deep Linear Networks [34.85235641812005]
  We reveal a surprising "law of parsimony" in the learning dynamics when the data possesses low-dimensional structures. This simplicity in learning dynamics could have significant implications for both efficient training and a better understanding of deep networks.
  arXiv  Detail & Related papers  (2023-06-01T21:24:53Z)
• Understanding Multi-phase Optimization Dynamics and Rich Nonlinear Behaviors of ReLU Networks [8.180184504355571]
  We conduct a theoretical characterization of the training process of two-layer ReLU network trained by Gradient Flow on a linearlyparable data. We reveal four different phases from the whole training process showing a general simplifying-to-complicating learning trend. Specific nonlinear behaviors can also be precisely identified captured theoretically, such as initial, saddle-plateau dynamics, condensation escape, changes of activation patterns with increasing complexity.
  arXiv  Detail & Related papers  (2023-05-21T14:08:34Z)
• Does Deep Learning Learn to Abstract? A Systematic Probing Framework [69.2366890742283]
  Abstraction is a desirable capability for deep learning models, which means to induce abstract concepts from concrete instances and flexibly apply them beyond the learning context. We introduce a systematic probing framework to explore the abstraction capability of deep learning models from a transferability perspective.
  arXiv  Detail & Related papers  (2023-02-23T12:50:02Z)
• Subquadratic Overparameterization for Shallow Neural Networks [60.721751363271146]
  We provide an analytical framework that allows us to adopt standard neural training strategies. We achieve the desiderata viaak-Lojasiewicz, smoothness, and standard assumptions.
  arXiv  Detail & Related papers  (2021-11-02T20:24:01Z)
• The Principles of Deep Learning Theory [19.33681537640272]
  This book develops an effective theory approach to understanding deep neural networks of practical relevance. We explain how these effectively-deep networks learn nontrivial representations from training. We show that the depth-to-width ratio governs the effective model complexity of the ensemble of trained networks.
  arXiv  Detail & Related papers  (2021-06-18T15:00:00Z)
• Developing Constrained Neural Units Over Time [81.19349325749037]
  This paper focuses on an alternative way of defining Neural Networks, that is different from the majority of existing approaches. The structure of the neural architecture is defined by means of a special class of constraints that are extended also to the interaction with data. The proposed theory is cast into the time domain, in which data are presented to the network in an ordered manner.
  arXiv  Detail & Related papers  (2020-09-01T09:07:25Z)
• Learning Connectivity of Neural Networks from a Topological Perspective [80.35103711638548]
  We propose a topological perspective to represent a network into a complete graph for analysis. By assigning learnable parameters to the edges which reflect the magnitude of connections, the learning process can be performed in a differentiable manner. This learning process is compatible with existing networks and owns adaptability to larger search spaces and different tasks.
  arXiv  Detail & Related papers  (2020-08-19T04:53:31Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.