NIPS 2002


December 13 and 14, 2002
Delta Whistler Resort, British Columbia, CA

In mathematics and theoretical computer science, exhibiting counter examples is part of the established scientific method to rule out wrong hypotheses. Yet, negative results and counter examples are seldom reported in experimental papers, although they can be very valuable. Our workshop will be a forum to freely discuss negative results and introduce the community to challenging open problems. This may include reporting:

  • experimental results of principled algorithms that obtain poor performance compared to seemingly dumb heuristics;
  • experimental results that falsify an existing theory;
  • counter examples to a generally admitted conjecture;
  • failure to find a solution to a given problem after various attempts;
  • failure to demonstrate the advantage of a given method after various attempts.

Prospective participants are invited to submit one or two pages of summary. Theory, algorithm, and application contributions are welcome. We also welcome tutorials or historical presentations on negative results and counter examples that pushed the frontiers of neural network and machine learning research as well as tutorials on scientific methodology making use of negative results and counter examples.

In preparing your submission, please remember that reporting negative results and counter examples does not mean reporting inconclusive results. One may report experiments that failed because of an invalid design or an invalid theory, if a tentative analysis of the reasons of failure is provided and the subject matter is potentially of interest to others. But, the failure of an experiment will be consired a potentially interesting negative result only if some conclusions can be drawn.

If you are introducing the community to a new open problem, it is desirable that you provide both (i) a high level introduction stating the context of the problem and its fundamental and/or practical importance, and (ii) a formal mathematical statement of the problem, if applicable.

Email submissions to:

                                         Saturday morning session

7:30 am   Welcome and introduction, Isabelle Guyon.
7:45 am   On the impossibility of learning a continuous distribution in a covariant way, Timothy Holy and Ilya Nemenman.
We discuss a question of whether it is possible to infer a continuous probability density and other quantities in a reparameterization covariant way. An explicitly constructed reparameterization example gives a negative answer to this question. The conclusion does not dependent on a particular learning scenario used. We present arguments that explain and further strengthen the result.  Finally we argue that approximate reparameterization invariance with respect to a class of "weak" reparameterizations is possible, and the quality of the approximation depends on the number of samples, as well as on the assumptions about probability densities involved.
8:15 am There is no unbiased estimator of the variance of K-fold cross-validation, Yoshua Bengio.
K-fold cross validation is probably the most commonly used method to estimate generalization error (or to perform model selection), especially when there is little training data. In order to compare learning algorithms, it is important to estimate the uncertainty around the estimation obtained by cross-validation. Such uncertainty estimates (either the variance, a confidence interval, or a p-value against the null hypothesis of no difference) are made more and more mandatory by reviewers of machine learning papers involving experimental validation of new learning algorithms. Unfortunately, we can prove the following negative result (and we will explain its basis): there exists no universally (for all distributions) unbiased estimator of the variance of the K-fold cross-validation generalization error estimator, using only the outcome of the K-fold cross-validation experiment (i.e. the individual errors). However, understanding the source of this problem can hopefully help us choose among a variety of candidates, or resort to estimators based on multiple K-fold cross-validations.
8:45 am   On the number of modes of a Gaussian mixture, Miguel A. Carreira-Perpinan and Chris Williams.
We consider the following question: given a Gaussian mixture in D dimensions with M components, what is the maximum number of modes that it can have? As far as we know, the answer to this is only known for particular types of mixture and/or particular values of D and M. The question remains open in general. We conjecture that if all the covariances of the Gaussian mixture are isotropic or equal to each other then it can have M modes at most. This intuitive conjecture does not hold when the covariances are non-isotropic. We will review some related results in statistics and scale space theory and also discuss algorithms that attempt to find all modes of a Gaussian mixture. Aside from its theoretical relevance, the problem is practically important for statistical machine learning, in models such as kernel density estimation or the generative topographic mapping, and in algorithms such as a recent method for sequential data reconstruction or mean-shift algorithms for clustering.
9:15 am   Break.
9:30 am   Discussion: Negative results.
Informally share negative results. Get feed-back. Know what you do not need to try.

