With input from Andrew Ng, Peter Dayan, Daphne Koller, Sebastian Thrun, Bruno Olshausen, Yair Weiss, and Bernhard Schölkopf
Edited by Max Welling and Zoubin Ghahramani in 2013.
In this informal essay, we describe some of the criteria that will be used to evaluate NIPS submissions. This document should not be construed as an official NIPS policy statement; but through it, we also hope to give advice for writing a good NIPS paper.
We'll take for granted that your paper will be clearly written, be technically sound and correct, and reference previous work. Thus, we will not further dwell on the issues of clarity and soundness, despite their importance. We will instead focus on how one might shape a paper's content so as to maximize its chance of being published and influencing others.
A few notes:
We ask authors to submit one or more keywords for their papers. We use these keywords to assign area chairs and reviewers. Note that the keywords marked by the authors do not bind the PC in assigning reviewers to the paper. For example, if a paper proposes a new algorithm, but contains no empirical assessment, marking it as a learning theory paper will not necessarily lead to more likely acceptance.
NIPS is an interdisciplinary conference that covers both natural and synthetic neural information processing systems. It is often the case that strong NIPS papers appeal to both parts of the community: for example, by using modern analysis methods developed on the synthetic side to study natural systems, or by investigating algorithmic aspects of methods used by natural systems. While a broad appeal tends to strengthen a NIPS submission, there are also many strong NIPS papers that are more specialized, and thus only fall into one of the two categories described below.
A significant fraction of NIPS papers either describe or study new machine learning methods or theory. Examples of such papers may include: a paper proposing a new learning algorithm; one that describes a solution to a difficult application; or one that proves bounds on the error of some learning method. Such papers are expected to make a significant (i) algorithmic, (ii) application, or (iii) theoretical contribution. NIPS seeks to publish papers that will have a high impact in the world---|both within our research community, and beyond. Whenever appropriate, papers will therefore be evaluated on the basis of the following five criteria:
1 Novelty of algorithm. For example, a paper that gives an elegant new derivation for an algorithm; or one that proposes a new approach to an existing problem.
2 Novelty of application/problem. For example, a paper that addresses an important application that has heretofore been little-studied at NIPS. Or, one that introduces a novel machine learning problem (some past examples include ICA and structured prediction) and proposes an algorithm for it.
3 Difficulty of application. For example, an application of machine learning to a difficult, important, and "real" application, that takes into account the full complexity of getting a non-trivial system to work.
4 Quality of results. Whether the algorithm is rigorously demonstrated to give good empirical performance on the task considered (here, "real" data or "real" experiments may be more effective than "artificial" or "toy" experiments); or whether the theoretical results are strong and interesting; etc.
5 Insight conveyed. Whether the paper conveys insight into the nature of an algorithm; into the nature of a practical application or problem; into general lessons learned; and/or into theoretical or mathematical tools that might be used by others for future work.
Not all papers are expected to address all of these criteria, and a paper that is extremely strong on only one of them may well be acceptable for publication. For example, a learning theory paper that studies an existing algorithm may be reasonably expected to address only the last of these criteria. However, in some cases where the research can be reasonably expected to address more than one of the criteria above, a paper may have a better chance of acceptance if it does indeed address them. For example, a paper that gives an elegant mathematical derivation of a new algorithm (Criterion #1) may fare better if it is also demonstrated through rigorous empirical evaluation to do well (Criterion #4), or demonstrated on a real/non-trivial application (Criterion #3). This is because such experiments can help build a significantly stronger case for the algorithm's actual utility. Similarly, a paper describing an impressive application of machine learning (Criterion #2 or #3) may fare better if beyond reporting success, it further elucidates the structure of the problem or algorithm that made the application work, and thereby conveys insight (Criterion #5).
For empirical studies, a good result can lie along many different axes, all of which compare to the best state-of-the-art algorithm. These axes may include: better accuracy, better ROC performance, faster, less memory, more generally applicable, easier out-of-the-box usage, much simpler to code. If an algorithm does not excel along any of these axes, a reviewer may wonder why it is worth publishing at NIPS.
Although NIPS strongly encourages interdisciplinary work that spans multiple topics, we now also describe some evaluation criteria that are more specialized and may apply only to individual topics.
(e.g., clustering, dimensionality reduction, feature selection, nonparametric Bayesian models, graphical models, kernels, boosting, Monte Carlo methods, neural networks, semi-supervised learning, deep learning). Authors of papers that propose new algorithms for well-established, existing problems are encouraged to provide evidence for the practical applicability of their methods, such as through rigorous empirical evaluation of their methods on real data or on real problems. For example, a paper about a new mathematical trick (or about a beautiful new mathematical derivation) would be stronger if it is supported by empirical evidence that the resulting algorithm really helps on a problem. We also encourage submission of papers that describe algorithmic or implementation principles that may have a large impact on applications or on practitioners of machine learning.
