Numenta HTM Hackers' Hangouts, 5 Jul 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: HTM Hackers' Hangout - July 5, 2019

Description

HTM Hackers’ Hangout is a live monthly Google Hangout held for our online community. Anyone is free to join in the discussion either by connecting directly to the hangout or commenting on the YouTube video during the live stream.

More info on all these topics at Discuss at https://discourse.numenta.org/t/htm-hackers-hangout-july-5-2019/6240

A

Hello and welcome to HTM hackers hangout is July 5th. We do this pretty much regularly every every month. So thanks for joining everybody have a few people watching online. Just a reminder: if you're watching live, give the video a thumbs up. That would be helpful. It's YouTube really likes it when people thumbs up the video, even if you're watching later give it a thumbs up.

A

So this is. This format is going to change, because this Hangouts online thing is going away. Youtube just told me as soon as I hit broadcast. It's like this is no longer gonna be around so there's if I change next time, there's something called YouTube webcam. So we'll see how this goes next month, when we try and do this I just want to take a moment and say: welcome to new subscribers.

A

I know: I've got a lot more traffic and subscribers since we had that Lex Friedman interview with Jeff, so if anybody's watching that came from there. Thank you for subscribing and watching I do a lot of live streams on this channel, which you are free to ignore. If you like, or- and this is one of them- we do this once a month and it allows us to talk with community, so I always invite the community to to actually join.

A

It's got Marty, which is the teapot and Marcus from Demento, is joining so have some discussions with them. Otherwise there is a forum post, child I'll, put in chat right now. I think this is it yeah and that sort of has a summary of what we're going to talk about. So we will get right into that, so, first off and once we get through this, we'll open it up to anybody who has who was joined. It wants to talk her and I'll, be watching chat.

A

So if you have any questions in chat, I'll address those once we get past these main topics, so the the first thing I want to talk about is the Mintos research direction. So we have over the past year, you know sort of been leaning more towards machine learning and the application of what we've understood about the particularly you know: the neocortical circuit and sequence, learning of spatial pooling, let sort even the old stuff.

A

You know that came out in 2013, which is spatial temporal memory and how to apply that to current machine learning system, specifically deep networks and we're continuing to go in that direction. We're getting traction there, which is great, and so so that's where, since we're getting that traction, we're going to looks like my bit rates a little bit low. So you might you guys, might get a little shaky video. So sorry about that.

A

But anyway, we're gonna keep going in that direction and we're going to stop doing hardcore neuroscience research for the time being. I think not that we're going to give up entirely, but but we're gonna focus pretty much all of our engineering research resources to.

A

When wait sorry mark says, please repeat this: when chad is live, may be worth dropping and weekly to introduce the lincoln name, change. I'm, sorry, I, don't know what you talking about mark um anyway. We're we're moving towards the machine learning, so you're gonna see us all of our new research stuff and our live streams and everything are gonna, be focused towards the application of HTM ideas. Sorry, that happens no window cam in this legend, so you're gonna see the research meetings focused towards machine learning.

A

You know how to apply particularly sparsity to deep networks and to get the benefits of sparsity there. On that note, we also I want to talk about torch. Let me share my screen here and go to new pic torch. That sort of the this is where a lot of our emphasis is is occurring.

A

On this machine learning side, we have a a project called nupoc torch, which is basically creating machine learning architectures that are using the ideas of HTM, particularly sparsity, but as we move along we're going to continue to create new models- and you can see there's there's models in here. If you want to take a look at them right now, where you've got sparse, CN, n and then and more will probably be coming, the thing I wanted to point out on this is this recently Lewis has sorry. Now we have a firetruck.

A

This will take a moment. Cuz, it's gonna be really loud. So let me just pause.

A

It's going right by the timeline window.

A

Okay, so Lewis has recently added more examples, and the cool thing is that there are Jupiter notebooks. So we can. We can take advantage of that in this new collab environment that Google has has made available. So all of our all of these examples can be opened up in one of these collab environments and trying to make this bigger, and you can run this on a GPU which is crazy right.

A

So let me just show you if you, if you open this up and go to runtime, you can change it and add GPU and then essentially run all, and it will run this on Google's GPU, which is pretty cool, so I think I just hit run. Oh there. It goes so it just so pop through here.

