Numenta Live Streams, 18 Mar 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Catching up on a week's worth of HTM Forum conversations

Description

Broadcasted live on Twitch -- Watch live at https://www.twitch.tv/rhyolight_

A

I am going to be using some of this time today on my weekend to catch up with a week's worth of HTM forum conversations and first I will go to the forum. It's how long that's what I'm doing so I have only recently started doing twitch stuff and then I took a week off and I was totally like offline for week. At my kid science, camp, okay.

A

On forum conversations, if you feel like it I think, maybe I should put the my your own that too again just in case. If it forgets well, let's I don't want it to be, and I wanted to get giant link every time, so I'll just put it n up so now, if anybody happens to be online because look, there's usually people online.

A

So now I'm going to look and see on all the unread topics that I have. Let's see if this works I'm still figuring out all the twitch stuff so like on. So this is.

B

A

This is the broadcaster set up, so I should see like that chat. They you see the chat pop up, but it's all white, so you can't really actually see it. So I should probably fix that, but I don't really know I'm I'm still figuring it out, so it doesn't doesn't, doesn't show up on the white background very well, but that's okay, back to the forum stuff, so I'm, a professor like twitch, telling me I'm, live-streaming. Obviously, okay, so I'm gonna go through the forum and start answering questions because I missed all these topics.

A

So there was a lot of nice. Interesting conversation, probably we'll find out so duh how much trust swarm results. This is Sam, so we've been going back and forth on this. It started with the question about the adaptive scalar encoder, because the swarm was returning an adaptive, scalar and coder in it really shouldn't be because we don't even use the adaptive scalar encoder anymore and he was asking if he should trust it and the and I told him no. But when now we're dealing with some other issues, I don't even know.

A

What's going on here, everything's, like seems to be plotted in different resolutions and I'm totally confused about. What's going on and I asked for Suba tide to answer this, but I know he's been super busy. He was gone, a cosign for like a week or more, and he just got back, and this swamped with stuff he's got a paper he's working on and I know. This is not a high priority for him. I don't even know if he saw this message honestly, but I don't know the answer. This said, why is the bucket index?

A

That means there's a function somewhere might have function like this work to convert so yeah I mean that function might work, but that there's it's definitely not the right thing to do. I, don't know what to say about this. This is a tough one.

A

So there's some bug in in Sam setup and I can't figure out what it is. I don't understand why, when he sends in a new row of data, he gets back a prediction that that is referencing the encoders bucket index as the value, the most likely value and not the actual encoded Rd coded value of that. So that's! So that's the problem, um but I. Don't know why I'm gonna have I will have to dig into this. I really don't know the I'm, not even answer it. Cuz I, don't know the answer.

A

This question so I'm just gonna leave it I hate, leaving things like this, but I just don't know the answer. The question installing new pick on a Raspberry Pi. Oh these topics, people have been installing new pick on a Raspberry Pi since, like night 2015 I think was the first time I saw somebody do it, which was pretty impressive, because it's really hard to get all of the C++ binaries compiled properly on an arm because you have to bring in so many external libraries onto it yourself- and this is totally not my forte.

A

It looks like Paul Paul's definitely got a lot of experience on the on this aspect of new pic by the way new, because then it meant a platform for intelligence computing. It's the Python, / C++ is a C++ part of it for a speed implementation of hierarchical temporal memory. So there's a lot of traffic on this form about nucleic like technical things, it's software I was able to start each team service with basic basic functionality without errors on a Raspberry Pi 3 B+ cool in total I spent not much more than two hours.

A

I also found two mistakes: okay,.

B

A

So I think this is how he's pointing out some things in Paul's, tutorial and Paul. Are you like this I, like it too good job, I'm gonna, call him a community helper. If you point out, if you find some bad documentation or something and pointed out, that's that's good juju! That's a community helper!

A

Okay! Thank you! George I! Don't think I need to respond to that and bouncing back here. Cuz like I, can't tell how many users are watching. So there's three people watching: that's nice! Anybody, if you have a question about anything, feel free to jump on chat on I will take questions, path, anomaly, detection, using grid cells and temporal memory.

A

This is Marty Marty's project Marty is doing some really cool stuff. I of this I still need to figure out how to get this running. So this is this is a he's implemented this grid cell, like encoder or so, let's see, what's that we made a video I haven't seen this yet this gonna work.

A

Okay, so any this running, because that looks cool so he's doing some type of grid grid, so module type encoding looks like you can see the actual encoding on the top. Here it says grid cells, and then this is the prediction of where it's gonna be next, so he's running simple our memories. Long as he's got a nominal score.

C

A

Something's right just didn't see me it should be doing, but.

