{"forum": "B1g5sA4twr", "submission_url": "https://openreview.net/forum?id=B1g5sA4twr", "submission_content": {"authorids": ["preetum@cs.harvard.edu", "galkaplun@g.harvard.edu", "ybansal@g.harvard.edu", "tristanyang@college.harvard.edu", "b@boazbarak.org", "ilyasu@openai.com"], "title": "Deep Double Descent: Where Bigger Models and More Data Hurt", "authors": ["Preetum Nakkiran", "Gal Kaplun", "Yamini Bansal", "Tristan Yang", "Boaz Barak", "Ilya Sutskever"], "pdf": "/pdf/2313f8e4c1bbcb174b1e34904fbc5f638c589efa.pdf", "TL;DR": "We demonstrate, and characterize, realistic settings where bigger models are worse, and more data hurts.", "abstract": "We show that a variety of modern deep learning tasks exhibit a \"double-descent\" phenomenon where, as we increase model size, performance first gets worse and then gets  better.  Moreover, we show that double descent occurs not just as a function of model size, but also as a function of the number of training epochs. We unify the above phenomena by defining a new complexity measure we call the effective model complexity, and conjecture a generalized double descent with respect to this measure. Furthermore, our notion of model complexity allows us to identify certain regimes where increasing (even quadrupling) the number of train samples actually hurts test performance.", "keywords": ["deep learning", "double descent", "optimization", "SGD", "complexity"], "paperhash": "nakkiran|deep_double_descent_where_bigger_models_and_more_data_hurt", "_bibtex": "@inproceedings{\nNakkiran2020Deep,\ntitle={Deep Double Descent: Where Bigger Models and More Data Hurt},\nauthor={Preetum Nakkiran and Gal Kaplun and Yamini Bansal and Tristan Yang and Boaz Barak and Ilya Sutskever},\nbooktitle={International Conference on Learning Representations},\nyear={2020},\nurl={https://openreview.net/forum?id=B1g5sA4twr}\n}", "full_presentation_video": "", "original_pdf": "/attachment/51a9c89ab14e0b8ab44670eaabb35678344db624.pdf", "appendix": "", "poster": "", "spotlight_video": "", "slides": ""}, "submission_cdate": 1569439393835, "submission_tcdate": 1569439393835, "submission_tmdate": 1583912040571, "submission_ddate": null, "review_id": ["H1lrZBqMoH", "H1xTYre6tB", "rJxvG0y6KH"], "review_url": ["https://openreview.net/forum?id=B1g5sA4twr&noteId=H1lrZBqMoH", "https://openreview.net/forum?id=B1g5sA4twr&noteId=H1xTYre6tB", "https://openreview.net/forum?id=B1g5sA4twr&noteId=rJxvG0y6KH"], "review_cdate": [1573197053258, 1571779972894, 1571778063458], "review_tcdate": [1573197053258, 1571779972894, 1571778063458], "review_tmdate": [1573197090186, 1572972482986, 1572972482941], "review_readers": [["everyone"], ["everyone"], ["everyone"]], "review_writers": [["ICLR.cc/2020/Conference/Paper1329/AnonReviewer1"], ["ICLR.cc/2020/Conference/Paper1329/AnonReviewer2"], ["ICLR.cc/2020/Conference/Paper1329/AnonReviewer3"]], "review_reply_count": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "review_replyto": ["B1g5sA4twr", "B1g5sA4twr", "B1g5sA4twr"], "review_content": [{"experience_assessment": "I have read many papers in this area.", "rating": "8: Accept", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "title": "Official Blind Review #1", "review": "The paper defines the effective model complexity(EMC) that defines the complexity of the model. EMC depends on several factors such as data distribution and architecture of the classifier. \n \nThe paper empirically shows that the double descent phenomenon occurs as a function of EMC. \n\nThe paper provides interesting perspectives of their experiments and gives the intuitions for these observations. However, these are mainly hypotheses.\n\nI like the paper for its rigorous experiments and providing intutions from their observations.