The Impact of Geometric Complexity on Neural Collapse in Transfer Learning

Many of the recent remarkable advances in computer vision and language models can be attributed to the success of transfer learning via the pre-training of large foundation models. However, a theoretical framework which explains this empirical success is incomplete and remains an active area of research. Flatness of the loss surface and neural collapse have recently emerged as useful pre-training metrics which shed light on the implicit biases underlying pre-training. In this paper, we explore the geometric complexity of a model's learned representations as a fundamental mechanism that relates these two concepts. We show through experiments and theory that mechanisms which affect the geometric complexity of the pre-trained network also influence the neural collapse. Furthermore, we show how this effect of the geometric complexity generalizes to the neural collapse of new classes as well, thus encouraging better performance on downstream tasks, particularly in the few-shot setting.

Paper

Similar papers

Peer review

Reviewer pQCH5/10 · confidence 4/52024-07-06

Summary

The paper studies the relationship between geometric complexity (based on penultimate layer features) and the neural collapse phenomenon (especially, the variability collapse property). The main theoretical result is a bound on the CDNV-based NC metric using geometric complexity (Proposition 4.1). This is then employed to extend generalization and transfer learning bounds (proposed by previous efforts) to rely on geometric complexity. Experimental results on classification tasks (along with transfer learning settings) are presented to validate the proposed relationship.

Strengths

The unified analysis of neural collapse and geometric complexity presents an interesting line of research for the community. Especially, in terms of studying seemingly disjoint phenomena under a common lens.

Weaknesses

1. There seems to be a mismatch in the $\log(2/\delta)$ term in Proposition 4.2 vs $\log(1/\delta)$ term in Proposition 4.3, which makes the main results inconsistent (yet fixable). Additionally, why do we have an "expectation" over the "error" term in the LHS of Eq (10)? Isn't the "error" itself an expectation? 2. Although the paper presents generalization bounds, a discussion on the generalization performance of the classifier (VGG-13 for Figure 1) seems to be missing. Only the neural collapse and geometric complexity plots are illustrated but a deeper relationship between the trends and the test performance is lacking. Is there an upper limit of "good" learning rates, and batch sizes above which the geometric complexity can reduce but the test performance can get worse? 3. Similar to the above point, currently it seems like increasing the learning rate, increasing batch size, and increasing the regularization parameter value results in lower neural collapse values. Ideally, neural collapse is applicable during the terminal phase of training (TPT) where the training accuracy is $1$. Does this mean that all the experiments (at least for Figure 1) lead to TPT during training? nit: Please fix the notation for the norms as it is unclear if $||.||$ represents $||.||_2$ or $||.||_F$.

Questions

1. Since the geometric complexity-based generalization/transfer learning bounds are the upper bounds of existing neural collapse-based bounds, what can be a potentially new implication of your results? since it is common practice to do a hyper-parameter search over learning rates, batch sizes, and regularization parameter values. 2. The paper does not discuss any explicit regularization mechanism that aims to reduce the geometric complexity. What are the potential challenges in designing such a regularization mechanism?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

N/A

Reviewer N73X5/10 · confidence 4/52024-07-13

Summary

This paper examines the relationship between geometric complexity (GC), neural collapse (NC), and transfer learning performance in deep neural networks. The key contributions are: * Proposing geometric complexity as a measure that connects the flatness of the loss surface and neural collapse * Deriving theoretical bounds showing GC controls NC under certain assumptions * Empirically demonstrating that mechanisms affecting GC (learning rate, batch size, regularization) also influence NC * Showing lower GC during pre-training leads to better transfer learning, especially for few-shot tasks * Proving a new generalization bound in terms of GC * Demonstrating GC can be efficiently estimated on small samples The authors argue that GC can serve as an informative proxy for transfer learning potential and provide a unifying framework for understanding implicit biases in deep learning. By bridging concepts from optimization geometry, representation learning, and transfer learning, the paper aims to provide deeper insights into the fundamental mechanisms driving the success of modern deep learning techniques.