A

So this is a an example of how to use this model and it will run all the training and I'll do the training and the testing and then give you like error rates and introducing random noise and all this stuff, and you can just sit here and watch it working and not. You know overload your local CPU with this stuff. You can do it all on Google's servers, which is outstanding.

A

So if you want to try any this stuff out, you can do it right here, which is a great way, I think to to try out what we're doing with with Nick torch, so stay tuned for more examples, as we sort of evolve this, and hopefully that's exciting to you guys this for me, so I'm gonna be playing around with this and hopefully I could put together some.

A

Perhaps you know video tutorials on on what what exactly is going on here and how it relates to you know more standard machine learning or deep network architectures going forward. So there's that which is nice so.

A

The next thing is, if you're, some sort of talking to our forum right now at HTM forum, we've had I've noticed a big increase of spammers on the forum. So if you notice that there's spammers on the forum, you can flag the post so there's in the options you know under each post, there's a bar with a little flag and if you think something spam, please flag. It they're getting really good at this, with like crazy good.

A

But if you're, if you keep up with you know, what's going on in machine learning and I can't tell if they're, just dirty tricks or if they're, like really smart applications of the latest NLP machine learning tactics. You know because someone, a new post will pop up and it'll, seem completely reasonable and they'll be a link to something that also seems reasonable.

A

But then, if you, if you go to the users profile, typically they're advertising a product and somehow in the profile, like the user name, resolves to a product name or company name or there's a link in the user profile that is basically pitching a product or something that typically seems to be the spam tactic right now and they're playing a long game.

A

Here too, you know so they're there they're posting without any explicit links to get you to like click bait, they're, not posting a direct click, bait I think they're posting initially to get trust in the system, because the forum software has this trust system and I think they're doing this to see real content so that in the future they can, they can do the clay and the spam stuff. So just keep an eye out for that. If something seems weird, let me know, I mean I'm, always looking, but sometimes I miss it.

A

Sometimes I miss the spammers and I almost missed one just recently. Okay, so that there's that quick, I don't want to talk too much about twitch versus YouTube, but I'm gonna be live streaming on YouTube. Now, for the foreseeable future, I think that's where the audience is so like. This, obviously, is a YouTube live stream. If you go back and look at the history, almost all my live. Streams recently have been done on Twitch and then moved over to YouTube as just a video but I've been investigating. Youtube live streaming and I.

A

Think I can do everything I did on Twitch. I can do on YouTube, so I'm going to keep so I'm going to sort of move. My process I have some technical tasks to achieve to. You know update my tooling to apply it to YouTube like the chat interactions and stuff like that, but it'll all it'll all be back and just like on twitch, so which is nice because most most workplaces allow YouTube access, but they do not allow twitch access.

A

So if you want to watch your interact at work, you'll be much more likely to do that. Okay, I hear.

B

What you're saying Marc.

A

Brown thanks for doing that, the last thing is I'm still working on building HTM systems. That's my current, like livestream that I'm heavily streaming and we're doing spatial pooling stuff, so so we're partially through it. We got to do finished up active duty cycles. We're gonna do overlap duty cycles soon, which I haven't even talked about or explained at all, so still working through that and with that hello. Lucas welcome.

A

Okay, with that we'll get right into sort of questions on the forum and respond to some of those maybe share this at least so you guys can see I think this'll work if anybody's watching here's the forum thread were was just talking about the summary so bit King was asking about a summary of capsules versus HTM. You know, and he cut caught on this thing, we're talking about like grandmother, cells and I I. Don't quite understand that I thought Marcus when we were talking about grandma capsules.

A

That was that wasn't quite I mean that we don't really want think it's working that way right all right, sorry, I'm laughing, not know I, don't know! If you hear a sign outside my window now, yeah I do they're following us anyway, throughout my train of thought.

A

So I can address a couple of those things go for it. So I guess, since you brought this since we're on the topic of grandmother capsules. I can talk about that first, so yeah, the the notion of grandmother cells, the idea that there are single neurons that represent, like you, have an era devoted to when you see a grandmother or something like that and the idea that that's happening in the brain, etc.

A

He's reacting to in capsule networks, typically how you'll train them is at the top you'll devote one capsule to each trained object, and so, for example, so grandmother capsules is like a shorthand for describing that I understand.