\n\nHowever, I am very new to this area of research and will only provide my review on my understanding.", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A"}, {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #2", "review_assessment:_checking_correctness_of_derivations_and_theory": "I assessed the sensibility of the derivations and theory.", "review": "This paper provides a valuable and detailed empirical study of the double descent behaviour in neural networks. It investigates presence of this behaviour in a range of neural network architectures and apart of identifying it as a function of the model size it also identifies it as a function of training time which I believe is novel. Overall I think the paper presents results valuable to the community. At the same time it has several issues that need to be addressed. If the issues are addressed in a satisfactory manner I will recommend acceptance of the paper. \n\nIssues and questions: \n\n** \"Such a phenomenon was first postulated by Belkin et al. (2018) who named it \u201cdouble descent\u201d and demonstrated it on MNIST with decision trees, random features, and 2-layer neural networks with `2 loss.\" This is not a correct statement. The authors cite works (Advani & Saxe 2017) and there is also \"Stefano Spigler, Mario Geiger, Stephane d\u2019Ascoli, Levent Sagun, Giulio Biroli, and Matthieu Wyart. A jamming transition from under-to over-parametrization affects loss landscape and generalization. arXiv preprint arXiv:1810.09665, 2018.\" that both identified this behaviour prior to the Belkin et al. (2018) work. There is even a much older line of work identifying the peak in the generalization error by Opper's group: \nhttp://www.ki.tu-berlin.de/fileadmin/fg135/publikationen/opper/Op03b.pdf (1995)\nSiegfried B\u00a8os and Manfred Opper. Dynamics of training. In Advances in Neural Information\nProcessing Systems, pages 141\u2013147, 1997.\nhttp://www.ki.tu-berlin.de/fileadmin/fg135/publikationen/opper/Op01.pdf\n\nMoreover, as is only said in the supplementary material, Geiger et al. also tested on the CIFAR dataset, while the introduction of the present article only mentions previous experiments on MNIST. The fact that previous work also observed this in CIFAR should be moved in the introduction. \n\n** The authors claim to define the \"effective model complexity (EMC)\" and claim this as one of the main results of the paper.\n\nPresenting this as a new measure is strange, this exactly is what Geiger et all call jamming transition/threshold in Stefano Spigler, Mario Geiger, St\u00b4ephane d\u2019Ascoli, Levent Sagun, Giulio Biroli, and Matthieu Wyart. A jamming transition from under-to over-parametrization affects loss landscape and generalization. arXiv preprint arXiv:1810.09665, 2018. And very closely related to what Belkin calls \"interpolation threshold\" (to define interpolation threshold we could fix the model and vary the number of samples, Belkin et all fix number of samples and vary size of the model, but one is just the dual of the other). Giving it a yet completely new name seems to create more noise than value.\n\nThe relation between EMC and the jamming transition is discussed in the supplement. However, not in a accurate way. The authors say \"a \u201djamming transition\u201d when the number of parameters exceeds a threshold that allows training to near-zero train loss.\" but jamming is inspired by the physical phenomena when spheres are added into a finite volume box and the more spheres we have the harder it gets to fit them all in until it is not possible anymore. The analogy here is that fitting sphere in = reaching training error zero. Then number of spheres = number of samples. Thus the more samples the harder it is to get the training error to zero, leading to the jamming transition. Both in training and jamming the number of samples at which this happened depends on the details of the protocol/training, thus it does depend on things such as regularization.\n\nThe only aspect that I have not seen covered in the jamming analogy is the epoch-wise double descent.\n\nIn any case the discussion in the paper needs to be adjusted and these relevant relations to previous work corrected and moved from the supplement to the introduction.  \n\n** \"Informally, our intuition is that for model-sizes at the interpolation threshold, there is effectively only one model that fits the train data and this interpolating model is very sensitive to noise in the train set and/or model mis-specification.