Merken

Q&A with Jeff Hawkins

Zitierlink des Filmsegments
Embed Code

Automatisierte Medienanalyse

Beta
Erkannte Entitäten
Sprachtranskript
Europe from our I I opt to stick to the extent that it so this is known as a very all this is that was what I have here all get that's the key is found the use of what we can do this last time I I really enjoyed that we just had a curious session of jazz and values in the system that we started our new temporal pooling layer so that point but for this to be idea what I would like to use invite you have the questions that you want to stand here and even if it is a general question in the room just really to to explain the case so this is sort of a chance to take advantage of the close proximity to the head of of danger that there had been so and of the sum of the right to know that the only 1 that you use this knowledge awards by so this is your only chance to talk to to me if you want to start with the reverse chance but part of that and I'm sure that any time you like you you know this
is this influences from the left and everything that needs to be some black question during that I'm not going to so what from the the face of it that your review that so but they have to jump in and the we have a question about some of the data of them when this knowledge of the year so I mean or anything in any questions you know this is due to the the fact that we start the and ideas this would be something simple and in a way that no 2 with the result of of the of all of the hundreds of time when the world series in and in 19 times series of time you have a lot of the uh well Hong is a kind of model and each year was the sort of when they had the feeling that I understood something to learn from the presentation of these and then you can do this through the form of of the drug has the effect of this is really a question so that the charter and the charter is just a couple of to catch with 1st of all the the model we'd like it to be shipped out of memory model is a model it's related to services in Europe where it can predict what's election no time so it's just say up this was next week and during the next system the system that that the fluid region so where where the greatest amount net something something of that happens because it covers right or wrong and can predict next so inference model so if you try to infer that you try to learn sequences and regular you actually don't need to encode time itself all you need to do is making a prediction of the next the the constant and that so that was just predicting 1 data point at a time into the future but it really doesn't work so that's that's for inference I only action codes time per say that kind of model for uh just it called transition right using the approach of you 3 each him with the answers to the time series of society and the use of all the 3 that sequence which is commonly followed from the novel of the of all of this through the at the time you know I
mean that you can all of the idea of a series of ratings referees question and I am just learning that she and ontology should you may not be thinking that we don't actually find itself because the oldest 1 curve 30 ms of water is sense of like absolute times greater than all of the result of a series record a variant of the model that just interested to learn that and then you know the series is basically matches to produce all of these things on and what its constitute predict what's going on so from a different point of view is that we using when the sequence right beside you will see that it was just a sequence of stories that distributed fashion and all equal the same memory structure to constantly people think that time but but that's you just basically sequence and then when you feel a record of finding that can I with sequences in basically make correct prediction what's the cost and keep track of the higher order of this problem I think it was a separate problem would push back point that recall but robot motivated and we have nearly doubled the price so a lot of the stuff we have right now actually placed actually which he could do it artificially of so that would require some sort of like you know it's just my own security next that there's a lot to do with the different religions anomaly-detection with the community and 1 at a time that help you know due to some extent evidence that there are some features of the time when there were a number of moles socialist and connections these kind of matches the distal connection is also on the the 1 close the body of the approximate inference so there are a number of in the universe have somebody will tell the story because of all the points on the presentation which was the product of light OK and so the real of tools and here's the deal tools were part of the time so OK yes so so we we the the synapses racks and then this is gets a people forward right and that's not and then they would like think that this this just the little yeah the poem coincidence detectors right and I really didn't like connections from of about so basically using saying this cell is learning to predict what's going to happen based on the other cells actually got so it's says you might recognize state a moment for our collective and then I completed and in front of the user is going to see that this is in contrast to the so is this is all about is learning where you get the Minnesota become active based on some very high or maybe of very 1st part our states as 1 of the of the 23rd this melody I reckon that important next so so those thoughts of these actions yes or related tool tools or they can be you know depending on you know this 10 step before you were some prior before so these basically Americans universal nearby and have at any point time due to the broader but
some of birds so become active become active they say no 1 who was active nearby regions and all of the connections them and that that because my predictions based the cells like seals in by that that they were exactly quite active the fire that I'm going on so that's the this is the transition memory you well example of this is that this is not an end in all these questions came in my mind and we really need to know what the thinking in the health sciences they call all of you know the memory of what the a number and you know something which is I don't know what the few seconds basically so we can memory the the past user or maybe even more than what we this is the toughest topic we start talking about even biological worry about all the time scales at which we can have holdings in our head are people think that for coming to terms the temporary memory comes at the price of gradually you that we talk about that but I think the the basic thing and basic a question about worst time story stuff is still in the in the majority and absolute that the lack dendrites that the schools and those those Russia basically transition states and it is also 1 of the things I was trying get through my it's really about real-time like can out of that something properties it's probably have different color that some questions right from the simple things that you know where we learn that the transitions and time versus sequence storage through here E cell is learning on its own when it becomes active 1 of the things that you can do