This Kaggle competition looks great! I suggest the following as a starting approach for competing in it:
Train two models, one for Ictal (seizure) and one for Interictal (normal). With learning on, feed training data to the respective model. Then, with learning turned off, feed a test sample to both models, and record the anomaly scores for each model on the test sample. Classify the sample based on the lower anomaly score between the two. I'm willing to help out with this, if anyone is interested in participating. On May 19, 2014 at 4:24:05 PM, Doug King ([email protected]) wrote: I think the criticism is valid if we don't have repeatable proof of results that can be evaluated against other machine learning methods. We don't seem to have a good framework for that. Unfortunately all the benchmarks out there all are based around static object recognition. Does anyone know of any time based / temporal datasets being used as benchmarks? I haven't checked in with Kaggle for a while. To my surprise, today they opened a competition that is a good fit for NuPic and something we have discussed here before as a worthwhile and interesting demo. It is the UPenn and Mayo Clinic's Seizure Detection Challenge. It looks like it is open to anyone, not just high rankers, and the purse is $8000 USD. Competition closes Aug. 19 and so far about 8 teams have signed up. https://www.kaggle.com/c/seizure-detection -Doug On Mon, May 19, 2014 at 1:50 PM, Pedro Tabacof <[email protected]> wrote: While DeepMind's results are impressive, as far as I know their deep neural network was static, and the games they did really well were somewhat "static" games, with little long-term dependence between the game states. "Dynamic" games, such as Pitfall, perform rather poorly, so this could be an interesting problem to apply reinforcement learning (or motor control) with Nupic. On Mon, May 19, 2014 at 5:43 PM, Craig Quiter <[email protected]> wrote: I think it would be interesting to see how NuPIC scores in the Arcade Learning Environment. http://www.arcadelearningenvironment.org/ DeepMind did very well here with a combination of convolutional neural nets for spatial pooling and Q-learning with some clever tricks to do temporal pooling and driving of actions. http://www.cs.toronto.edu/~vmnih/docs/dqn.pdf They have a neuron per action at the top of their net, much like classification in NuPIC works. On Mon, May 19, 2014 at 11:24 AM, Chetan Surpur <[email protected]> wrote: I agree that we need to show the CLA's performance on standard benchmarks. However, I don't think those benchmarks are MNIST or ImageNet (as Yann LeCun suggests), at least not at this point in time. Instead, we need to find good time-dependent datasets that exercise the targeted capabilities of the CLA. On May 19, 2014 at 4:38:03 AM, Nicholas Mitri ([email protected]) wrote: I would classify those as more honest (albeit blunt) than rude. Realistically, HTM/CLA hasn’t earned its stripes yet but that’s what we’re trying to do as a community. Unfortunately, there’s a lot of ‘going with the flow’ in the efforts I’ve been aware of. A more critical approach is necessary on our part as a community. There are very well known benchmarks any algorithm like CLA needs to perform well on and, at this point, it should be an utmost priority for us to put CLA through its paces and start to validate its claims rather than accept them at face value. I think it would be great if a part of the community took it upon itself to do just that (I’d be the first to volunteer!). Focus on breaking down CLA into its individual components and validating the theories on those levels. Not only does that grant it merit should its claims hold true but it’ll also reveal a lot of ways in which it can be improved or optimized. best, Nick On May 19, 2014, at 12:25 PM, Fergal Byrne <[email protected]> wrote: Yann LeCun (Director of AI Research at Facebook) gave a fascinating AMA on Reddit [1] a few days ago. As usual, a few people asked him about Jeff, HTM, Numenta and NuPIC. As usual, his answers varied from the somewhat complimentary [2] to the downright rude [3]. The whole AMA is well worth reading, but I was wondering what people here think about the scorn many leading lights in Machine Learning and AI heap on HTM/CLA. [1] http://www.reddit.com/r/MachineLearning/comments/25lnbt/ama_yann_lecun/ [2] http://www.reddit.com/r/MachineLearning/comments/25lnbt/ama_yann_lecun/chisjsc [3] http://www.reddit.com/r/MachineLearning/comments/25lnbt/ama_yann_lecun/chj9pwm -- Fergal Byrne, Brenter IT Author, Real Machine Intelligence with Clortex and NuPIC https://leanpub.com/realsmartmachines Speaking on Clortex and HTM/CLA at euroClojure Krakow, June 2014: http://euroclojure.com/2014/ and at LambdaJam Chicago, July 2014: http://www.lambdajam.com http://inbits.com - Better Living through Thoughtful Technology http://ie.linkedin.com/in/fergbyrne/ - https://github.com/fergalbyrne e:[email protected] t:+353 83 4214179 Join the quest for Machine Intelligence at http://numenta.org Formerly of Adnet [email protected] http://www.adnet.ie _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org -- Pedro Tabacof _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org _______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