A

I'm missing an English word here, my brains, just not recalling its.

B

Work concern I, would yeah yeah.

A

The concern what we'll call that I understand like the young, the concern there that yeah, it seems silly the idea that you even have like a cell devoted to each concept, or that word that's like how learning occurs. Do you just the grandmother cell and have a coffee cup cell or similar with capsules? That is certainly like, that is almost definitely a hack of sorts. What is going on at the top of a capsule network as a hack but whatever, but what's going on between the bottom and the top is not a hack.

A

It's this quite nice and like this idea of yes at the top, the capsules are sort of grandmother II, but all the ones in between are in the end, between layers they're, more like, like these reusable components. They're like almost a now I mean it. It is almost analogous to the spatial pooler's many columns, where the spatial cooler learned sort of these, like course, features where each mini column is start of the course feature, and in that sense these capsules being devoted to one feature or one component, one type of component.

A

It can be elegant if it's all learned, if, like if a population of cells is capable, I would say, a population of cells contains a thousand capsules and uh and.

A

To learn a good set of features that part is nice and so yeah. What's going up on up top.

B

A

But there yeah I'm pretty sure that everyone missed, maybe the last twenty seconds of what he just said. The part that was nice, okay,.

C

A

Would you repeat that, okay, so yeah repeat- and you say you can hear me now- yeah.

A

The part that is nice is sort of like is sort of like how suppose you have a thousand capsules in a network. There's say you have a layer of capsules, a population of cells thousand capsules. It needs to learn a set of features of that set of a thousand features. Basically, that are useful for representing the input, and it's so like using the statistics of the objects to pick a good set of features. That's all really!

A

Okay, that's all like a idea, so at the top of a capsule network having this kind of hacky, one capsule per object is like kind of a placeholder to get to then take an advantage of a bunch of useful properties.

A

If you were to build a more advanced system that that, instead of using this fully supervised de proach, that's passing in something up top if it were a little more similar to like a reinforcement, learning system where the task of the system is to like output, the desk of the system is to, like is to identify objects, but it does that more as it does that more as like a behavior as more of it as an output. If, if instead the network, word like this big black recurrent box and and.

A

You you could basically remove the hack is, what is what I'm thinking here is like once you once you make the system more and more of a black box work somehow internally, it is figuring out that that's a grandmother or that that's a coffee cup and then then it acts in a way. It speaks the word grandmother or something like that that that that that can be one approach to removing the hack this. So it's like yeah grandmother capsules are like a temporary.

A

They are like a placeholder, so we figure out, like the bigger system thing.

B

A

Like figuring out, the bigger system involves action in some way motion or something probably yeah. Yeah I would say that I mean like your brain. Doesn't.

A

If you're asked to identify what's in this photo, you don't answer the question by having the person decode your brain. You answer the question by by by speaking the answer, and so like everything. So basically there doesn't have to be a grandmother capsule in your brain, because you have another way of reading it out through through speaking and so so. So.

A

My point of view is that the grandmother sells grandmother capsules, all that is a placeholder hack 2 to get to kept in a bootstrap the whole field of research, but but eventually it will go away. So that was a long answer to just one part of his to one part of that. So yeah I mean summary of capsules versus our model.

A

I would say, I mean, of course, there have been multiple hours of research meetings where I talked about this recently and one one specific part I would point to is in the most recent research meeting, which is, there was last Monday part to counsels part to I off to the right side of the white bar board to the toward the end of the research meeting. I drew like this on this spectrum of design constraints like of the you'll.

A

If you go to the video you'll, see on the right side of the board, I have these just this axis of yes-no, yes-no and and I talked about capsules versus our model and in the the the thing that is varying between. That is the underlying assumptions about wonder what neurons can do and what a small number of neurons can do um in capsules.

A

It uses a small number of neurons to represent, like the location and orientation of objects, of features of components of objects and part of a working assumption of our model is that that requires much more than 16 neurons. It required like these. These.

A

You know these ring attractors, these continuous attractors, all of the all the populations required to represent, like you know when, where something is like this orientation, where it is how far it is all that stuff that takes many more cells than then a capsule has, and so you so suppose you take a capsule, give it as many cells as is needed to represent all of that, and and now it's like now, you're really making a strange use of cells you're, making a you're like giving a capsule these thousands of cells to represent where something is, but then you then you're just giving it this sum.