that from the context in that cell and cell of course is usually used over and over again in many different representations in many different temporal context but it's always a combination which is unique part of the story that you know the only other result in the audience of a certain amount of over over again used all over again and ML many many different context but it's a combination which is know so that's why all members during the start of your everyone else this is 1 of the flow of content so of this is participating in a melody they could be produced in many many different knowledge many points in the knowledge the management of states of all i so you're going to give you the distance and the weak or talk about what that says that for all of them in sort of subsequent year here a year yeah we did that and it's not such short talk on using the community the good for you to do is to group you have the right well that they some parts of regional context is sitting on top of a whole bunch of other stuff and not really know knowing anything about you have that noticed is that those anywhere so this not just sitting there it doesn't know where it is not just another really recognize distance this and but when it does is it learns how to effect behavior by connecting something else below what that's doing that could be defined as a form hierarchy that here through reading and through these of course partly in New smoldering stuff that here you know a little cell here I can reject it down to the regional or any and all that kind go over the use of a drug control anything below but in this case to the on influence of that have and but this is the key to hold the key to the whole of all things in a higher than the stability of understanding that that this research project is modelling everything in the natural world and build tries to build a model of all this stuff begins and ends in projection layer 5 cells to cells generally and understand what social belief that the 1st phase of the war and during and even know what projected into but can also should to control some of the errors in this region control this guy at this thing or associative so I what the but I have to go for the the the the big insight for me was that if you could do this all associative linking not all of the social and this time models the behavior of the thing in the imaging control and that control of the novel which basically says that when you're learning explore areas and now I'm I'm going to invoke it what it would have been involved so I I can't through whole talking like this that you know this is what we do is we well what what I don't understand the key part of the it departs from with right now now is on 1 of the problems are 1 thing about this is that what we know is that you know his his his cortex Holocaust was applied the is so freaky just black in the center of mass other thing about all of the in the we'll have a pretty good idea of how this guy were to control this this is this behavior that where other cortical regions of things so that we have a document from and thinking about the potential of generating but I don't know yet announced that starting with houses decide what to to do and you introduce a golden behavior what we all know what we call the assigned says about the context presented as part of the province of portal and involving which you literally means lower stock and that affect the extended to select groups so that all words of little things that is that they go with to different names but it is probably going to the wall and what happened general belief is that the content presented that there is a uh a set of possible alternative behaviors gasses like these are things you can't know right and they
chooses 1 there's some magical mechanism that they had no that's going to execute that so there's a lot of existing psychology another theory about you know how we evaluate the behavior is how we choose behaviors related areas is that this pathway is a complicated things were so little region and because of the of the vessel and and all this this this basic idea that the game again decides which to execute this really almost those really the challenging that you think is introduced into its
elements like that this like kind of mysterious and so on and so on sorry so I struggle with it and how is it that we we were trying to to model cortex of how is it that we want to 1 of them was the desirable outcome how we used to come from the set of all we will weaken structure of something like that probably pick the right answer it's it's a speculative particularly the the selectable what mechanisms which are in place in my head reflects the coated with but then this idea how to generate you know what is the what is the the the the function for an optimized on how to do that we become support structure that the designed to do this because apparently according to some of their own and that is also 1 of the problems of so with and if I look at that look at the of the regions in the layers and of means for you to really go on on content 6 and the cells are generated using the a power of 1 the work with the complex they also reject up the hierarchy while they EHEC input from my a pretty may also be seen in the volume of my 611 finances a lot of circuitry here that I don't have time to do and have anything related to a kid gets a basic thing but that was a very common in the literature that solidifies right then in response to but is it OK well aware of the to work with just those that have the right so you can find them I don't really understand relationship period by faculty it looks to me like where 5 and 6 are almost almost competing with the formal review the the forward of the content of the book and sex and actually find solace from here and what we important to really look at the of power law is this is this is almost to begin with so this sort of which is that the life should you just it's just joke so you yes it is 1 thing that understanding that have happened in the way I but it it it's not there's not there I saw lot of was on the right of your really close to the realism standing within the social and the and and the a naive implementation of be used and objective function present generic reward yeah you you see that when he what I was doing my career that right just like simple to write 10 today are what we don't know the emotional states of all of them have do something with simple problem to be a useful and share the idea that we don't have these guys I have no intention of this it's kind of weird stuff it's got all kinds of and this is like Moms 1 little thing but still it's this still is no if I could just hear that somebody like that they need to solve the a problem is a very very clear problems and often the the trick to solving result of things is to get the right exact probabilistic drops off and it actually part that and so I struggle to a transfer what what exactly is exactly what were what would be the minimal problem we can solve that actually is the heart of the problem and and our hope be something reasonable but it's it's complicated so it's not that there's a lot of this is the room around here it's it's accurate talk to intelligence you know I a lot of pieces of information in real life how it can be that hard and making it possible to predict the work that is also 1 of the examples but anyway so the pilot during or playing around different models of this kind continue players in most environments in which for nutrients and figure out what's going on exactly here and some of these technologies that we have I think you have just for you to do it but they actually have much to do so well I want to find that so that's what might demo was devoted to think it makes more sense if you are launches at this could be the the importance of ensuring mutual relying on the moon shot people have told me really happy about that have happened the of knowledge workers in the sense that random processes you will be so what is it about the 1st thing is just imagine what to we to be the 1st thing