                                         Saturday afternoon session

4:00 pm   Panel: Suggested Directions of Research.Y. Bengio, N. Tishby and other panelists TBA.
5:00 pm   Discussion: Open problems
Informally share open problems. Have they been solved them already? Get new ideas.
6:00 pm   Break.
6:10 pm   Impromptu talks.
Please contact I. Guyon in the morning if you want to present in that Section.
6:50 pm   Closing remarks.
Open problems and suggested directions of research will be summarized.


Journal of Interesting negative results in Natural Language Processing and Machine Learning. JINR is an electronic journal that gives a voice to negative results which stem from intuitive and justifiable ideas, proven wrong through thorough and well-conducted experiments. It also encourages the submission of short papers/communications presenting counter-examples to usually accepted conjectures or to published papers.

Forum for Negative Results in the Journal of Universal Computer Science
Current Computer Science research is primarily focused on solving engineering problems. Often though, promising attempts for solving a particular problem fail for non avoidable reasons. Due to the current CS publication climate such negative results today are usually camouflaged as positive results by non evaluating or mis-evaluating the research or by redefining the problem to fit the solution.

Science Makes Much Ado About Nothing
After decades of shelving studies with negative results, researchers around the nation are agog about not one, but two new journals that focus only on studies that demonstrate what doesn't work. 

Journal of Negative Observations in Genetic Oncology
In the pursuit of genes whose mutations drive the development of human cancers, most of the candidate genes will elicit negative results -- i.e., no mutations will be found. The dissemination of negative data is therefore a crucial component of a lean strategic plan for the genetic analysis of cancer. 

Journal of Negative Results in Biomedicine
This open access, online journal publishes papers on all aspects of unexpected, controversial, provocative and/or negative results/conclusions in the context of current tenets, providing scientists and physicians with responsible and balanced information to support informed experimental and clinical decisions.

Journal of Articles in Support of the Null Hypothesis
In the past other journals and reviewers have exhibited a bias against articles that did not reject the null hypothesis. We plan to change that by offering an outlet for experiments that do not reach the traditional significance levels (p<.05). Thus, reducing the file drawer problem, and reducing the bias in psychological literature. Without such a resource researchers could be wasting their time examining empirical questions that have already been examined.

Index of Null Effects and Replication Failures
The iNERF is an index comprised of short 1-2 page descriptions of experiments or replications that did not meet the traditional level of significance. The purpose of the iNERF is to provide researchers with an opportunity to disseminate information about their null studies without having to take up precious time writing up a full manuscript.

Teaching Problem Solving, Hypothesis Testing, Evolution, and the Meaning of Life Through the Marine Insects Question
The question of why insects are not as dominant at sea as they are on land is ideal for teaching how to form and evaluate scientific questions. Once a hypothesis is formed, we look for current examples that would contradict it. For example, the argument that insects can’t survive in the ocean because of water pressure doesn’t seem so good when you realize one insect species survives at a depth of 1,300 meters! Eliminating hypotheses by counter examples is a powerful approach in assessing hypotheses.

The Seven Open Problems of the Clay Mathematics Institute
In order to celebrate mathematics in the new millennium, The Clay Mathematics Institute of Cambridge, Massachusetts (CMI) has named seven “Millennium Prize Problems” with $1 million allocated to each. One hundred years earlier, on August 8, 1900, David Hilbert delivered his famous lecture about open mathematical problems at the second International Congress of Mathematicians in Paris. This influenced CMI's decision to announce the millennium problems. The clear and precise way the problems are exposed is inspiring.

Machine Learning Research: Four Current Directions
Tom Dietterich - 1997.
Machine Learning research has been making great progress in many directions. This article summarizes four of these directions and discusses some current open problems. The four directions are (a) improving classification accuracy by learning ensembles of classifiers, (b) methods for scaling up supervised learning algorithms, (c) reinforcement learning, and (d) learning complex stochastic models. 1 Introduction The last five years have seen an explosion in machine learning research. 

Contact information 
Isabelle Guyon 
Clopinet Enterprises
955, Creston Road,
Berkeley, CA 94708, U.S.A.
Tel/Fax: (510) 524 6211