A

But then it's just representing one type of feature.

A

It starts to be a strange use of of neural tissue.

A

You, you wouldn't want these capsules to be devoted to individual features. Individual objects you'd want, because there are so many cells required required for representing this other stuff to kind of avoid being wasteful. You want this capsule to represent many different types of features, and that really starts to change the architecture of the holder of the whole thing or the or the the the motivation. The the principles of the system- well, it once you have this this.

A

You know this extra-large capsule that is representing, pose and scale or whatnot, and then it's also representing, like object, identity at this point, because it's so bigger you now, rather than having it devote to one object, you want to devote it to a lot of objects. It starts to seem. Well, you could call it a cortical column. It starts to seem it's just different and and spirit, and so like at the core.

A

The connection that can the very easy connection to draw between capsules and our model is our compositional, our compositionality model, which and the frameworks paper that the section about displacement cells is about this and the idea that you know that cortex works by by composing components or by arranging component send two and two objects and the idea that, like so when you see a component- and you see another component, those kind of vote on a higher object, like all that is in common across a lot of things.

A

That's not just capsules, nuts, that's kind of a historical approach to computer vision and they both soso. Both have that core piece, though, where they have this neural circuitry devoted to doing that, and so it's really easy to draw parallels.

A

So otherwise, I tried on the board I'd tried to draw there's sort of this some continuous space of how many cells does it require to represent a location, orientation and everything. If it's very few and wanted into the spectrum you get capsules. If it's a lot, you get something more like our model, and you know maybe there are in between. Maybe maybe it's actually something in between and and I also made it clear in that meeting that.

A

A

Don't know the answer to these questions about how many neurons it takes to represent a pose and like and I'm just saying if it takes a lot our model, our model is kind of predicated on that. Then it takes a lot if it doesn't capsules, there's a lot. We don't know about neural tissue. There's a lot could be going on. Maybe cells are more powerful, more capable representing more than we think these are all open. Questions neuroscience is still there's still a lot.

A

We don't know so I'm, not I'm, not emphasizing correctness I'm, just laying out like if a then capsules, if not a then our model right.

A

Okay, thanks for the summary now I hesitate to move on to a completely other topic, because yeah I think I think we should I. Think you've said enough about the about the capsule discussion. There was a very high-level question from Barnett talking about object, object, recognition, specifically higher-level abstracted objects, and so, for example, we always talk about coffee cup like this particular coffee cup, but there's a difference between my coffee cup that I use every day and then just the general idea of a coffee cup.

A

Do we have anything to say about that at this point so to the piece he also gave me. The idea of apples versus free or like Apple, is a fruit ban as a fruit, but corn as a vegetable or you know it's really talking about abstract concepts and how real physical objects can can be sort of classified in these abstract ways.

A

Right, I, not not really I mean I.

A

We we don't have anything on this topic. This official I have little thoughts about like instances versus classes this. This idea, like you, have models of coffee cups, and then you have various instances of that of coffee cups, but I don't have any it at this point. It'd just be me, making stuff up so I.

A

Don't really add much by doing that, this is probably safe to say this is still in the realm of theory and we're still thinking about it and yeah I guess you mentioned the other like fruit and apples and such that one all is like. We don't have a very strong point of view on that right now.

A

The only thing that comes to mind is that's the type of test that that's the type of thing that is almost as easy for certain types of systems to do that are really simple it if you just have like you, know, associations, it's like you have fruit and it's associated with you. You have Apple and its associated with your representation for fruit, and you can draw these graphs so like this. Is that and and I don't know a simplistic approach to that?

A

It's just like associating representations with each other and in some ways that that class of problems is is not challenging, or rather there are approaches to it, but are simple I, don't know if those approaches are right, whereas the types of things we've been approaching are a little bit. Are things that can't be solved through just simple associations like that having to bring in these notions like location and orientation? Yeah I mean another I, don't know if this is what he's explaining.

A

But you know if you see a new type of object, a new object, you've never seen before it, and you trap your you have to think like to identify like what is it and what is its purpose? You know a lot of times that takes active thought.