C

A

Whole history, like a window I.

A

Don't see a like there's a.

C

C

A

So this should be a very low anomaly likely at this point, because it's just going in a circle king, so hopefully you'll move it now. We've changed the.

C

A

What Marty's gonna do is.

C

A

Don't know if you guys can see this and you guys see what I'm, what I'm, showing they click this video like it's full screen. Let's, let me make sure that you guys can see it so I'll do.

B

A

Theater mode that.

C

A

Hey dear, let me say that right here here, so you can see it full screen. Okay, you close, oh well good. Some hold on alright, so I'm guessing see that see that anomaly indication there. They like the history, he just moved and they not only went boom so so the so. The algorithm was learning this path in space 2d space and then, when he moved this over, the anomaly was like bumps up. Then we moved it back and out.

A

My goes down at school and did you see those legs in it because a little indicate those little things like that? That's probably because it it crosses every time. It crosses the path that it's been before just like. Well, that's, not a nautilus, so that is it moves it completely into a different place.

C

A

Is awesome Marty.

A

So for those of you, you're you're, watching or not in the community, so the community members are gonna know what I'm looking at sort of. But three Twitter's is that a good term Oh three people in twitch, one of the things we can do with one thing we know about grid cells and you can lookup grid cells. There's an HTM School episode about it. Just search for grid cells on YouTube.

A

The first video that comes up is me, I'm, very proud of that anyway, grid cells are really cool things that can help you map out space.

A

It's this, how it's sort of a way that neurons new brand map space or learn space through movement and once you're an adult you've already got these complete maps of space like you know how space works, because your grid cells are working and they have been working for years and years, and what Marty is doing here is he's using coding space in a two dimensional space using sort of a similar mechanism using things called good cell modules, and we can use properties of those to learn the movement of the thing over time.

A

An object through space and temporal memory, the part of HTM edge, hierarchical temporal memory, like the temporal memory. Part of that algorithm can remember that location space over time, based on how the grid cell modules update as the thing moves through space. So what he's doing is he's like sort of training a system severly training is just the system, is observing an object moving through time and and it gets used to it after a while. It's stuck because it's always pretty thing engines are, or your brain is always predicting everything. It's observing.

A

It's always pretty thing about what it's going to do next so again, like the HTML rhythm, is again always creating predictions about where what it thinks is going to happen next or what sensory input it's gonna see. Next, and in this case, when he moves the the track, you know the objects moving to it at the place. Then it's predictions all become wrong. You know this is it hasn't seen an object move over here yet so it takes a long time for it to see to see it.

A

Okay, okay I see it over here now, because that builds up over time. New synapses this that links locations in space like as it sees it over and over and over. It's like it's like dialing. It in you know, there's a new synaptic row every time it observed something and after a while it'll have learned like many different patterns. However many it might have observed and you can sort of bounce between them and that's one of the things that Marty's showing here.

C

A

He's jumping back and forth.

A

Htm' is a continuous online learning system, so every new location, all he's sending in is a location of this circle. At this point in time, and then the next step it moves a little bit and he sends in that new location and that location is encoded into a sparse representation using this grid cell mechanism, so grid cells sparks activations.

A

Just like all the neurons, the pyramidal neurons in your brain have sparse activations, meaning only a small percentage of them are actually active at any time, and so this, this location is basically converted into a sparse representation of space and that's what the memory algorithm is learning over time like this, this sparse representation of that space and what comes next and what came before it, and and because these cells are sort of actually representing neurons or the synapses between the neurons, and so as long as you have learning turned on, which is like just a parameter that you can you can you can change as you're running the algorithm?

A

It will continue to update of the synapse weights permanence weights. We call them so and continue to learn as something moves as long as you have it, so you can turn learning on and off, but but HTM is definitely a continuous learning system. It's very different from like Bayesian networks, deep learning networks where you have a huge space of spatial data and you sort of process it all at once and extract properties of it.

A

You know when you like facial recognition, software image classification, all the cool image stuff that's come out lately has all been these giant deep learning, networks, convolutional neural networks that can find features and images and and then call up all the different places it's found is seeing those features sort of in different places and put things together based on that, but it's all using probabilities and spatial techniques, spatial classification techniques. What we're seeing here is temporal prediction so through time, so an object moving in space, which is very different. It's a very different problems.

A

This is the way your brain evolved to understand the world through times the exploration of space their time. So it's a big part.

C

A

Anybody's watching us you can get to this HTM form is that discourse documented that augury just go to DeMint org and there's a link to it.

A

Really cool party.

C

A

C

100% to the process, okay,.

B

C

Immediately and.

A

This goes down.

A

So he's like speeding up time.

C

A

Remembers sort of the last things that it saw so.