\" This intuition is correct I believe. However, it should not be called \"our intuition\" as it already appeared in the line of work by Opper cited above. \n\n** The authors present as another main result the fact that under comparable training conditions training with more data samples provides worse generalization, examples of this is also included already in the papers by Opper et al. cited above. "}, {"experience_assessment": "I have read many papers in this area.", "rating": "6: Weak Accept", "review_assessment:_thoroughness_in_paper_reading": "I read the paper at least twice and used my best judgement in assessing the paper.", "review_assessment:_checking_correctness_of_experiments": "I assessed the sensibility of the experiments.", "title": "Official Blind Review #3", "review_assessment:_checking_correctness_of_derivations_and_theory": "N/A", "review": "I do not have much to say about the paper except that I like the sumilation, and found rather interesting. It shows a rather extensive set of simulations, that enrich the observations of the so-called double descent phenomena, and shows empirically its apparent generality. Also, the Effective Model Complexity seems to be a good description of what is empirically observe. All in all, i definitely support publication.\n\nI have, however, two comments: first, a minor one: there is a part of the story on double descent that is completely forgotten: the fact that there is a pic in the data for some classificators at #example close to #parameters was known at least in the 90! It is discussed in \"Statistical Mechanics of Learning\" A. Engel & C. Van den Broeck, 2001, page 61, with a plot that  comes rather from an old work of Pr. Manfred Opper (1995), can be seen here: http://www.ki.tu-berlin.de/fileadmin/fg135/publikationen/opper/Op01.pdf FIG 10 \nFrom this work, it is also rather clear that increasing the number of training examples can hurt performances (this is exactly what the plot says), at least without regularisation. Since the authors are quite thorough in their review of history -especially in the appendix- I thought I would point this reference.\n\nThe second comment is that the paper should acknowledge the work of the  Geiger et al more explicitly, and cite as well the related work of  Spiegler et al (https://arxiv.org/abs/1810.09665) rather explicitly! Virtually all they show was already shown in this work (which by the way PREDATES Belkin et al), albeit only for a simpler set of data.  I agree that the authors extend a lot on this papers, especially in terms of dataset and completeness of experiments, but they are definitely closely related, and the fact that it is not cited is a serious flaw to the current version.\n\nAlso, in the conclusion  \"We note that many of the phenomena that we highlight often do not occur with optimal early- stopping. However, this is consistent with our generalized double descent hypothesis:\u2026\". Again this was shown (empirically) as well in the above mentioned papers and also explained in \"physics terms\" where the peak is like a phase transition (jamming), if one stop the dynamics.\n\n"}], "comment_id": ["y7kBfIonCR", "HkxGyDxhoS", "BJeZYOWbjB", "H1gFwdZbjB", "BJebj9ZrqB"], "comment_cdate": [1576864693830, 1573811929797, 1573095545472, 1573095521045, 1572309657014], "comment_tcdate": [1576864693830, 1573811929797, 1573095545472, 1573095521045, 1572309657014], "comment_tmdate": [1576952792426, 1573811929797, 1573095545472, 1573095521045, 1572309657014], "comment_readers": [["everyone"], ["everyone"], ["everyone"], ["everyone"], ["everyone"]], "comment_writers": [["ICLR.cc/2020/Conference/Paper1329/Authors", "ICLR.cc/2020/Conference"], ["ICLR.cc/2020/Conference/Paper1329/AnonReviewer2", "ICLR.cc/2020/Conference"], ["ICLR.cc/2020/Conference/Paper1329/Authors", "ICLR.cc/2020/Conference"], ["ICLR.cc/2020/Conference/Paper1329/Authors", "ICLR.cc/2020/Conference"], ["ICLR.cc/2020/Conference/Paper1329/Authors", "ICLR.cc/2020/Conference"]], "comment_reply_content": [{"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}, {"replyCount": 0}], "comment_content": [{"title": "Remarks to Program Chairs regarding related works", "comment": "We thank the reviewers and the committee and are excited to present the paper in ICLR.\nThat said, we believe the PC is mistaken about \"original lack of proper references\", which we wish to clarify.