Strengths

**Originality**: The paper presents a novel perspective connecting several concepts in deep learning theory. Its original contribution is using geometric complexity as a bridge between loss surface geometry and neural collapse. This approach provides a fresh lens through which to view the implicit biases in deep learning, potentially unifying several strands of research. Using GC as a progress metric for transfer learning is innovative and could open up new avenues for improving pre-training techniques. **Quality**: The empirical evaluation thoroughly examines multiple datasets, architectures, and hyperparameter settings. The authors have made a commendable effort to validate their hypotheses across various conditions, including different network architectures (ResNet, VGG), datasets (CIFAR, MNIST, Fashion-MNIST), and transfer learning scenarios. The ablation studies exploring the impact of learning rate, batch size, and regularization on GC and NC provide valuable insights into the dynamics of deep network training. While limited in scope, the theoretical results give some formal grounding for the empirical observations and offer a starting point for further analytical work in this area. **Clarity**: The paper is well-structured and communicates the main ideas. The authors have presented a complex set of ideas in a logical flow, starting from the theoretical foundations and moving to empirical validation. The figures illustrate key results, particularly the relationships between GC, NC, and transfer performance. The use of color-coding and consistent formatting across figures aids in interpretation. The pseudo-code and detailed experimental setup descriptions in the appendix are beneficial for reproducibility. **Significance**: If the claims hold up, this work could provide valuable insights into the mechanisms behind transfer learning and self-supervised learning. The proposed geometric complexity measure may help analyze and improve deep learning models. The potential impact extends beyond just theoretical understanding. If GC can serve as a reliable proxy for transfer learning potential, it could guide the development of more effective pre-training strategies and architectures. Furthermore, the connection between GC and NC could help explain why certain training practices (like large batch training or specific learning rate schedules) are effective, potentially leading to more principled approaches to hyperparameter tuning.

Weaknesses

**Limited theoretical foundation**: While the paper provides some theoretical results, the assumptions required (e.g., Poincaré inequality) are quite strong and may not hold in practice for real datasets and architectures. The connection between the linear model analysis and deep neural networks is not strongly justified. This is a significant limitation, as it's unclear how much of the theory translates to practical scenarios. The authors could strengthen this aspect by providing empirical evidence that these assumptions hold in realistic settings or deriving weaker results under more general conditions. Additionally, the paper could benefit from a more thorough discussion of the implications and limitations of these theoretical results. **Experimental limitations**: The empirical evaluation focuses on image classification tasks and relatively small datasets. The generalizability of the results to larger datasets, other domains (e.g., NLP), or more complex architectures is unclear. This narrow focus raises questions about the broader applicability of the findings. For instance, it's unclear whether the relationships between GC, NC, and transfer performance would hold for large-scale pre-training scenarios like those used in modern language models. The paper would be significantly strengthened by including experiments in a broader range of tasks and scales or at least by providing a more detailed discussion of the potential challenges in scaling up the approach. **Lack of comparison to related metrics**: The paper does not thoroughly compare geometric complexity to other related complexity measures or generalization bounds from the literature. This makes it difficult to assess the relative merits of GC. There's a rich body of work on complexity measures for neural networks, including various forms of capacity measures, stability-based bounds, and PAC-Bayesian approaches. A comprehensive comparison would help situate GC within this broader context and clarify its unique contributions. Such a comparison could also help identify scenarios where GC might be particularly advantageous or potentially suboptimal compared to existing approaches. **Causality vs correlation**: While the paper shows correlations between GC, NC, and transfer performance, it does not conclusively demonstrate a causal relationship. Alternative explanations for the observed trends have not been thoroughly explored. This is a critical limitation, as it leaves the possibility that GC is merely a proxy for some other underlying factor driving NC and transfer performance. The paper would be strengthened by a more rigorous causal analysis, perhaps through interventional studies or by controlling for potential confounding variables. Additionally, discussing potential mechanisms by which GC might causally influence NC and transfer learning would add depth to the analysis.

Questions

1. How does geometric complexity compare to other measures of model/representation complexity like Fisher information, gradient norm, etc.? A more thorough comparison would strengthen the paper. 2. The theoretical results rely on strong assumptions like the Poincaré inequality. Can you provide empirical evidence that these assumptions hold for real datasets/models? Or derive results under weaker assumptions? 3. The paper shows correlations between GC, NC, and transfer performance but does not conclusively prove causation. Have you considered alternative explanations or confounding factors? 4. How computationally expensive is it to measure GC during training? Is it feasible to use it as a regularization term or early stopping criterion in practice? 5. The generalization bound in terms of GC is interesting, but how tight is it empirically? How does it compare to other generalization bounds in the literature? 6. Have you explored using insights about GC to guide architecture design or improve training algorithms? Demonstrating practical benefits would strengthen the paper.

Rating

5

Confidence

4

Soundness

2

Presentation

2

Contribution

2

Limitations

The authors acknowledge some limitations, like the focus on image classification tasks and relatively small datasets. However, they could go further in discussing potential shortcomings: * The strong assumptions required for the theoretical results may limit their applicability. * The computational cost of measuring GC is not thoroughly addressed. * Potential negative consequences of optimizing for low GC are not explored.

Reviewer 5tN36/10 · confidence 3/52024-07-21

Summary

The paper explores the relationship between neural collapse (NC) and geometric complexity (GC). It presents both theoretical and empirical evidence showing that geometric complexity is a robust metric across various variables. By substituting the NC metric with GC, the paper introduces a generalization bound based on geometric complexity. Furthermore, it highlights the significance of GC in transfer learning, demonstrating that by controlling the pre-training GC, downstream NC can also be controlled, leading to improved transfer learning results.