C

A

Just it's similar to like memory in your brain. You don't have it's not like you. You have to dump memory in order to learn something new, but all the time when you're learning something new I mean well. The old memories fade into the background, like you're, not exercising those synapses anymore, so they become less strong.

A

The more you practice, something the more you think about something. The more you observe something the stronger the connections get between the synapses involved in that perception, or in that sensory perception and the this is the more you easily. You can recall it.

A

So it's not like you just don't old memory, but if we do some is he's showing here is that he's got a circle and we can learn it learn it learn it learn it and then it can be predicting well enough that there's very low anomalies, but then, when he moves it to it at the place and he's learning over here over here and over here when he comes back to that place, it's not as so.

A

It's like it's more anomalous than it was when he had had it there before, because you'd been deserving a bunch of other stuff, so like some of those old synapses sort of will become overridden in a way because it's had its learn, this other stuff and there's actually settings and all these algorithms to decide how quickly synapses fade away. How quickly do we forget and there's lots of ways you can go in and trim or prune different permanence Azure synapses to try and optimize, but that's a very big topic.

A

You know problems Eric come back any time, I'm gonna stream on Monday. So hopefully I'll see you there. This is just a test for me. Marty that was really cool. I really need to get this working. Maybe I can maybe I can, since it's in C++ and I, followed C++. Maybe that maybe I can do this on Twitch I can try and get your project running on another one of my work sessions on Twitch and some of the C++ people that I've found there might be able to help me out.

A

Twitch is amazing because I can help. People with coding like I, can go in and say: oh I, see what you're doing here and people do that to me too, which is really cool. It's a great way to learn a new language.

A

B

A

Don't know what to say: Marty I hope, you're watching this I think this is really cool.

A

Theo was awesome, I really need to get this project piling.

B

Are you all ready, hopefully.

B

A

This is a bad thing about twitch because, like my half-baked thoughts, if I do this, while I'm, actually updating the forums, I'm gonna have a have big thoughts. I'm like yeah, maybe I shouldn't, say that. But then it's already live I, don't know, I was gonna, say next week, but I'm, but I really don't know what this I'll just say: I'll try I'll make it a higher priority. I'll leave it at that. Okay,.

A

Okay, so can hm previous inputs I think we've already answered this question Paul stepped in and helped with this yeah. You I think you can do this with sort of a backwards classic. We've got the classifier a back looking classifier this we like to be able to remove, so you can spend your time to create and run each TM in a programmatic manner. I think this is an old topic. Isn't it.

A

I like to fix errors, automatic.

A

And I'm going to edit your question so I can fix a couple spelling issues.

A

I'll, just obviously I shouldn't go I shouldn't be going correcting other people's good misspelling. Should I thanks Marty for answering him thanks Oh like the swarming, so this was while I was gone. Maybe you have any example for making swarming with the algorithm API this? Not really so if anyone watching swarming that isn't very compatible with the algorithms API you're gonna have to build your own network, API or algorithms swarming sort of gives you this canned file with a model parameter. It's like a configuration file that works with the OPF.

A

The OPF is the online prediction framework and I'm trying to move away from it. I, don't like the OPF. You have a lot more flexibility with the network API.

A

So if you do swarm and you find mount model parameters, you can certainly use those just look at them and then go through the API Docs like here on the new pic API QuickStart like here. We go no, not okay, algorithms, it guys so so then you could go through and if you look at this I think I should maybe I'll just link you to this quick start, because this QuickStart is great because it they.

A

It shows you using the exact same model parameters file in the amel, how to use those parameters to instantiate and opf model and then how to use those parameters to instantiate a model with the raw algorithms API, and then one the same file. The same configuration file to do that with it guys so I'm going to refer you to this.

A

Okay, Mack that I think pink. You will benefit from a review of the QuickStart.

A

It is nice plus. It shows you how you can build out all three types of models: oh yeah, algorithms and network, a good guy using the same model configuration which is in a bottle.

B

A

New pic implementation and actor model.

A

Seven days ago, how much of this did I miss.

A

So I have not done any work with actors out of actor models. I know, I, know this guy D live. You were doing the pony implementation.

A

Or was it relying elixir, I thought your tank Pony find out.

A

You know HTM pulling yeah, so he's got a pony implementation and apparently it's doing something of clicksor, but that's about experiment on down neurons and voting just each model. So this is a big topic because, with like Erlang HTM should be parallelizable at some point that the bottleneck there is really hardware, but in software there's probably more efficient ways we can. We could create HTML to where they could be distributed and I think moving towards the lik serve pony or something is definitely step in the right direction.