\n\nWe believe the comments on citations do not refer to the paper itself but rather to other related postings and tweets. The paper should be judged by its content alone.\nNevertheless, our blog post [ https://windowsontheory.org/2019/12/05/deep-double-descent/ ] contains several citations and the main citations are included even in the abbreviated version posted on the OpenAI blog.\n\nThe actual submission had always contained an extensive discussion of prior work, describing that our contribution is showing this phenomenon exists in modern deep networks and presenting new findings (epoch double descent, and sample non-monotonicity) as well. The reviewers pointed out two pre-Belkin works of Opper [1995, 2001] that we missed and are now added."}, {"title": "Acknowledgement of the answer", "comment": "I thank the authors for considering my comments and taking them into account in the revision. \nI am still not convinced about the relevance of the defined Effective Model Complexity. But the experimental part of the paper is very nice and the paper should be published in ICLR.  "}, {"title": "Response to Reviewer 3", "comment": "Thank you for your comments and references.\nAs we noted in the response Reviewer 2, we indeed missed several relevant citations to works prior to Belkin et al., as you mention, and we are grateful to you for pointing them out to us\u2014we have updated the Introduction of the paper to reflect the connections to prior work more explicitly.\n\nAs we note in the response to Reviewer 2, we believe our contribution lies mainly in the experimental verification of this phenomenon to modern deep networks, its extension to epoch-wise double descent, and a unified view on both these measures.\n\nRegarding early-stopping: We refer to Advani-Saxe and Geiger et al. in the context of early-stopping in the Extended Discussion of Related Works in the Appendix. Interestingly, as we note in the paper, we did observe at least one setting where double-descent still occurs even with optimal early-stopping (Figure 15). "}, {"title": "Response to Reviewer 2", "comment": "Thank you for your comments and references.\nWe indeed missed several relevant citations to works prior to Belkin et al., as you mention.\nThank you so much for pointing them out to us\u2014we have updated the Introduction of the paper to reflect the connections to prior work more explicitly.\n\nOur main contribution is demonstrating that these phenomena, which have been noticed in various settings in prior work (including Belkin et al., Spigler et al., Opper et al.), continue to hold for modern neural networks and domains. It is not obvious that behaviors that occur for linear regression or 2-layer networks will carry over to modern 18-layer convolutional networks, and so we believe that these experimental results are valuable. \n\nMoreover, we discovered a new instance of this phenomenon (\"epoch-wise double descent\") which as far as we know had not been observed before\u2014and we unify all these forms of double-descent under a generalized hypothesis (EMC).\n\nWhile we agree that it is related to concepts such as the \u201cjamming transition\u201d and others used in prior work (as well as to Radamacher complexity), our notion of \"Effective Model Complexity (EMC)\" is not identical to those prior notions. Indeed, as far as we are aware, it is the only explicitly defined complexity measure that both identifies the location of the test-error peak and also applies for all known instances of double-descent, including epoch-wise double descent. \n\nRegarding the novelty of \"sample-wise non-monotonicity\": Again, we acknowledge that this behavior may have been present in other settings in prior works (implicitly or explicitly), but our main contribution is explicitly stating this behavior and demonstrating it for modern neural networks, as well as using the double descent analysis to identify the regimes where it could hold.\n"}, {"title": "On the Mechanisms of Sample-wise Non-monotonicity", "comment": "Thanks for your comment! This is a good question, and the issue is subtle.