the basics of have all the rest of
this century this is the nearest directly translate to about itself it was OK but the guy's got what we have to put it in the context of the political economy of the activities of the service so that he could sell space to store all the rest of my result on the vertical neurons to double that make all knowledge is another only if I just turned itself in all cases so this is a generic cells generated and the application of I'm going to talk generally yes no you need to follow set of a so and each cell surface signal which comes in the reason that in the all will you know get a the go generally kind of wealth to this the ones generated various function by model more and more but the created they're injecting rises because this will destroy and self if you out of the current from 4 different cells in each of these such that we all have different slopes so the alright to rates of the cells body to the 1st part will cause which has that they the activation of the cells that are currently comes in comes from above proximal dendrites and this is and what's in there are added together for the period time leading to this this that so over time you can simplify that by just saying that there is this so important that to a certain started in this job on here so they they ideological as a and if the influences that 2 cells like that this will be our 1st cluster basic to the rest of the bases they have all the stories that model that's that's how all that's how we end up with 1 presentation of the results 1st numerically correct so that model I have a protocol works is that there are 2 types of during the normal curve under this so here which actually is so complex that's wrapped around the columns of the matrix this has something here which has also that that's what we would call BST is like the best thing is that valid now known as the argument resulted in fact this year and the tumor cell that's that's that's that's part of this in the following our son's interact in addition to the these guys and you they all respond to the forward In sum up to a point but you have to get the actually so what happens is is that assuming that most all of these cells are also that's the problem and the other is in the form of and we also have approximately the same response to the various like what we know this so here in itself has a faster response to the same input as entities so but that means that it's going to have a higher level than any of the individual cells in the so it will fire 1st now right now when it fires its sends a signal language to its brothers and sisters you know in the results and they all get killed off and somewhere over here before that signal reaches the same thing happens in the call that since the same 1 signal back this way and that means there and there's no activation OK so the cells are protected from being defined in this way from elsewhere so so far this is exactly right exactly right the continuing to get the same feed forward in their place which means they all get to get all the way up to top and they old fart comparison of all of the steps in sequence when they just continue because they're never going to stop and international onto yeah sorry the fastest 1 of our 1st and 2nd pass and that is because 1st yes that's right so long as they know things that you don't traditionally always the this is how all the work is that correct and the first one to 1st coal signal and its variants well basically having all of these cells still try hit the 1st time very quickly that 1 went out because still those of results and 1 of the ones that will be his 1st policy of well I allowed always fire that you have that when the virus gets the longest of the same thing whether they it's that there differential in these results fire you can get hold of with the knowledge that was the 1 norm of the object articles that the fall so you have things that can be used to get some of them have evolved into a typology of decision yes yesterday this includes the title so this is all of the young on the right is would know which you do have prediction 1 of these guys here that the priors for each these information correct it in addition therefore is vital for the inhibitors of viral that's why 1 cell and nominals that Congress president estimates some mechanism here but when you start fires of of this is where I will not let us in growing even after some of our first one is that when in the other so it's just generally the yeah my model-based says that it is not firing here that this which that's another recognition site of the associated with that happens at a later point that should the 1st and the image and and 1 the new structure I don't understand if if they fired and therefore not because they're still there this is still this mechanism that says to meself wires cables right and yet we're trying to learn out of the colonization the 1st people
that happened in all environments and what was once all fired in so this whole personality is not something is stable in number yeah OK medium with each step of the same thing happens again the need of the ideal of awareness that the other class to this thing is that would have so this cell when you get a prediction around OK that's going on we have so not usually better than that and this ask the and then also itself but the effect sizes of his stock-picking begin its really good quality but it was sufficient justification units of the data novel but this is the direct action potentials of of the of the rules of the result that we think of right so if you have predictions of the highly sparse presentation yes and what you get the same color murder absolutely so that's simultaneously executing code would have what we have to we will know the old the year old on the last several months but this is the basic CIA although and 1 of the coordinate is the 1st thing that happens is pretty well if it's good if there is 1 that just still happens on the basis of commonly out here but this happens 1st people that was that will always it this in this cell element polarized this paper is biased 1st and just and 1 of the ways that women were correct but in the implementation it's that's me so I I don't think part of of work you're implication of this is that if you have a person so if you have the wrong person and then all of the cells here learn the forward that matched the beginning there is no and so they tend to converge and around signatures to whatever the columns of represent and the you have predictive context sufficient it content and