Seeing Inside HTM Algorithms [DEMO #9]

Zitierlink des Filmsegments
Embed Code

Automatisierte Medienanalyse

Erkannte Entitäten
we want to talk about things that interest them and all my took 3rd there is no incentive to talk about my attempts to see inside each tendons and using commodity is that markets is already helpful introduced that stays the sun is very often misleading and his his comments that actually did a lot of substantial work on complicates this and and the integration and fantastic to have components running on a server on the JVM which has a lot more capacity than running across time wouldn't to and so on so that the begin thinking but it's because it's on get out and it has a it's able to run in the browser doesn't realizations of and what what what I just demonstrated that really quickly so and around this this is sensory-motor data stream which can have an so so to just get you oriented in this in this space on the left we have a representation of what the input is in this case it's just like a sequence of letters a B C D E F G and J J K like today and it's currently focusing on 1 letter represented by the black line so basically this letter C is just the input into 3 categories of K and because the this is sensory motor and case study the there's not additional leaving but which is the direction that we're going to move on the next time step the circadian you might call and that's coming from the so this region again it's like a category about the direction and amount of the the movement these fields here there's the activated columns in
this case these 2 layers platform 3 therefore receives the sensory input on the proximal 3 4 2nd segment and also receives the modern important as a distal input to the distal segments so that's just to get you some of them are into that what these delta that but it's not really well but I won't talk about this but I want a what do in my head was the temporal pooling algorithm has been trying instrument to properly algorithms explain it if you don't have a temporal willing and is you can look it up and the community to wiki and itself and basically the idea is that differ stable but distinct representations in higher level regions of layers so that when it does that by keeping cells active all over time while the underlying sequence is predictable so that the problem I ran into a been playing around with implementing this but as you as you go through the process you realize that there's a lot of it's not clear how to do it exactly and and you know what should be what should be the relative weight of all these parameters that you faced with creating and supporting realizing a need was an ability to visualize what is that what the different influences are on this on the cells of becoming active and because it's a sparse representation it is that's feasible there's only a relatively small number of active cell and each time step is usually the the numbers using like 2 % of the number of columns so you in this example but what I'm showing here on the right of the individual buyers here are a single cells that are becoming active so it is the total of Adzharia 12 or 15 cells are becoming active and each 1 of the color represents the and the amount of influence causing it to become active a case there red is just the proximal excitation is represented by these these lines here this approach much attention striving the young are related to other that's the boosting factor which is by the algorithm and that's very encouraging goes to become active hasn't had become active recently so it it it surprised me actually so that's how much effect of boosting is having an that's obviously I've said that money but I'd forgotten about it was invisible until now so this is at the beginning here is just proximal forcing but if we run this for a little while just 100 times that so so it will start to build products so that would be the 1st layer will start to be able to predict what what's happening and consists either in the time series plot so this plot although it's the same kind of structure showing a completely different thing that in this case it's showing the distribution of the states of the active columns the red being active and unpredicted active but not predicted to become active and purple is correctly predicted to become active and you continue over time more a higher proportion of the columns were predicted as you expect and just will just the emitting comes there and this this 1 sequence of letters of course it's been predicted again that also incidentally the light represents also constitute predicted but did not commit to connect so if we look now what's happening OK is going to turn out these and uh taken say on the on the lower level this cells which are predicted but the red and the these represent active cells the red indicates that there the feed-forward input that activating noses is not part this was not predicted income be rejected because it's coming from the input data into something that should not really relevant but if you look at the last 3 which is the next layer in in the stack and this purple here is indicating that what's forcing these cells to become active as input and just 1 so the the purple lines this the synapse representing that the proximal input but that's turning cells and it's coming from cells which were pretty predictable submits and aligned sequences predicted and in that case it adds to persistent level of excitation called temporal pooling level which again this plot is the green line so it's that which is being which is responsible for forcing some of these cells to become active in higher level
regions and if we go to just under stepping forward in time here at the you have a situation where the some cells become this set of cells become active and because it's from the predicted but then the following time step you a large amount of persistent excitation which is the green lines and that just decays over time until they cells turn off if you step forward and you got red
coming in at the means something something is being unpredictable not not predicting label or in the temporal pooling tends to turn off mic the so
this is how it needs to be just and calibrate or check that nothing crazy is going on there were just crazy things like the number of active synapse coming in activating cells is a weighted like 5 synapses so this is good to be able to check that that sanity check out and to start to get a feel for what the their relative influences of the temporal continuing temporal pooling activation and now you know how fast that's dropping off with coding causing to drop off as you know I haven't really explore this much but that's just some very got to have of giving you some idea of you know the concepts involved in the algorithm they found in questions depend on to the island lying on the charter right down here and I notice is served to going forward and in that case it is wrapped around the world that's gonna commentators something yeah but just so I'd like to do is there what what we should have as it is multiple words and then you can have microcircuits prevented and high-level circuits going to the next word so you can start actually when I was originally hoping to look at was forming a higher levels sequence memory might pretty actually pretty the sequence of words on top of like in the next layer and again now I realize that I've got no fundamental issues to do with how the algorithms implemented for a consultant so what's the impact of the predicted I line so you're predictions are the letter that is going to be seeing the so this is not what is it that contribute you minimize because that did online yeah that's not a prediction that's actually what would be the next movement is going to be right or some so it's not a prediction but it's it's the next move but is that that is also yes this since this this to input 2 inputs 1 is sensory input once when employed in the where the input is only given in as distal input but it's always anticipatory input that interesting and different ways in different choices that remain there as well about whether comes in approximately distantly where also have lateral and that'll distal connections in life or I'm not familiar with the current thinking and actually down and say hey process thus as as true but I see the this is then Js jobs refrain was compiled objectives for the closure enclosure I did you have to put the entire algorithm for did you get that interested in this course has been doing some more questions few
there I meant to say something about its is it what it is that a page has a link to to his is just as an interactive Wonderment different types of examples of the particular in
my life and I found myself family in my family
Folge <Mathematik>
Kategorie <Mathematik>
Sehne <Geometrie>
Zusammenhängender Graph
Folge <Mathematik>
Prozess <Physik>
Gewicht <Mathematik>
Total <Mathematik>
Zellularer Automat
Plot <Graphische Darstellung>
Rechter Winkel
Arithmetisches Mittel
Einfach zusammenhängender Raum
Algebraisch abgeschlossener Körper
Folge <Mathematik>
Prozess <Physik>
Interaktives Fernsehen
Zellularer Automat
Binder <Informatik>
Objekt <Kategorie>
Prozess <Informatik>
Wort <Informatik>
Quelle <Physik>
Familie <Mathematik>
Demo <Programm>


Formale Metadaten

Titel Seeing Inside HTM Algorithms [DEMO #9]
Serientitel 2015 Spring NuPIC Hackathon
Anzahl der Teile 19
Autor Andrews, Felix
Lizenz CC-Namensnennung 3.0 Unported:
Sie dürfen das Werk bzw. den Inhalt zu jedem legalen Zweck nutzen, verändern und in unveränderter oder veränderter Form vervielfältigen, verbreiten und öffentlich zugänglich machen, sofern Sie den Namen des Autors/Rechteinhabers in der von ihm festgelegten Weise nennen.
DOI 10.5446/18063
Herausgeber Numenta Platform for Intelligent Computing (NuPIC)
Erscheinungsjahr 2015
Sprache Englisch

Technische Metadaten

Dauer 13:58

Inhaltliche Metadaten

Fachgebiet Informatik
Abstract Felix shows off some really interesting visualizations of HTMs using Comportex and ComportexViz.

Zugehöriges Material

Ähnliche Filme