A

It's just like just not something that automatically happens, but you can make associations, or at least you can say well, it looks sort of like this thing that I've you know my eye can match it towards a sense as I get when I'm looking at this type of thing or that type of thing. But it's not so simple, as just like performing a what we would call a union without everything that we know of.

A

That's all I say about that was I cutting out earlier. During my long talk, I realized my laptop had had connected to a ridiculous Wi-Fi network that was I was like I'm sitting next to a Wi-Fi, router and anyway and I connected to a different network Oh only.

B

The one time, okay, question.

A

I think, hopefully the same for everybody else: hello Falco, because respect goes popped in we're at the part of the Hangout, where we open up discussion to the community. I know of Marty and and Falco that Jim had joined in so I wanted to ask either. If you had anything, you wanted to bring up.

A

Perhaps not got a good amount of viewers right now. 14 15 people watching Lucas welcomed you. You would have any. You want to talk about what you're working on at all ah yeah.

B

Sure can you hear me fine, yes,.

A

B

Right now, the research direction we're working on is trying to combine machine learning with the HTM principles we've just called in the last year's and so we're starting from the beginning. So we have, we have a bucket placed of a lot of things. We want a taco and the first thing we're gonna solve is sparsity so right now, machine learning models are mostly dance and we believe it they should be sparse, as the brain is so right.

B

Now, that's the work, we're doing we're, trying to apply a sparse, teach several machine learning models and see if the properties we have in sparse distributed, representations like robustness and efficients, they will translate to machine learning models as well, so that's a very short key stupid and then we'll probably move on to other problems like continuous learning called encoding structural plasticity. That's something we're working on right now, and so we have like a large bucket list of things.

B

We want to apply to machine learning and all those things came from the research that has been done in the last 15 years. In momentum.

A

Cool thanks Falco has a question so open it up to you, you can it's your turn. Okay,.

D

Thanks guys, one thing that I am wondering about often during the research meetings is, when any of you guys talk about child objects. So you you have this combination of features and, and you have the you call it child objects but I'm, trying when I'm, trying to make sense of HTM I see as a as features as points the smallest amount of sensory information that any of our sensors can get.

D

So if you're, considering visual information, I consider the smallest point that the retina can detect and then it's channeled through whatever means towards our neocortex and then it starts the computation the same for our skin. When we touch something with with the tips of our fingers that there are probably a you know, hundreds maybe thousands of on one top tip of our finger and that's but okay I'll cut it short. The idea is when you guys talk about child's objects, are those already combination of inputs or is a feature just that a point sensed?

D

A

Okay, so yeah. This is a good question.

A

So when we talk about child objects, we are talking about so most of our papers so far have been really concerned with child objects. We just call them objects.

A

In that context, we call them child objects when we're once we've gotten up to the level of talking about compositionality, so the child objects here are the things talked about and columns paper, the theory of whatever that paper was called, how columns in the neocortex learn the structure of the world or whatever that and and then the follow-up paper of locations in the neocortex, the theory or sensorimotor object, recognition using cortical grid cells. Remember that title because I'm, the first author, those in those we talk about objects and and.

A

Those objects are the child objects and what they are. It comes from from moving a sensor around like it like a like code, fingertip or retina, and at each location recording what you sense at that location and.

A

So it's like really a building a map hug like at this location that sense this at this location that sense this or like this location, a sense that this location I sense that and you and you can think of in these cases you can think of the input being whatever is coming in on the sensory patch. You can think of it as like, if you wanted to think of it as just a set of zeros and ones of some kind.

A

Just like a big like a bitmap of input bits, that's one way to think about it. So we typically think of this. As the set of input bits comes in spatial fooler maps it to some sparse set of mini columns and basically that's what a quote unquote feature is, and the system a child object is what feature do I sense at this location. What sense what feature to us and said that location, so you were asking about like? Is it point by point, or is it larger than a point?

A

The literal answer is it's: it is a spatial pooling of the set of input bits and that's basically you what it is. So it's sort of like a spatial pooling over a set of points and that's basically, it.

A

So the the idea of a feature being something that takes up the idea of a feature being something that, like spans, an area or spans a volume. That's that the way we the way we I'm just reading this kind of spatial, a spatial pooling of a set of points as an object or as a feature.