also especially on this feed-forward input and within that context that going yourself so the learned specialize in the forward so you have predictive systems spatial houses put them into the of that's the reason that this because the act of template really happens when sparseness is meant and the cells learn to tell the difference between the sparseness and bursting In this concept that we would basically relies on prediction you want to call it detected that users with partial internal limits 1st that so far easier for you to think of the basic issues of gender so you for right column here 1 and in the case where there was no states and you can solve region to have all this is that the use of now when you have correctly predicted transitions everywhere was you of 40 cells that 64 and when the person in the so yes tree you have these cells which have learned to rely on the challenge of the data it's going on fire do so for retirement sampling and that was occurring phrases and up and we talked about that so far there is that actually layout for so I'll think it's here but I we can tell it's it's sort of like exactly how it works it wouldn't have been in some but I'll see you in this that if I had been interested in understanding of which is the shortest amount of time but now the internet internet relating to the same thing happens at the the same time you're in yes yes I got all of this is called the opinions of the lower classes have a way of being 1 of the products 127 so in here you really were those phonemes only 40 activation of very specific versus a very specific ontology editor that only a very small number of the connected nodes for himself and his you know who that and then the frickin literature but we can think that because I think that's actually going in this model can have a very small number samples that you start with that of and real time until it was too much no part of this is that the cells are connected to the very small number of these well yeah transitions in here you have the signal is so poor that those cells are basically they're going to connect to several of the things that don't change for example and the ground also inhibit neighbors so that their neighbors on learning any of these currents this process transition so they're going to be an advantage to how this relates recognized as father of 1 is that when you have sparse and you get as you get temporal pollutants in literature tree you don't have sparse that's as well that's right as well this is what you get non-sparse hidden layer treated well yes and that gets transmitted and later following OK so basically what happens is is that the system is trying to have sparseness everywhere test the predictive models of instructed trying to have started everywhere and we can do that because 1st thing that propagates well that's not necessarily true of integrated because of it's not clear where we're probably going on with result and it's not clear that collection of them the direction right I happens is that the the cells and these things 1st and then here that causes 1st here nest that that was going on and off I was not in this regard and therefore screw reducing what is going to be tree and what it's trying to do is generate behavior that maintains the sparseness in layer tree so it's genuine behavior and learns to associate whatever instead layer 3 which is the origin of the world where the driver from you can concentrate on the ground right of layer I going signals from it's not clear that that's the of phylogeny I invite
you to do it if you want to want to talk about how do think that's the purpose of this nature of the data that you can use the propagation loss and there are many ideas in general overview of the general idea is that you have spotted stations injured created with input from the that's what was used to propagate you will so you can do that depends on the wavelength of the unintelligible theory the question that we have a accessions exactly how that works 5 not clear it's not clear that those that it such that the center of propagation 1st in common sense knowledge of what it's doing its own thing and so I would only apply proximal so what is the final model and that way we could sort of predictive signal so I think there's a lot of organizational research on the school sessions of and I'm not against talk about this but I don't think that you know therefore 1st getting copy of the of the incoming sensory motor it's it's also getting a representation of 1 or not as stable as throughout what I'm saying is that it's generating behavior that associates keeping that cycle components and trying to be here and that leads to a prediction yes I bridges holding interprets as sparse signal coming from the guy who had percentage of that you guys are over there is 1 of the gold in services that make predictions comfort you or make you more predictable so but the question still is a question about how we evaluate multiple behaviors and how do you know what has to be designed to you all i so back to the problem at hand for my suggestion is that the answer is that it is saying that this is a forest if this is part of a new that in layer 5 it's getting the same resources His on the spot of sparse signals from their linear 6 pages of the sensory motor so and it's trying to also maintained its fast and stable representation of the tree and so its representation in all areas where there is a sequence of very sparse representations in the following the generated varies cities yes there is very small subset of possible behaviors that were maintained in the by trying to read all of the details of your heart and natural and that's 1 of the things that they did not agree with but I think this is 1 of the things we do not see same there cells here people from the golden varies from the ones on the well they didn't they want to know I think you can explain that understand throughout the body that has to be a part of the it's not just that they're the ones that are related to that it is that some of the things that think that part of the important things going on there is and how you have the support of the particle and the modifications in human human situations and in general and this entire in general this goes more stable and this already know most of and so as I go through sequences here in layer 5 this is the time to like this sequence is 1st year of the Family what isn't through that these all these layers of cells have and interest of little 1 all of feedback from this region all and essential happening is as both of these sequences in a stable representation of the 1 on the move but this is a higher-level concept is stable and longer sequences here and every 1 of these cells can be associated with this stable rotation so in the future not right approach is stable and here he is always located in How would cells become people and what what that union-represented that union represents in some sense all the possible states of all the sequences associated with this so that we might be it multiple data so you can tell a goal is to employ of what can happen multiplication compared to get attention on its and all the time about the kitchen and basically associated with multiple sequences here the individual and the different steps and we get the kitchen 1 case the