Authors of papers that propose new algorithms for existing problems (such as solving MDPs) are encouraged to provide rigorous empirical evaluation of their methods on real problems, and show its relevance to real/difficult decision making or control tasks. For example, rather than demonstrating your idea only on a grid-world or on mountain-car, also show if it works on a more challenging task. The other comments for algorithmic papers also apply here. Learning Theory
Any Learning Theory paper should have a theorem about learning and a proof. Leaving out the proof is not an option in a double blind setting! Several styles of papers exist:
1 Propose a new natural model of learning and algorithm for this model (examples: Bayes learning, statistical learning, PAC learning, Online learning, MDP learning, Boosting).
2 Propose an algorithm with an improved analysis in some standard setting.
3 Prove that some learning task people have been attempting is hard or impossible.
4 "Other". Meta-theorems about learning theorems, etc. Technically difficulty or novelty is not the goal. Impact on the process and practice of learning is the goal. Experimental results are nice but not necessary in general.
Application papers should describe your work on a "real" as opposed to "hypothetical" application; specifically, it should describe work that has direct relevance to, and addresses the full complexity of, solving a non-trivial problem. Authors are also encouraged to convey insight about the problem, algorithms, and/or application. For example, one might describe the more general lessons learned, or elucidate (through an ablative analysis/lesion analysis, which removes one component of an algorithm at a time) which were the key components of the system needed to get the application to work. A NIPS application paper should be comparable in quality to paper in the corresponding application domain conference: for example, a text paper should be acceptable to SIGIR, EMNLP, or other appropriate conference Application papers should not only present concrete application results, but also contain at least one of the below elements:
Authors of vision papers are encouraged to provide rigorous empirical evaluation of their methods to demonstrate value added not just for a few selected images, but more broadly. Ideally, a NIPS paper proposes a machine learning algorithm or system that can be used by a computer vision researcher to help solve a difficult computer vision problem. NIPS papers in this area should be comparable in quality to those accepted in the major computer vision conferences, such as ICCV or CVPR.
Similar to computer vision, a NIPS paper should solve a difficult audio, speech, or other signal processing problem via machine learning; and be useful for a signal processing practitioner. The quality bar for NIPS is higher than those of a typical signal processing conference (such as ICASSP or ICIP): the NIPS papers are 30% longer, the reviews are more detailed, and the acceptance rate is about half. Therefore, a NIPS signal processing paper should be more significant than the average ICASSP paper.
In addition to describing a successful implementation, a NIPS hardware paper should also convey insight into the underlying principles behind your implementation that serve as useful lessons learned to non-hardware researchers, such as computer scientists or neurobiologists.
A significant fraction of NIPS papers, comprising mainly ones from the neuroscience, biological vision, or cognitive science, either describe or study natural systems. Examples include a paper proposing a new model of human decision making, a paper describing evidence for a neural code, and so on. Papers submitted in this category should make significant contributions to the computational, psychological and/or neural understanding of an important biological and/or behavioral system or function. Such papers will be evaluated on the basis of some or all of the following seven criteria:
1 Novelty of model. For example, a new account of a popular issue such as the representation of uncertainty in neural population codes.
2 Novelty of method. For example, a new analytical analysis of a phenomenon (say phase locking in oscillatory networks) that had previously only been studied using simulations.
3 Novelty of results. For example, a re-analysis of data on input-output functions of auditory cortical neurons, showing a new facet of their tuning to spectral contrast.
4 Novelty of system or function. For example, a model of a neural region (say a hypothalamic nucleus) that has not hitherto been analyzed.
5 Fit to data. For example, whether the suggestion evidently accounts for a wide range of data that have resisted previous approaches.
6 Explanatory power. For example, whether the suggestion links different (Marrian) levels of analysis, maybe showing the control-theoretic or Bayesian soundness of a well-known psychological learning rule.
7 Appropriateness of model. For example, if a proposed model or mechanism is supported by multiple data points or experiments.
A good neuroscience model should make testable predictions - and they should be interesting, too. An interesting prediction is something you may not have thought about otherwise: a prediction that is non-obvious, or does not derive directly from the limitation assumptions made in the model. A neuroscience model should give you a new way of looking at the system, which inspires new experiments. NIPS neuroscience papers should either be neuro-scientifically or computationally well-grounded, ideally both. The paper should make a serious attempt at connecting to state-of-the-art neurobiology, and/or provide a rigorous mathematical treatment or comparison to a state-of-the-art engineering method.
Papers on this topic tend to fall between the natural and artificial systems categories. A good brain imaging paper may lead to neurobiological insight, or it may propose an experimental method for obtaining new kinds of measurements. A good brain computer interface would either be useful as a computer interface, or also lead to neurobiological insight.
These criteria were selected with the goals of encouraging good research, and of maximizing NIPS' long-term impact. Note that this is not as simple as accepting papers with high-expected impact. For example, a paper that makes ambitious but poorly substantiated claims may have high expected impact---|largely on the off-chance that the claims turn out to be correct---|but is still likely to be rejected. Some of these evaluation criteria exactly address this issue of providing evidence for the utility of one's work.