NeurIPS 2020

Deep learning versus kernel learning: an empirical study of loss landscape geometry and the time evolution of the Neural Tangent Kernel


Review 1

Summary and Contributions: The paper provides experimental evidence on CIFAR-10 and CIFAR-100 that the deep network determines the final basin of low loss within an epoch, and NTK corresponding to a trained network (only 10% of total training) can outperform full network training.

Strengths: - The paper defined numerous measurement metrics as tools to study the landscape. - The definitions in this paper are clear. The visualizations are good. - The topic is clearly relevant to the NeurIPS community.

Weaknesses: - The figures end with epoch 10^2, which is not the end of the training as mentioned in the appendix. Clearly the training error is not 0 at epoch 100 as shown in the figure, which means the training is not finished. This makes me doubt the conclusion that NTK corresponding to trained network can outperform full network training. - The paper only runs tests on CIFAR-10 and CIFAR-100. These two datasets are all image datasets and pretty similar. The results on other datasets (e.g., language dataset) can be very different. - The paper cannot draw any solid quantitative conclusions. The qualitative phenomenon has been known.

Correctness: It seems the neural network is not trained to have zero training error in the paper.

Clarity: The paper is overall well written.

Relation to Prior Work: The paper gives enough related works.

Reproducibility: Yes

Additional Feedback: Why the figures only show the first 100 epoch? ---------- post author response comments -------- The authors have addressed my main concerns of this paper (misleading information in Figure 6 and related section). I highly suggest that the authors do a major modification, especially on the conclusion. In addition, I agree with other reviewers' opinions that this paper conveys an important message to people that are not aware of this phenomenon. Therefore, I increased my score and happy to see this paper appearing in NeurIPS if the authors modify the paper accordingly (which I assume they will).


Review 2

Summary and Contributions: This work gives a large-scale phenomenological analysis of training based on various metrics related to the loss landscape and data-dependent NTK. The experimental results clarify that training dynamics within an epoch determine the final basin. After the basin fate is determined and within 10 epochs, data-dependent NTK acquires sufficient information to decrease the training and test error as a Talyorized model.

Strengths: *This work gives a unified and more sophisticated visualization of various well-established metrics that have been considered independently. In particular, the sophisticated evaluation of the parent-child spawning reveals that the decision of basin fate happens within 1 epoch. The parent-child spawning also leverages the evaluation of other metrics. *As a novel metric, this work investigated the approximation accuracy of Taylorized models with data-dependent NTK. It reveals that this NTK acquires sufficient information within 10 epochs. *These empirical findings will give rich insight and suggestions for further developing the theory of loss landscape and NTK.

Weaknesses: *The method of parent-child spawning [6] is not original to this paper. Besides, as remarked by authors, the early decision of final minima has also been argued by some other studies [9,10]. One can even easily imagine this phenomenon from the studies of mode connectivity [3,4]. Therefore, this paper's pure contribution seems limited to the evaluation of Talyerized models (that is, Fig. 6). *Some ambiguous explanation of experimental settings may prevent readers from correctly understanding the main claim. I will discuss it in the "Clarity" part.

Correctness: I find no incorrent description or methodology in this work.

Clarity: In Fig. 6, the linearized model with the data-dependent NTK out-performed the original network in the final phase of training. This is interesting but seems non-trivial (or strange) in the following two points. (i) The rightmost point in each figure appears to be given at 90 or 95th epoch. The rightmost green point takes lower errors than the rightmost red point. What happens if you take both points very close to 100th epoch (e.g. 99.99 epoch, 100 epoch minus 1 "step" )? I guess that the green point should be close to red one because the parameter update (1) should be not so large in the final phase (due to the small gradient, or scale learning rate by hands or Adam) and the Taylor approximation (2) will become more accurate. It sounds highly non-trivial that the green point always out-performs red one even at epochs very close to 100. (ii) The same problem also appears in the blue points. If you take the blue point sufficiently closer to 100th epoch, the blue point will become closer to the red one because both will share almost the same parameter. It would be necessary to confirm the consistency of the obtained results in a more thoughtful way arournd 100th epoch. One possible approach is to depict figures corresponding to Fig. 6 with the x-axis of [100th epoch, 100th epoch - 1 step, 100th epoch - 2 step, ...]. I believe this unclear point is not a serious flaw. But, if this point keeps ambiguous after rebuttal, I might be forced to decrease my score.

Relation to Prior Work: As I remarked in the above, the parent-child spawning is an established idea [6]. It would be better to discuss the difference between the current work and [6] more carefully. Certainly, the previous work [6] would not have been dealing with the function space. However, one can see the basin fate just by seeing the parameter space (Fig. 5 (left two)). Fig. 4 (right far) suggests that the tangent place in the function space is insufficient to represent the error. In Figs 4 and 5, is there any positive advantage to use the evaluation in the functional space compared to the previous studies in the parameter space [3,4,6]?