\n\nFor test error to increase with sample size, at least one of the following must occur:\n(A): Training error increases with sample size\n(B): Generalization gap (= Test Error - Train Error) increases with sample size.\n\nNow, all possible settings can occur in practice:\nThere exist cases where (A) is true and (B) is false,\nwhere (A) is false, and (B) is true,\nand where both (A) and (B) are true.\n\nFor cases where (A) is true: As you mention, in Figure 26 (intermediate model) the train error increases and the generalization gap does not increase. \n\nWhere (A) and (B) are both true: This occurs in Figure 9b, for the Smaller Model in the range of samples 0-10k. Here, both the train loss and the generalization gap increase.\nHere's the plot including train loss for Figure 9b: https://i.imgur.com/ITJHMeG.png\n\n\nWhere (A) is false and (B) is true: This is most evident in the case of Random Fourier Features (Figure 11). Consider a RFF model with embedding-dimension d = 1000. This is simply linear regression in the embedded space, and this model can exactly interpolate (train error is exactly 0) up to n <= d samples.\nWhen n <= d, the linear model is the min-norm interpolating solution of the (embedded) train set. (This is the solution that would be found by gradient descent/flow).\n\nThis plot shows 1D slices of Figure 11: https://i.imgur.com/NBqJjul.png\nNow, compare n = 500 to n = 1000 samples. Both have 0 train error, but the generalization gap grows (and peaks at n=1000). In our framework, moving from n=500 to n=1000 shifts the model from being \"very overparameterized\" to being \"critically parameterized\".\n\nAt n = 1000, there is exactly one interpolating solution -- and intuitively, since the model is misspecified (ie, the true distribution is not linearly separable under the embedding), forcing this model to interpolate the data does not perform well. Whereas for n < 1000, there are many interpolating solutions, and choosing the min-norm one acts as a good regularizer.\n\n\nNote that sample-wise nonmonotonicity occurs even without random-features: it occurs for standard linear regression, when the data-distribution is (for example) gaussian with observation noise. In this case, the peak in test error occurs for similar reasons -- because the variance of the interpolating estimator grows at n = d.\n(A related analysis is in: https://arxiv.org/abs/1903.07571)\n\nOur intuition is that this setting (where (B) is true) is most indicative of the mechanisms of sample-wise nonmonotonicity: for the critical number of samples, the models must \"try very hard\" to fit the train set, which can destroy their global structure. Whereas for fewer samples, the models are overparameterized enough to fit the train set while still behaving well on the distribution."}], "comment_replyto": ["ir53kmmQyi", "BJebj9ZrqB", "rJxvG0y6KH", "H1xTYre6tB", "B1lSR-pl5r"], "comment_url": ["https://openreview.net/forum?id=B1g5sA4twr&noteId=y7kBfIonCR", "https://openreview.net/forum?id=B1g5sA4twr&noteId=HkxGyDxhoS", "https://openreview.net/forum?id=B1g5sA4twr&noteId=BJeZYOWbjB", "https://openreview.net/forum?id=B1g5sA4twr&noteId=H1gFwdZbjB", "https://openreview.net/forum?id=B1g5sA4twr&noteId=BJebj9ZrqB"], "meta_review_cdate": 1576798720695, "meta_review_tcdate": 1576798720695, "meta_review_tmdate": 1576800915903, "meta_review_ddate ": null, "meta_review_title": "Paper Decision", "meta_review_metareview": "This paper experimentally analyzes the double descent phenomenon for deep models. While, as the reviewers have mentioned, this phenomenon has been observed for some time, some of its specificities still elude us. As a consequence, I am happy to see this paper presented to ICLR.\n\nThat being said, given the original lack of proper references as well as the recent public announcements about this paper giving it visibility, I want to make it absolutely clear that this paper is accepted with the assumption that proper credit will be given to past work and that efforts will be made to draw connections between all these works.", "meta_review_readers": ["everyone"], "meta_review_writers": ["ICLR.cc/2020/Conference/Program_Chairs"], "meta_review_reply_count": {"replyCount": 0}, "meta_review_url": ["https://openreview.net/forum?id=B1g5sA4twr&noteId=ir53kmmQyi"], "decision": "Accept (Poster)"}