time of the the when is can you book all the watches and you know all kinds of errors that led to this more stable patterns and then when it went into account in each 1 particular 1 having variety of that and they will pick up 1 those sequences exactly and in the and all this kind of people have been you know people they should but I will write specific instance during sense the sampling that kind of statement from the list so I think it's a good kitchen is a higher level you know I'll be around 1 achieved by the union of all the paths do that and then I have a specific and what comes in more here is like cigarette itself those personality and I don't actually need to starting to exactly what you find out in the Soviet Union of these things you know what the politicians and so on so there's a lot of this this operation of about this basically involves the union of behavioral sequences of images that then 1 particular it becomes a universal exactly what do this follows as long as the same thing here is called so that satisfies obviously if I sort of 1 and similar things happen lately so I don't know and what was actually that's saying I'm expecting you see something on thinking you perceive something like look for that along with the you for the cat and involved I sent union of cells that are all polarized that all the possible patterns that you might use of all that so now I'm going to come in and even if it even if it's from ambiguous in nature in that context and yield a higher that's part of presentation media you can come and that should is but because people as the cells from right need for that that is biased in this case the idea being that was going to change that happened but this is exactly what happened to have to live 3 cells that's happened only once again so that suggests that this problem starting with an environment where they really is related essentially pattern and if virus is of order of and and so you're looking at the kinds of of in this year's looking sequence of
inference and so all this is 1 of pieces that that this would happen and they still they still have the whole picture wrapped up spills story but all of these variants and emotional problems I think maybe honestly this what you talk about the idea we should let some of people have a chance to ask questions of the Japanese couple the approach to understanding of inhibitory cells and their role in the temporal pooling I don't have a fully specifically yeah the here we just think about so still working on exactly how we think that people so that's the hard part of well I'll have a series of brief introduction to the Semantic directives of them and we talk about some of the world mentioned into so this is a general rule 75 per cent of cells in the cortex oxide and Poindexter it is not general will be generally the individual cells are not if they are not going to be a and a lot of reasons suggested this is your change but this is this is a general science believe it could you are right we have described it lot of other topological evidence suggests it is true that in the sense that we'll look at connections between inside cells they're very very sparse like you you almost never see 1 the very rare since I something synapses on another 1 but it's really common in inherited itself so it's all about the politics of ourselves from ultraviolet brown and in looking at an image and sound children's inter-cell like connect many many synopsis all these guys in ways that just click the shut of that no innovations will nothing going on the inhibitory cells they they they they impact on their land and in 3 places they they can form synapses on on the cell body itself which excited resonances never do only have interest in Cyprus and is always on the the and they can also form synapses on what they called the very beginning of the of the acts of the whole place in the interaction with the actual action potentials start and they think they form synapses there a very strong base of polarized you are you're not generous by that is not part of the market so these are these are multiple while from individual cells the density carry information other than just shut down and the 3rd place in is along along the dendrite to be part of the world and I believe those are changing the threshold of the difficulty of building a model so that this is the way that this is a project in terms of prediction turned out not to be generated to image to image cells in generally not to the information carrying a lot of reasons for that are there are there many classes of individual cells depends on counting but I think what final some of the most common 1 is the 1 that is shown here that they just want cell becomes active reactants inter-cell interactions and region of virtual on very fast this new wireless for sparse optimization this is essential for vision pairs of variable documents there are also inhibitory cells which have a very strong common African and so they literally defines bungled and interact almost literally surrounded by many people and those 2 classes of those things were relying on the part of the way to basically well what they but they seem to be doing some people say the cells and the cells so if you that at some common words and the other ones which essentially the quality of Internet anymore all the cells here are not given this what would lead to a comma activation because you haven't heard that inventory of inhibitory cells in column fashion and so those are those of the general idea is that although there's lots of other flavors which are complex and it's going so we want to here we realize the students about this predominantly in both the spatial pooling function of selecting the column and then it is quickly and then selecting 1 the cells when the world was talking about so what we think is what happens is the inhibitory colonization and quickly itself will so far in this is the part that will so they're not 1 you have specific intelligible speech with you but I'm getting naturalness let's talk about we are still working through and will we will still have the right how we want to work with it and we model it mostly on only 3 and 4 and so we're basically saying here we're saying that if I have a fine the university 1 of the 4 position columns that should be that actually probably straight up until a really 1st account here it's said nothing nothing happens from all of the general the challenges is when you have a good predictive model in the form of being a very sparse activation for the most part of the reason for the but now you want the to a stable here and sort of the 2nd 1 based 1 hand this many of these changing and here matter that there will be we started off and this is totally just you know work parts of this very moment of political I can turn of the lower back and that's about and his invention about your specific questions and this is a little bit later on this idea this last 1 question is when cells are likely they trying to so must have to go public which almost all the approximation to the right here and what this analysis of multiple patterns down here and we see that the meaning right so we can find out how the capacity guys and can itself we did the experiment this this week 1 of the things we couldn't much higher that but the a person level and and and what we're celebrating history resistant OK when you can have 1 cell in and that 1 of the spots columns where the partition of one's own found now working on model and actually might be in morning and that might have like 1 cell and quality and
