• Andre Beinrucker got 1st prize for oral presentation for the presentation of his research and of the MASH project at the interdisciplinary PhD symposium of the Potsdam Graduate School.
• Charles Dubout got the Idiap Paper Award 2011 for his NIPS paper Boosting with Maximum Adaptive Sampling.

After a first testing round, the contests are now ready to accept your submissions.

Congratulations to wenqi, the winner of our those first contests, with the heuristic hue.

A number of new features were added to the MASH platform during the past few months, and internally beta-tested. You'll find a summary of them below.

• Detailed classification experiments results

  Additionally to the results of an image classification experiment, the following informations are now accessible:

  • Number of features used from each heuristic
  • Images of the worst mistakes (both false positives and negatives)
  • Per-label analysis of the correct classifications
  • Per-label analysis of the incorrect classifications
• Classification-based contests

  Any public heuristic can now enter our contests. There are currently three contests, of increasing difficulties.

• Clustering of the heuristics

  The similarities between the heuristics are now analysed, to produce a graph of the space of heuristics. You can now see which heuristics are close to yours, and thus who your friends (or foes) are.

The future developments will focus on goal-planning.

• An evaluation mechanism for newly-uploaded heuristics. This will give contributors a rapid feedback on the usefulness of their work in various scenarios.
• Based on this evaluation, a ranking of the heuristics.
• Display detailed informations when a heuristic crash (or is considered problematic by the platform). This will help you to debug your heuristics. You will also be notified by e-mail.
• Permanently run public experiments using the public heuristics.

Note that all the heuristics are currently checked and evaluated. Once done, the heuristics that were previously public will be published again.

The next update of the platform will focus on displaying more detailed experiment results.

To introduce those new features, we had to delete all the existing experiments. It is very likely that we'll have to do that again during the next update. After that, this should not happen again.

Please report any problem you might encounter during the following days (for instance on the forum). Thanks.

We released two screencasts to introduce the MASH platform.

The first one explains the usage of the website:

The second one talks about the development of a new heuristic:

• An evaluation mechanism for newly-uploaded heuristics. This will give contributors an immediate feedback on the usefulness of their work in various scenarios. A first version of this system is already available.
• Based on this evaluation, a ranking of the heuristics.
• Display detailed informations when a heuristic crash (or is considered problematic by the platform). This will help you to debug your heuristics.
• Display more detailed experiment results (for instance: a ROC curve, the worst mistakes, ...).
• Add more experiment types (object detection and goal-planning).
• Permanently run public experiments using the public heuristics.