Reproducibility: Yes

Additional Feedback: Minor issues *Visualization method of Fig. 5: I am not sure how the authors depict this paper. Is it based on PCA of trajectories? It is also unclear why linear lines give these trajectories. *Line 21 "a kernel machine can actually out-perform...": The term "kernel machine" is misleading because authors do not perform kernel regression. It is just a linear regression with the Taylorized model (2). More technically speaking, when we use data-dependent NTK in a linearized model, the positive definiteness of this NTK is non-trivial and the equivalence to the kernel regression becomes unclear. *The term "chaotic": Authors sometimes refer to the training process in the early times as ''chaotic'', but does this mean literally "chaos" as a technical term? Is there any evidence for an exponential sensitivity against the initial condition? *The batch size used in experiments seems not mentioned in the manuscript. Batch size is known as a key factor in determining the shape of global minima [24]. It would be better to add any discussion on the dependence of the obtained results on the batch size. * Line 213 and Section B: The escape threshold has already been discussed in the context of neural networks [LeCun et al. "Efficient backdrop", 1998]. In the context of modern deep learning, [Karakida et al. https://arxiv.org/abs/1806.01316] revisited the same quantity and obtained its generalized form for the training with momentum. It would be better to refer to them. *Authors showed some experiments without batch norm in Supplementary Material, but not mentioned in the main text. Is there any discussion on the effect of batch norm? Batch norm is reported as a key factor in making the loss landscape smoother ([Santurkar et al. https://arxiv.org/abs/1805.11604]). Line 381: Reference [33] is missing in the main text. ===After rebuttal=== Thank you for your kind reply. Only after reading Reviewer 1's comment and your reply, I realized that total epochs were not 100. Since the current Fig. 6 (i.e., comparison with NN of epoch 100) causes misunderstanding, I recommend the authors to replace this figure with a new one shown in the rebuttal (i.e., comparison up to epoch 200). This new figure is insightful enough as a measurement of the geometric structure of training. I agree with other reviewers that the authors should do a major modification on some explanations regarding Figure 6. For instance, "within 10 percent of training time, the learned kernel as a kernel machine can actually outperform full deep learning" (which appears in abstract and conclusion) is highly misleading. It does not outperform the NN model of epoch 200. Although there are not a few inadequate descriptions, they do not flaw the significance of this paper, and I keep my overall score. I am looking forward to seeing a fully revised version.


Review 3

Summary and Contributions: This paper empirically studies many quantities related to the neural tangent kernel, including the dynamics of the kernel, the loss landscape, etc. It is claimed that after roughly 1 epoch of chaotic change in the learned kernel, the final basin containing the end point of training is basically determined. Then using the kernel learned after roughly 10 percent of the total training time, we can achieve good training and test errors.

Strengths: I think this paper provides some insights. For example, the kernel becomes stable after a long enough training (Figures 2 and 3, C-E), and the learned kernel after only 10 percent of the total training time can already achieve pretty good training/test error (Figure 6).

Weaknesses: I think more details and discussion are needed to justify and evaluate some claims in the paper. 1. In Figure 5, it is said that "two children spawned at an early time t_s in the chaotic training regime arrive at two different loss basins", while "two children spawned at a later time t_s in the stable training regime arrive at the same loss basin". While I do agree with this claim, a simple explanation for it may just be that as t_s becomes larger, the distance between a child and its spawning parent becomes smaller and smaller. To see this, note that as the risk decreases, the gradient norm also becomes smaller, and the child is trained for T-t_s epochs, which also shrinks as t_s becomes larger since T seems to be fixed. Therefore we can expect the child and parent to eventually lie in the same basin as t_s becomes larger, and I wonder what is the new conclusion we can draw from this experiment. 2. In Figure 6, at the last points of the curves, is \tilde{t} equal to T? If so the green curves should meet the red curves since there is no additional training, while if not I wonder the details. Regarding the claim that the learned NTK can outperform full deep learning, it seems that the kernel can indeed drives the training error smaller, but the improvement in test error is very small and may be negative as in Figure 12. I also wonder if the curves in Figures 6, 11 and 12 represent the average of multiple parent and child runs?

Correctness: Discussed above.

Clarity: I think the paper is well-written, but more details would be helpful, as discussed above.

Relation to Prior Work: The discussion of prior work is thorough as far as I know.

Reproducibility: Yes

Additional Feedback: Reply to the feedback: Thanks for the response! I agree it is interesting that the weights further travel a distance of 43 after epoch 30 while the learned NTK at epoch 30 is already able to give a good test accuracy. I think such details should be mentioned in the paper. On the other hand, I still think the current empirical results are not enough to conclude that the learned NTK can outperform full deep learning, particularly when considering the test accuracies. More discussion on this would be interesting.


Review 4

Summary and Contributions: The paper summarizes and compares three areas that aim fora better theoretical understanding of deep neural networks. In detail local, global loss landscapes as well as neural tangent kernels (NTK). Based on that the authors establish several criteria to measure how networks compare in terms of loss landscape and NTK. They are used to see if a child networks take at some point during training ("spawning" a network) diverge after further training. The results suggest that the first (5) epoch(s) are determining the final "loss basin" and how the network will converge. Furthermore, the authors show that a data-dependent NTK can trained to good accuracy (compared to a randomly initialized one). ---- read the reviews and the rebuttal and want to keep my score ----

Strengths: * Interesting results that likely contribute to the discussion in this field. * The empirical evaluation seems to be extensive.

Weaknesses: (I'm afraid I feel not familiar enough with the literature to give good critique here.)

Correctness: I am not familiar with the literature, but the paper makes a good impression. Experiments are outlined and detailed well and I would assume one could

Clarity: Yes, fairly easy to read.

Relation to Prior Work: I am not familiar with the literature, but the handling gives a good impression.

Reproducibility: Yes

Additional Feedback: As mentioned I am not too familiar with the literature in this area, but to not waste the review: My educated (phd in ML) opinion is that the submission is solid and by reading it twice it seems sound to me.