with the world War distributed less sparse representations of reasons for that so that there was a still appears that essentially saying I'm so basically they where's the histories so that you going into some until the stuff and that the telescopically important version of this stuff because I don't want to be in a lot of anything but I think he's trying to sell people where is where is the inhibitor so can play a very simple level we're going to a lot of these guys again here because we say well these kind of response representation active and all of the cells and where there official you by the so far so that's what I'm not sure if you follow would result anywhere else in the next so maybe you have some specific thing in the basically where where cells like each 1 neuron at number of the these results was wondering how that activation stops when the Patriots pretty how home want something so in general the general will be worse so far is that we call instead of always it's there it's not a transitory state it it is a temporary state where all the cells 5 2 is 1 of just over 5 1 9 and 1 and when you take all of those and in addition to all of the equipment by I 1st and then I still want so that's always going to have that happen only 2 of the 1st right that's that's just quantities along the and treaties that we want to generate that mention as some of the situation resignation of knowing that a a lot of what you really want that but I think it's a lot of the others in the questions and because of trying to present fault that it's not really fall under the individual flair for pompous so the sensory data come into layer for a place through both proximal and distal but is that just implementation details in the code well at the moment this is the implementation of the moment we're solid form not sentences coming to the end of this century voted this knowledge is close to interest rate given the current sensory signals signal and 1 of the many about what it's interesting that we feel that this all uh yeah but not matter a in the there were in the problem the idea the answer that and so have cell here in the forward which would think that sense of right and then on the of the disorganized you're getting a you're getting more and more on your instructor well what you were thinking about what we generally what you want to initialize is you want and you want have a labor force and here the other 4 cells represent high order state and
that's what we really want to predict wanted a higher state of viruses virus but here that you don't do that because it's not our pattern can some people see sequence so that you won't have a national or self which we don't have the rate so if usually you know we know that something like 60 miles if you want a sense of the Senate and people in absence of labor still come from 1 of most of the time I and we know that there is a motor commands coming into contact through the analysis of the subject but we don't know where they don't identify so work relating to there on the of religion the will want read from but that's that's what be here and the idea behind this is roughly this research OK we don't want to have it documents that of course but I will you know about the law the beginning here you would you want feed into the little synapses is not so that wasn't very 1 but them so the color representation so you wanna say I'm looking I'm about to the right and and and I couldn't see and I and I don't want to some high which they want to stay on and I know about you but I have know of the I and the right and so my original thinking is that when want small century there you really want How tolerant of the observation the that we coming from the set up but lots of folks so this this sort of the models OK this is a lot higher state of the current but this is more convenient then can predict what's going to happen that there are some complications here we do know that most of the conventional here we more along with well you know there's a few connections from 1 of the but monolog but there's something I don't we can start to think about what you're doing so you know that's the state of all of this stuff this number here and this year makes a lot of sense in the fact that the cells are not highly interconnected for the loans all of that was there but we probably don't understand the non-standard was on the connections and we don't we don't we have work all these of how this is working for on models we basically take column representation used probabilistic so waste basically works in this case the but closed and now you have a slightly different implementations of exciting for starters sequence summary of about well so of multiple cells yet well need a model so the columns represent what we actually do have also some already in that that was the year that we create the high water sources say that says this is all I have gotten by with so you can have these so these are the things you need to
represent spatial relations that this is what this is not just my time with below that of the of different from the so do you need to know the variance you have a lot of so called over the loss of the same thing here is that this is just the beginning of the book book new things and not always special state of an individual person out is what was teleconnections here there's some small percentage or as an state was going to have to get speculation was and nobody knows what states but it's nice to know that this is getting it was almost it's not 1 the worst of all of the forces of people from so we can find the most successful in the success simple cells so you can say OK in some non-Irish states from the states which experienced by the way that it will be important so so this so widely accepted intonational just like words and the pattern that was used to some of some of things them we all and it really well know well this is very much in line with of the the fact that you could turn out to different groups of people but most of it will also involved if the time for 1 more question all and it's all 1 part of the system of classification that was just another person good way expression so so you can always write you get different world spaces that you know the brain is very efficient to it's the like phonemes generally follows the consonants and then and then you get words in the sentences that you that that paragraph model also and along the way there are choices to be made to follow you pay attention to the time frame in which you want to create a hierarchy you are what might be the mechanism year as to which the different time spans the entire interest press is different integral got this higher in 1 of them in the right here and part of different time of like that of links and warehouse that have decided what is I think the answer is it's it's all over that every region layer and went to have past we have some amount of passive passive recipients of the responses optimal pulling arm and it does as much as they can and then and then so that's all I can do that then it says OK the next guy pixel