Thank you phil. I will consider your advice and have a try recently. Thank you again. :) Yajingfu------------------ ???????? ------------------ ??????: "Phil iacarella"<[email protected]> ????????: 2015??9??28??(??????) ????0:48 ??????: "FYJ"<[email protected]>; ????: Re: How to calculate "field contribution"
Yajingfu, Regarding question 2: I??m facing a similar dilemma where I plan on using multiple fields concatenated together to predict another field. Using the Network Classifier, I would have fields A+B+C+D -> predict E. My approach is to encode fields A-D to have the same length (n) and the same number of on bits (w). Using the default values in the RDSE (n=400, w=21, forced). If its category data then use the SDRCategoryEncoder with n=400 and w=21. Not real sure how to set the potentialRadius and potentialPct ? This will produce a 1600 bit input vector to the SP. My thinking is for any TP ActiveState: A+B+C+D, A+B+C2+D, A+B+C3+D, A+B+Cn+D would have the same overlap count (or nearly so) with one another. This would be true for holding any 3 fields constant and changing just one. This would allow (I think) separate fields to provide some predictive contribution of ABCD that AB doesn??t provide and not producing wild differing activeStates. This might be completely wrong. I have not tested this yet. I??m curious if anyone has and what others think of this approach? -Phil On Sep 26, 2015, at 10:00 PM, Matthew Taylor <[email protected]> wrote: Sorry, I don't know the answer to your feature detection question. You might try asking the nupic-theory list. As for your installation problem, I think there is a mismatch between the python headers used to compile the binary nupic.core installation and the python version running the program. My best advice is to uninstall everything and compile from scratch. Matt Sent from my MegaPhone On Sep 26, 2015, at 5:01 PM, ?????? <[email protected]> wrote: Hello NuPIC Are there anyone can answer my questions? Thank you very much. Yajingfu ------------------ ??????件 ------------------ ??件??: "??????";<[email protected]>; ????????: 2015??9??24??(??????) ????9:26 ??件??: "??????"<[email protected]>; ????: ?????? How to calculate "field contribution" Thank you Pascal. From the website, I think the formula is?? C=(E1-E2)/E1 which C represents the contribution of field B for field A, E1 presents the error by A itself, E2 represents the error of A with the help of B. Is it right? If my understanding is right??and assuming there are 5 fields (A,B,C,D,E), and we want to predict A. Does it just calculate the combination of A, AB, AC, AD, AE individually? Moveover, I know for a sequence A1, A2, A3 ... An, it will change the performance (increase or decrease) between synapses and remember a time sequence. How to use sequence B1, B2, B3 ... Bn to help decreasing sequence A' error? I want to know more about the key algorithm. Which code do I need to learn? Are there any suggestion for me? Thank you so much. :) Yajingfu ------------------ ??????件 ------------------ ??件??: "Pascal Weinberger";<[email protected]>; ????????: 2015??9??24??(??????) ????6:03 ??件??: "??????"<[email protected]>; ????: Re: How to calculate "field contribution" For 2) https://github.com/subutai/nupic.subutai/blob/master/swarm_examples/README.md Regarding one, try to completely redoing the installation according to the wiki :) The latest release 0.3.1 should work :) Best, Pascal ____________________________ BE THE CHANGE YOU WANT TO SEE IN THE WORLD ... On 24 Sep 2015, at 05:23, ?????? <[email protected]> wrote: Hello NuPIC It's me again. I have several new questions and hope for explanation. 1. It's about my last question "How to update the newest NuPIC version". I pasted the output of command "py.test --version" but I didn't get any reply later. Are there any new ideas or suggestions? 2. Now I have a dataset with 50 attributes (including attribute 1: date, and attributes 2-50: scalar value). I want to make prediction(attribute 50, for example). If I don't take sequential relationship into consideration, and just use linear regression, there are relationship between attribute 50 and attributes 2-49. If I want to predict attribute 50 at time T depending on previous information, I think I need to use TP of HTM. But I still want to make full use of attribute 2-49 at the same time. I found there is an item "field contribution". I want to know how does it calculate. Does it range from 0 to 1 and 0 represents no contribution and 1 represents strong contribution? 3. Maybe I will get a low contribution of these contributions. Is it possible to encode attributes 2-49 into one long encoding, and predict attribute 50 depending on attributes 2-49 in TP mode? Hoping for your reply and thank you so much. Yajingfu
