Matt, here is a transcript. Some names may need to be corrected, ??? means I did not recognize the speaker.
There are some good discussions here. Last comment around 54:00 is a great idea and something I was thinking about: We should share PSO tuning setting from the community to find out what are common setting ranges that always come up in any model. It would be efficient to minimize the work the the PSO has to do. Maybe a common directory under source control? -Doug ========================================= Summarized and paraphrased interpretation (not literal) transcript of "NuPIC Office Hours - Oct 23, 2013" video Apologies if I mangled a name or talking point. - Doug King 1:50-3:00 ??? : Questions about goals of community members and Numenta. Are people only focusing on specific practical problems? 3:20 Jeff: Nupic goals, Jeff’s goals, practical vs. research advances 5:04 ??? : 5:40 Jeff: long term goals - push theoretical into practical 7:00 ??? : Community is focused on getting system up and running right now, concepts are hard, we need to reduce startup/adoption friction in the toolset. 8:20 ??? : Impressions of community is not a lot of interest in theory / big ideas, as opposed to interest in getting system running. 9:00 Jeff: don't be afraid to post theoretical questions, we need that kind of discussion. It's appropriate for the list. 10:40 Jeff: "I’m working on sensory motor integration" - next big theoretical problem. Open to discussions / question on the list. 11:30 Chetan: Questions about how regions are wired up in a hierarchy? 12:15 Jeff: complicated, condense time up the hierarchy, pooling time up the hierarchy. We see this in biology. Not being used in the NuPic CLA, but code seems to be there, but disabled. 15:00 Jeff: want to focus on sensory motor integration, this can be done today, hierarchy is not easy and will take time to implement. 15:50 ??? : should we use active columns and decode at each level of hierarchy? 16:00 Jeff / Chetan: discuss biology and model, vision, language 17:40 Jeff / Subati: discuss some earlier experiments with Numenta doing hierarchy 19:10 Jeff: explains specifics of how he thinks hierarchy should work, all active cells wired up to next level, no classifiers between levels. 21:00 Jeff: more specifics of how to do hierarchy, subsampling, simplify first level CLAs, etc. 22:10 technical / audio problems 23:40 Anaka question - Is using NuPic as reference for his own version of CLA, wants to know, how does multiple variable / value input impact number of columns and topology. 26:15 Subati / Jeff: too many values give diminishing returns. 4-5 max input per region seems best. 27:40 ???: use strong signal values / fields 28:30 Jeff: For more than 4-5 inputs, use ensemble approach, multiple models / metrics, multiple CLAs, combine results with traditional ensemble techniques. Always get better results from Ensemble techniques. 29:30 Chetan: would using hierarchy allow you to effectively use more input signals? 30:10 Jeff: explains difficulty of doing hierarchy correctly. 31:17 Chetan: 31:35 Jeff: 32:00 Chetan: clarify question – do you mean using all active cells vs. active columns for feeding next level of hierarchy? 32:40 Jeff: explains - state of CLA is all the active cells, which retains temporal information. 33:26 – 35:00 technical diff / audio 35:00 Chetan, last man standing :-) 35:10 Marek question: how do you interpret / reconstruct hierarchy? 35:40 Jeff: can't reconstruct top level of hierarchy easily. 36:30 ?? & Jeff & Subati debate reconstructing top level of hierarchy. All cells determine state all the way up, no decode /reconstruct at levels, vs reconstructing (with classifier) each level and feed up. Jeff on principles says you can’t reconstruct at each level (based on biology), but Subati concedes you could deconstruct at each level and re-encode. There is a difference of opinion about models / implementation of hierarchy among participants. 41:00 interesting views on why certain approaches were taken at Numenta - 41:50 Jeff: brings discussion back to biology / hard to reconstruct hierarchy. Temporal memory causes problems. Without it, hierarchy would be easy. 43:20 Grandmother cells, oh my! 43:50 Marek gets audio back, applause. 43:50 Marek: does hierarchical CLA have less value then a single large region in the accuracy of prediction? 45:30 Jeff: You can make detailed predictions with hierarchy 46:36 Anaka: CLA white paper def. about pooling / classifier / sliding window 47:50 Jeff: Anaka: What are the implementation differences between classifier and pooler ? Jeff: temporal pooler represents encoding of sequences over varying periods of time and creates stable output. Different from what classifier does. 49:20 Anaka: how is PSO being utilized? 49:50 Subati: explains PSO and parameter tuning 59:40 Jeff: PSO is similar to "evolution" by tuning basic cortex algorithm. 52:40 Jeff: New version of Grok product does not use PSO, but a few base models that have been proven to work well. 53:40 Jeff: Surprising that some settings can be optimized without PSO, always end up around a center weight. 54:00 ???: Someone proposes a good idea: We should share tuning setting from the community to find out what are common setting ranges that always come up in any model. It would be efficient to minimize the work the PSO has to do.
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