over migratory humongous monstrous so during the last years the water on a lot of in 1 a lot more than 1 then the ones we have to include basically the war that 1st reading before it to the next and but it turns out that the state Spanish in theory you could only a single where the neocortices of everything that it would take the average train you have to those everything everything it and we have thinking in the knowledge management and the memory required to do but if you know if you have an infinite time availability outside the lines and you can always trained in the traditional monsters she would be 1 and only 1 and only 1 but what happens is that you can can do that so I'm going to start the use of the general idea is that have a regional context here and it's getting the primary auditory primary constraint but instead of Gaussian distributed fashion over something during the whole moon virus coming out of this guy essentially locally at times a lot is money sequences can and it has a similar there are a lot again the work environment more complex and more more shortly after the meaning of the text and it's really not topological and that's it so that certainly in in the trunk he will acquire wants equal 1 and it's a little more about so these guys are as much as they can and then they use usable over this then represent the sequences that more stable patterns and they become the input the next guy and that's outside the same thing in a trial and inland waters sequences and this over and over governance where the words appearing here will depend on how much memory complex of the all this and I think there's like a word you this level and of all the so far cycle and what it can be wherever they can naturally max out and in the hierarchy basically solids solvable in real time from an optimum but you do get built in the in the short answer your question there is only a short break for things that you I think about words or something and I give you lots of examples like you know the same thing here the original model that word right now it's patterns that you're not thinking about you have heard much to school you very beautiful long vowel I mean was that there was going to write about the number of transitions not right
now saying that all us that some of the things you know it's something that we think is 90 ranking into little phonetic sound a little sound transitions right not takes me to discuss the number 2 this is the unit price and quantity and on its international model I think the timing and circumstances that kind of stuff that talk about the other time the exact time out would of the your question the question is where and are how you decide which parts of you would use in the use of the machine where but you have to direct models now I think other things that I did not understand why and it's just looking at sequences and there's all the contents of the projected onto a node of the context he doesn't know what this is all I can do is look at the statistics of the data and say alright what was the appeal sequences here where you know how many or more can learn before and I'm really how long before the memory and then so I could imagine for for example imagine that a language became model languages very few falling into very simple words of running I don't have a proper architecture to them something and you end up with a very different representations here that I had some other things the 2000 phonemes and you know it just makes it up as it goes to some of the structural model that much of the and in that sense of wonder that nobody knew about all these languages of all users were given example of you know where like the day today is really 1 to right now and the start of something to work group then that that becomes 1 of the things we found that all come together to form phrases you know that there's no there's no clear breaks we compose those things after the fact to try to describe the language that linguists right variation of these are words the landing on his life a subset of ontologies in language because they were put in the structural analysis of selected really exist and also the more you learn more recent good they usually starts off to see that the 2 words in it becomes 1 sequence right but what is you never use so that not all the good and there's also works against the you might want to do with this solution to your the and you don't know language the language to analyze the frequency at which components were the independent of each other like novels in corollary frequently and many words yes you're going to create a hierarchy around those which would make it easier to allow the confidence that exist in the morning the right and uh because there's so many many students but you know we try to teach kids the structure like that I haven't by the the first one value that someone might want to learn was stronger and of love and I'll get you to think of you know that was just anything that was still it wasn't to her mind building that is 3 word now you know is just 1 thing and then the politicians make make tension and and all the time but the all sorry out but I don't think that the visual the whole of of the tackles the essential role of the way the authority of the book and the way we impose language structures after the fact that we don't understand why grammar rules we wouldn't be passed by working with something so this again on the layer all those languages all of its components down elements plot that really true uh it's an observation of parent structure that just is something that the language just use what what works and why the use of words that were part sending with music by by the late in the day all you know these are the types of phrases in the theory of the of the of resolutions calling for regression types of the distances on the reality of is the way that started playing like here and now that you know that there is a lot of analysis of the regression models of time but because of that they start non-starter that like in a language like that to start applying the sound of living a battle of substructure delicate marking to market over the years that will listen to the stream of audio and then will let like even through it and then you lose part of the videos of is like a very precise was later reported that has been developed yet there'e structure and the frequency that this neurons that have really fast learning he left his dream of the yet so I think I want to solve all this sort of thing you want to know what the answer we're every time I was the last hour and I think you talking discounted were prepared for the sake of I have academic taro over God to to say that had talked to to to to
Punkt
Gewichtete Summe
Rechter Winkel
Reverse Engineering
Mereologie
Physikalisches System
Abstand
Maßerweiterung
Schlüsselverwaltung
Quick-Sort
Schreib-Lese-Kopf
Resultante
Subtraktion
Folge <Mathematik>
Punkt
Gewicht <Mathematik>
Inferenz <Künstliche Intelligenz>
Momentenproblem
Wasserdampftafel
Gruppenoperation
Besprechung/Interview
Absoluter Raum
Zellularer Automat
Kombinatorische Gruppentheorie
Metropolitan area network
Weg <Topologie>
Datensatz
Informationsmodellierung
Prognoseverfahren
Zeitreihenanalyse
Vorlesung/Konferenz
Kontrast <Statistik>
Maßerweiterung
Kurvenanpassung
Datenstruktur
Grundraum
Hilfesystem
Soundverarbeitung
Einfach zusammenhängender Raum
Transinformation
Sichtenkonzept
Ontologie <Wissensverarbeitung>
Computersicherheit
Reihe
Physikalisches System
Bitrate
Quick-Sort
Roboter
Dienst <Informatik>
Rechter Winkel
Festspeicher
Mereologie
Codierung
Ordnung <Mathematik>
Aggregatzustand
Resultante
Vektorpotenzial
Punkt
Selbstrepräsentation
Gruppenkeim
Extrempunkt
Computeranimation
Datenmanagement
Prognoseverfahren
Phasenumwandlung
Kraftfahrzeugmechatroniker
Zentrische Streckung
Kategorie <Mathematik>
Temporale Logik
Ruhmasse
Zeiger <Informatik>
Kontextbezogenes System
Dialekt
Betrag <Mathematik>
Menge
Rechter Winkel
Festspeicher
Projektive Ebene
Schlüsselverwaltung
Fehlermeldung
Aggregatzustand
Subtraktion
Stabilitätstheorie <Logik>
Folge <Mathematik>
Gruppenoperation
Schaltnetz
Hierarchische Struktur
Zellularer Automat
Zahlenbereich
Term
Physikalische Theorie
Bildschirmmaske
Informationsmodellierung
Spieltheorie
Äußere Algebra eines Moduls
Inhalt <Mathematik>
Abstand
Speicher <Informatik>
Bildgebendes Verfahren
Schreib-Lese-Kopf
Binärdaten
Einfach zusammenhängender Raum
Assoziativgesetz
Zwei
Datenfluss
Quick-Sort
Flächeninhalt
Mereologie
Gamecontroller
Wort <Informatik>
Kantenfärbung
Informationssystem
Resultante
Fakultät <Mathematik>
Besprechung/Interview
Hierarchische Struktur
Zellularer Automat
Wärmeübergang
Kombinatorische Gruppentheorie
Komplex <Algebra>
Gesetz <Physik>
Informationsmodellierung
Globale Optimierung
Reelle Zahl
Endogene Variable
Spezifisches Volumen
Inhalt <Mathematik>
Datenstruktur
Tropfen
Figurierte Zahl
Schreib-Lese-Kopf
Leistung <Physik>
Kraftfahrzeugmechatroniker
Lineares Funktional
Videospiel
Ein-Ausgabe
Frequenz
Dialekt
Quick-Sort
Stochastischer Prozess
Arithmetisches Mittel
Generizität
Menge
Rechter Winkel
Mereologie
Information
Programmierumgebung
Aggregatzustand
Resultante
Matrizenrechnung
Prozess <Physik>
Gewichtete Summe
Punkt
Formale Sprache
Familie <Mathematik>
Kartesische Koordinaten
Element <Mathematik>
Raum-Zeit
Richtung
Übergang
Eins
Internetworking
Netzwerktopologie
Einheit <Mathematik>
Prognoseverfahren
Prozess <Informatik>
Stützpunkt <Mathematik>
Nominalskaliertes Merkmal
Umwandlungsenthalpie
Parametersystem
Kraftfahrzeugmechatroniker
Addition
Lineares Funktional
Güte der Anpassung
Temporale Logik
Systemaufruf
Strömungsrichtung
Ideal <Mathematik>
Partielle Differentiation
Schwach besetzte Matrix
Mustererkennung
Ein-Ausgabe
Bitrate
Kontextbezogenes System
Biprodukt
Elektronische Unterschrift
Frequenz
Entscheidungstheorie
Texteditor
Menge
Geschlecht <Mathematik>
Rechter Winkel
Information
Programmierumgebung
Message-Passing
Aggregatzustand
Subtraktion
Web Site
Computervirus
Folge <Mathematik>
Gruppenoperation
Mathematisierung
Klasse <Mathematik>
Besprechung/Interview
Zahlenbereich
Implementierung
Zellularer Automat
Kombinatorische Gruppentheorie
Code
Gauß-Funktion
Knotenmenge
Informationsmodellierung
Differential
Flächentheorie
Stichprobenumfang
Endogene Variable
Inverser Limes
Datenstruktur
Soundverarbeitung
Ontologie <Wissensverarbeitung>
Protokoll <Datenverarbeitungssystem>
Schlussregel
Physikalisches System
Paarvergleich
Quick-Sort
Objekt <Kategorie>
Druckertreiber
Echtzeitsystem
Basisvektor
Mereologie
Kantenfärbung
Normalvektor
Bit
Resonanz
Einfügungsdämpfung
Inferenz <Künstliche Intelligenz>
Virtualisierung
Minimierung
Natürliche Zahl
Selbstrepräsentation
Ausbreitungsfunktion
t-Test
Familie <Mathematik>
Schreiben <Datenverarbeitung>
Bridge <Kommunikationstechnik>
Drehung
Richtung
Formale Semantik
Internetworking
Homepage
Eins
Übergang
Netzwerktopologie
Prognoseverfahren
Mustersprache
Maschinelles Sehen
Umwandlungsenthalpie
Lineares Funktional
Schwellwertverfahren
Befehl <Informatik>
Approximation
Reihe
Schwach besetzte Matrix
Ein-Ausgabe
Intelligentes Netz
Dichte <Physik>
Arithmetisches Mittel
Teilmenge
Rechter Winkel
Projektive Ebene
Information
Computerunterstützte Übersetzung
Programmierumgebung
Varietät <Mathematik>
Fehlermeldung
Aggregatzustand
Instantiierung
Rückkopplung
Computervirus
Folge <Mathematik>
Ortsoperator
Mathematisierung
Klasse <Mathematik>
Besprechung/Interview
Interaktives Fernsehen
Zellularer Automat
Sprachsynthese
Kombinatorische Gruppentheorie
Term
Physikalische Theorie
Variable
Bildschirmmaske
Multiplikation
Informationsmodellierung
Arbeitsplatzcomputer
Zusammenhängender Graph
Materiewelle
Grundraum
Bildgebendes Verfahren
Analysis
Einfach zusammenhängender Raum
Assoziativgesetz
Wald <Graphentheorie>
Kanalkapazität
Schlussregel
Mailing-Liste
Partitionsfunktion
Quick-Sort
Flächeninhalt
Mereologie
Hypermedia
Dreiecksfreier Graph
Wort <Informatik>
Partikelsystem
Resultante
Addition
Momentenproblem
Selbstrepräsentation
Besprechung/Interview
Versionsverwaltung
Implementierung
Zellularer Automat
Zahlenbereich
Strömungsrichtung
Schwach besetzte Matrix
Bitrate
Code
Übergang
Arbeit <Physik>
Bildschirmmaske
Forcing
Rechter Winkel
Endogene Variable
Vorlesung/Konferenz
Ordnung <Mathematik>
Große Vereinheitlichung
Aggregatzustand
Einfügungsdämpfung
Selbstrepräsentation
Gruppenkeim
Komplex <Algebra>
Gesetz <Physik>
Raum-Zeit
Eins
Monster-Gruppe
Arbeit <Physik>
Arithmetischer Ausdruck
Datenmanagement
Mustersprache
Kontrollstruktur
Auswahlaxiom
Gerade
Kraftfahrzeugmechatroniker
Güte der Anpassung
Quellcode
Kontextbezogenes System
Bitrate
Ein-Ausgabe
Arithmetisches Mittel
Forcing
Rechter Winkel
Festspeicher
Programmierumgebung
Aggregatzustand
Lesen <Datenverarbeitung>
Nebenbedingung
Subtraktion
Computervirus
Folge <Mathematik>
Wellenpaket
Rahmenproblem
Wasserdampftafel
Gruppenoperation
Implementierung
Zellularer Automat
Zahlenbereich
Hierarchische Struktur
Physikalische Theorie
Informationsmodellierung
Migration <Informatik>
Endogene Variable
Luenberger-Beobachter
Ganze Funktion
Varianz
Widerspruchsfreiheit
Analysis
Einfach zusammenhängender Raum
Pixel
Relativitätstheorie
Physikalisches System
Binder <Informatik>
Quick-Sort
Unendlichkeit
Echtzeitsystem
Mereologie
Dreiecksfreier Graph
Wort <Informatik>
Kantenfärbung
TVD-Verfahren
Formale Sprache
Selbstrepräsentation
Formale Grammatik
t-Test
Gruppenkeim
Element <Mathematik>
Videokonferenz
Streaming <Kommunikationstechnik>
Einheit <Mathematik>
Lineare Regression
Kontrollstruktur
Regressionsanalyse
Bildauflösung
Statistik
Güte der Anpassung
Gebäude <Mathematik>
Kontextbezogenes System
Frequenz
Teilmenge
Rechter Winkel
Festspeicher
Grundsätze ordnungsmäßiger Datenverarbeitung
Folge <Mathematik>
Klassendiagramm
Gruppenoperation
Besprechung/Interview
Hierarchische Struktur
Zahlenbereich
Entwurfssprache
Physikalische Theorie
Virtuelle Maschine
Knotenmenge
Informationsmodellierung
Bereichsschätzung
Datentyp
Vererbungshierarchie
Luenberger-Beobachter
Zusammenhängender Graph
Abstand
Inhalt <Mathematik>
Datenstruktur
Analysis
Autorisierung
Videospiel
Graphische Statik
Ontologie <Wissensverarbeitung>
Schlussregel
Quick-Sort
Mereologie
Wort <Informatik>
Computerarchitektur

Metadaten

Formale Metadaten

Titel Q&A with Jeff Hawkins
Serientitel 2014 Fall NuPIC Hackathon
Anzahl der Teile 19
Autor Hawkins, Jeff
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 10.5446/18080
Herausgeber Numenta Platform for Intelligent Computing (NuPIC)
Erscheinungsjahr 2014
Sprache Englisch

Technische Metadaten

Dauer 1:11:02

Inhaltliche Metadaten

Fachgebiet Informatik
Abstract Hackathon guests chat with Jeff Hawkins about HTM theory.

Ähnliche Filme

Loading...