Strengths

* The paper is overall well-written and easy to follow, the interpretation of each theoretical results is clear and intuitive. * The relationship between NC and GC is quite interesting. * Most of the empirical results align well with the theoretical part, demonstrating the validity of the results.

Weaknesses

Some empirical results can be better aligned with the theoretical counterpart and additional experiments could be added to better support the claims. For example: * In proposition 4.1, NC is upper bounded by geometric collapse up to some constant scaling. What's the range of this constant $c$, does it purely depend on $Q$? If yes, it would be better to directly plot NC and RHS in (6) in the same plot to demonstrate the validity of the bound; and if no, it would be interesting to show how different settings examined in Figure 1 affect $c$. * The empirical evidence presented in Figure 2 clearly demonstrates the robustness of geometric complexity. However, I have some concerns regarding the Lipschitz notion. Neural networks are known to be not so Lipschitz, and making them more Lipschitz is an active area of research aimed primarily at improving robustness. Is Proposition 4.2 numerically verifiable by calculating the Lipschitz constant of the tested neural networks?

Questions

* In equation (7), does the distance $d$ represent the norm distance? This should be stated clearly. * Both the paper and the prior work (Galanti et al, 2022) rely on the data assumption where the source and transfer datasets come from the same distribution $Q$. However, in practical settings, source and transfer datasets typically have domain gaps which can potentially violate the data assumption. Would the results in the paper still hold in this case? For example, pre-training on ImageNet/Cifar-100 while fine-tuning on DTD?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer pQCH2024-08-09

Thank you for the responses and also the additional experimental results. - In my opinion, the authors discussion on explicit GC regularization is key to improve the contribution. Although it is computationally inefficient, it is always helpful to have experiments which validate the claim in this paper. Based on the rebuttal PDF, it seems like NC indeed reduces when explicit GC regularization is employed. - I would request the authors to focus more on this aspect (techniques, limitations, tuning etc of such a regularizer) to further strengthen their paper. Simply showing correlations by tuning common hyper-parameters might not fully justify your work. I have increased the score. Good luck.

Authorsrebuttal2024-08-09

Thank you!

We are glad that we were able to address your concerns. And we agree, this in an important (and potentially useful) aspect, despite the inherent computational inefficiencies. Including these experiments has helped to improve the exposition. Thanks again for the suggestion!

Authorsrebuttal2024-08-12

Thank you again for your review. We hope that the concerns in your original review were sufficiently addressed in our response. In particular, the justification of the Poincare inequality assumption for data distributions, the demonstrated empirical tightness of the Generalization bound (our Propn 4.3), and the explicit GC regularization experiments exploring causality of these phenomena. We also hope we addressed the concerns regarding experimental limitations particularly with respect to the applicability to modern language models and LLMs. As the discussion period ends, we wanted to see if there is anything that needs additional clarification. Thank you.

Authorsrebuttal2024-08-12

Thank you again for your review. We hope that the concerns in your original review were sufficiently addressed in our response, particularly, the application of the Poincare inequality and the role of the constant c in our Propn 4.1. As the discussion period ends, we wanted to see if there is anything that needs additional clarification. Thank you.

Reviewer 5tN32024-08-13

Response to rebuttal

Thanks to the authors for the detailed response. Given that the authors said "c ~ 1000 (lower bound) for CIFAR-10 in our setting", would this make the bounds in Proposition 4.1 and 4.3 vacuous?

Authorsrebuttal2024-08-13

This is a good question but indeed, taking the value c near 1000 does not make the bounds in Proposition 4.1 and Proposition 4.3 vacuous. Figure 1 in our paper plots the Geometric Collapse (i.e. the RHS of the inequality in Propn 4.1 excepting this constant c) in the middle column and the Neural Collapse (i.e. the LHS of the inequality of Propn 4.1) in the right column. From these plots, we see that by comparing the magnitude of the Neural Collapse and the Geometric Collapse that indeed c ~ 1000 for CIFAR-10 in our setting. We also explored directly the tightness of the bound derived in Proposition 4.3 during the rebuttal. Please see the 1-page pdf which we attached to our rebuttal response. Figure 2 of that rebuttal pdf shows empirically the tightness of the bound in Propn 4.3. For that experiment we trained a VGG-13 model on CIFAR-10 with 5 random seeds. The plot shows the LHS of the inequality in Propn 4.3 (i.e. the nearest mean classifier error on the test set ) in blue and the RHS of the inequality in Propn 4.3 (excepting the Lipschitz term which we expect to be negligible due to Fig 2 and (8) of our paper) in orange. Note that here we take $c = 1000$, $p = 1024$ (which denotes the embedding dimension of the feature map) and $m_c = 5,000$ (which denotes the number of samples per class). Indeed, the bound is not vacuous and in fact appears quite tight! Thanks again for your careful and thoughtful read. I hope we've been able to successfully address your concerns.

Reviewer 5tN32024-08-13

Thanks, that has addressed my concern. Raising the score now.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC