Align, then memorise: the dynamics of learning with feedback alignment*

Maria Refinetti, Stéphane d’Ascoli, Ruben Ohana, Sebastian Goldt

TL;DR

This paper introduces a theoretical framework for understanding Direct Feedback Alignment (DFA), a biologically plausible alternative to backpropagation. They reveal that DFA learning in shallow networks follows an "Align-then-Memorise" two-phase process, where weights first align with random feedback signals and then focus on fitting the data, breaking degeneracies in the loss landscape. This insight explains DFA's success in some architectures and its failure in others, particularly convolutional networks.

What Problem Does This Paper Address?

Training deep neural networks typically relies on backpropagation (BP), which, while effective, has limitations. BP requires symmetric weights for forward and backward passes, making it biologically implausible. Furthermore, its sequential nature hinders parallelization, which is crucial for training increasingly large networks. Direct Feedback Alignment (DFA) offers a potential solution by using random feedback weights, enabling parallel updates and avoiding the biological implausibility of BP. However, DFA has shown inconsistent performance, successfully training some state-of-the-art models like Transformers but notoriously failing with convolutional neural networks (CNNs). This divergence in performance has lacked a clear theoretical explanation, leaving a gap in our understanding of when and why DFA works.

Methods

The authors propose a theoretical framework for feedback alignment algorithms, focusing on DFA. Their approach involves:

  • Analytical Description of Shallow Networks: They derive equations of motion for order parameters in shallow, non-linear networks to analytically describe DFA's learning dynamics. This builds on existing work for BP in similar settings.
  • Two-Phase Learning Model: They identify and formalize an "Align-then-Memorise" learning process. The "alignment phase" involves weights adapting to align with the random feedback signals, approximating the true gradient. The "memorisation phase" then focuses on minimizing the loss function.
  • Degeneracy Breaking: They demonstrate that this two-step process leads to "degeneracy breaking," where DFA preferentially converges to solutions that maximize gradient alignment among multiple low-loss minima.
  • Deep Linear Networks Analysis: They extend their analysis to deep linear networks, uncovering the "conditioning of alignment matrices" as a key factor for gradient alignment. This framework helps understand the impact of data structure on DFA.
  • Numerical Experiments: They validate their theoretical findings with experiments on deep non-linear networks using standard vision datasets, observing the "Align-then-Memorise" process and degeneracy breaking.

Datasets

  • MNIST: Used for training and evaluating deep non-linear networks.
  • CIFAR10: Used for training and evaluating deep non-linear networks.
  • Synthetic Regression Task: Used to study the impact of data structure on alignment, with controlled correlations and variances in targets.

Key Insights and Ideas

  • "Align-then-Memorise" is the core learning dynamic: DFA doesn't just blindly approximate gradients; it first aligns its internal representations with the random feedback signals (alignment phase) before focusing on fitting the data (memorisation phase). This is a fundamental departure from the continuous gradient matching of BP.
  • Degeneracy Breaking is key to DFA's success: In loss landscapes with many equally good solutions, DFA's alignment phase acts as a "filter," guiding the network towards solutions that are not only low-loss but also well-aligned with the feedback mechanism. This explains why DFA can find specific solutions that BP might not.
  • Conditioning of Alignment Matrices is crucial for deep networks: In deeper linear networks, the "conditioning" (how well-behaved) of matrices formed by combining feedback and weight information dictates how well the gradients align. This provides a quantitative measure for understanding alignment.
  • Data structure significantly impacts alignment: The correlation and variance of input and target data can either help or hinder the alignment process, directly affecting DFA's performance. This is a critical insight for understanding why DFA might fail in certain scenarios.
  • Layer-wise "Align-then-Memorise" in deep networks: The "Align-then-Memorise" process occurs sequentially from the bottom layers to the top layers in deep networks, with earlier layers aligning first and later layers memorizing.

Key Results

  • Theoretical prediction of learning dynamics: The authors provide ODEs that accurately predict the test error dynamics of shallow networks trained with DFA, matching simulations.
  • Demonstration of "Align-then-Memorise" and "Degeneracy Breaking": Numerical experiments on MNIST and CIFAR10 confirm these phenomena in deep non-linear networks.
  • Explanation for DFA's failure in CNNs: The theoretical framework suggests that the constrained structure of convolutional layers makes it difficult to satisfy the weight alignment condition required for gradient alignment, thus explaining DFA's notorious failure in this domain.
  • Impact of data structure quantified: Experiments show that correlated or imbalanced output statistics (low $\alpha$ or $\beta$) hamper both weight and gradient alignment.
  • Layer-wise alignment dynamics visualized: Figure 7 shows that alignment occurs sequentially from lower to higher layers in deep networks.

Limitations and Future Work

  • Focus on linear and shallow networks for theory: While the theory is extended to deep networks, the core analytical derivations are for linear or shallow settings. Future work could explore more complex non-linear deep network dynamics.
  • Exploration of the interplay between DFA and loss landscape: The paper touches upon how degeneracy breaking affects the loss landscape but suggests more detailed exploration is needed.
  • Generalization of findings to other architectures: While CNNs are discussed, the paper could further investigate DFA's performance and limitations in other specialized architectures.
  • Practical implications for hyperparameter tuning: Understanding how data structure affects alignment could lead to better strategies for choosing feedback matrices or network architectures for specific tasks.
  • Fitting vs. Generalization: The paper notes that high gradient alignment at the end of training is not a sufficient condition for DFA to match BP's generalization performance, suggesting further research into the trade-offs between alignment and expressivity.

Summary updated

Abstract

Abstract Direct feedback alignment (DFA) is emerging as an efficient and biologically plausible alternative to backpropagation for training deep neural networks. Despite relying on random feedback weights for the backward pass, DFA successfully trains state-of-the-art models such as transformers. On the other hand, it notoriously fails to train convolutional networks. An understanding of the inner workings of DFA to explain these diverging results remains elusive. Here, we propose a theory of feedback alignment algorithms. We first show that learning in shallow networks proceeds in two steps: an alignment phase, where the model adapts its weights to align the approximate gradient with the true gradient of the loss function, is followed by a memorisation phase, where the model focuses on fitting the data. This two-step process has a degeneracy breaking effect: out of all the low-loss solutions in the landscape, a network trained with DFA naturally converges to the solution which maximises gradient alignment. We also identify a key quantity underlying alignment in deep linear networks: the conditioning of the alignment matrices . The latter enables a detailed understanding of the impact of data structure on alignment, and suggests a simple explanation for the well-known failure of DFA to train convolutional neural networks. Numerical experiments on MNIST and CIFAR10 clearly demonstrate degeneracy breaking in deep non-linear networks and show that the align-then-memorize process occurs sequentially from the bottom layers of the network to the top.
BibTeX
@inproceedings{refinetti2021align,
  author = {Maria Refinetti and Stéphane d’Ascoli and Ruben Ohana and Sebastian Goldt},
  title = {{Align, then memorise: the dynamics of learning with feedback alignment*}},
  booktitle = {International Conference on Machine Learning},
  year = {2021},
  url = {https://openreview.net/forum?id=ktYwwwXcEbF},
}
Figure