A

That's a that's a feature. A.

A

We talk in terms of cortical columns. A cortical column has a receptive field instead of a set of inputs, a set of like receptors on the finger or or like retinal ganglion cells from the retina, and that's just like this big, like 2d 2d input and you put a spatial pool or over that set of input bits that output is, we call that a feature and, and so the child objects are, are what feature do I sense at this location? What feature do I sense in that location, etc?

A

And then, then, to bring in this idea of.

A

Okay, yeah, so just and just briefly, to bring in this notion of features that span an area that take up space.

A

We call those child objects and the way we have this notion of like, if you think of a feature, if you think of like the tip of this, this whole this'll tap this pen, the whole top part of the pen is being a feature we instead, we we approach the problem as like: okay, I said: here's what I sense here, here's what I sense here.

A

Let's call that a child object and and now in order to learn this higher object like the full pen, I'm gonna, I'm, gonna, rearrange these child objects like the pen top and the pen bottom into a into a compositional object, and so, when we talk about something that spans an area, that's where we use our child object. Parent object mechanism; okay, great thanks, Marcus, I'm, gonna, move on to Marty, who also had a question to ask you're ready.

C

And so my question is that since humans are now focusing on deep models and they're, only.

B

C

Hours is going to be models yeah, so how how much HTM Theory will be developed and is there a possibility for research on deep models going back to HTM or what's the pathway for that I.

A

Mean the only answer I have is that.

A

Some parts of a lot of our work is, is a lot of our work already goes into thinking about the core problems the brain has to solve and then or, for example, thinking about like what is vision. What what is vision, what must the brain be doing to solve the problem of vision and just thinking about that problem in the abstract and then it's a whole other set of work to map it on to cortical columns and to map it on to neural populations?

A

Oh so, the one one reaction I would give is that, like we're, still gonna be thinking about the core problems of vision, our core problems of you know, learning sequences, and so what it would is likely to come back to HTM, as is just new new things. We learned from thinking about the problems, will be spending I, think less time focusing on biological neurons or no. No, that's not the right way to say it.

A

I would say anything that anything new that develops is going to be the type of thing that came from thinking about the problems and not so much from from you know.

A

Looking at the embracing ourselves in the biology, so I think there is a real chance that something good will come of that. But so is it so yeah it can be to spend some time living in the analytical world of problem-solving for a while is so good. Things might come about.

B

Yeah can I can I add something I think it's. It's also an interesting way to validate the theories we have right now, so we don't think about applications of extending too deep models and extending to to the benchmarks we use in machine learning. It's also an opportunity for us to validate what we have done in html5, so right now, I'm working on sparsely, but we have a Jeremy who is working on a HTM applied to continuous learning, so I think there is.

B

There is a path between HTML machine learning and we are going to that path. It's not like a straight line, so we're going back and forth and you experimenting. We have that sdrs or experimenting with temporal memory, but it's it's been I. Think it's been fruitful to see how the theory is applied to like larger data sets to larger models and I also believe we're going to learn a lot for that, and some of that can come back and somehow help us on there on the brain part of the neuroscience research.

B

You know h10 research, it you know, that's that's my opinion.

A

Thanks Lucas I'm excited about this direction because for the longest time for years and years since we open sourced one of the big arguments, critical arguments against HTM and what we're doing was what value are you adding to current system? So what can we do right now? That's that's valuable and the answer has always been well little here and there, if you can figure out how to do it and the direction that we're going right now is basically to take everything.

A

We've we've learned and find out where we can add the most value to current systems with the least amount of effort and I think that could make a big impact. You know finally get us a little bit of attention for our NASA amount of research that we've done over the past 10 15 years and bring it to the deep learning machine learning world. So I'm excited about that I think it's going to be enlightening.

A

And anything else from from anybody before we wrap it up.

A

Ok, I think we'll wrap it up thanks everyone for joining HTM hackers, hang up July 5th, just a reminder for all you guys watching. Please give the video a thumbs up and subscribe to our YouTube channel. It's encouraging and it helps validate, but I'm doing basically doing these live streams and trying to be as transparent as possible through for you guys all the community for my bosses so cuz it lets them know that what we're doing here is valuable.

A

So please give us a like and, and a subscription and I will see you all on HTM forum. Take care.