A

Okay, so he's just asking if there's been any progress and he's rebooted the pony our line, I like this before I left, Baxley I have a prototype of a scalar coding special for when we found this for month, winding down.

A

This is going faster than I thought problem running the swarm in the hot gem example on Windows, first time user, hello, boy.

A

So this error is because the file path is long.

A

Someone's trying to run a swarm but they're trying to run it on a data file and it's not where you the program, thinks it is and I think that's a simple of aesthetics. You say the same problem, iTunes error, 9. Why would you get.

A

See I'm unable to have access to the iTunes, so the swarm on the Mac is not support. I is this: is this spam there's like really weird tricky spam? That is.

A

Itunes air 9, what both hardware and software wow I've never ever heard of this error. Tommy Shelby you ever okay, seen one day ago, close to three days ago, joined three days ago, read less than one minute: I have a feeling I'm a Shelby I.

A

Can't tell if you're a spammer I can't help your spammer okay, I'm gonna, I'm gonna, err on the side of caution, I'm, not gonna, I'm gonna, think I'm gonna assume you're a real person, although I get there. This looks like a spammers message because usually it's like a very short message and there's a one link to it to a suspect page in this page of expiry suspect so I don't know so I'm actually like.

B

A

Hey some days, evolution matters some big day error.

A

That's a I have never seen.

A

B

A

Ran that cause the error. Alright, if you're a real person, hopefully you'll, be able to answer that question ideas for HTM on FPGAs.

A

This is an older threat. Oh no, it's not too old, alright, so Marty! This is the guy who had the YouTube video earlier of the dot moving around.

A

He's going to be implementing temporal memory on an FPGA for embedded AI contest, that's cool.

A

Feeling excited for a girl or something excited the HTM core will be working, might not have access to the main DRAM due to some limitations of the embedded platform.

A

Only hope we can solve it or it or it ended up. You know, I should unfortunate ly do to.

A

Jeez I don't know this is honestly the optimization stuff I can't I, don't usually spend too much time on because there's because the theory is still so deep, like there's still so much to parse in the theory honestly Marty you're gonna feel bad, but I'm not I'm, not even gonna I mean I, know what you're doing here, but I don't think I'm gonna read through this some some of these things, I love that you guys are talking about it, but I don't want to get sucked down to optimization discussions, so I'm LaVon whoops.

A

Let me go back to and read influenced spatial cooler with connections, back-end Oh.

A

Resnick is also named mark Oda, Hall he's long-term long-term commit her on new pic and as as heavily involved in the current community. Fourth of music. The idea is to ditch separate code for SP and just use the TM oh yeah yeah. Yes, so basically he's saying this is done in the community book already, which is great I love. What you guys are doing in the community fork, not oscillations. Travelling waves, hey Gary's back I haven't seen Gary in a long time.

A

Gary did some really cool dotnet animations in this virtual world. This little icon do you. You should go check out some of the stuff this guy did in his post activity. Probably it's probably yeah I look at this stuff. It was amazing, I love the see why this is so cool anyways. If you look at there's some videos of the stuff running- and it's just really neat so he's really interested in oscillations. A lot of this stuff deals with oscillations.

A

Okay, they're talking about a paper: oh oh yeah, cuz mark brown, our bit King just posted this nature, paper and I. Don't think I've read this, don't think I've seen there was recently the I just. Why did it pop it down all the way? There was recently a something that came across my desk, but it wasn't this one so Gary says he's still confident. I don't want to go through this whole video. It's only 26 seconds, that's cool!

A

A

That's very cool, so these are like directional arrows and you can see sort of the wave propagation in the brain. This is really interesting.

A

All these circuits I mean look at these they're all circuits right, that's so cool I'm, so confident, HTM theory will work before it's a floral related learning that must occur at each place in the cortical math yeah.

A

Yeah, so both Gary and and Mark Brown I, think are on similar page here. They have some other higher-level theories and and HTM will hopefully play a role in those theories that made perhaps a local level but they're sort of putting together, like the whole, more of the whole brain. How does this region connect to this region I'm interested in in that too, but I can't focus too much about it.

A

Right mark is exactly agreeing with them saying there's my references.

A

So it's saying this is suspicions that these oscillations arise in a thalamus and are injected into the cortex and layer for that could be Casey Casey. No, this seems to be pretty well informed about neuroscience stuff. From what I've read, some oscillations are sent from higher-order thalamus to primary cortex and l23, possibly l5a and another element there's so many connections.

A

Like ours does this ii-if because he goes off on the detail tangent he puts it in this little thing. Casey knows a lot about like Beryl, cortex and rats rat whiskers. This is a great, a great way to study. Cortical columns is rat whiskers because each rat whisker seems to have a direct connection to a one cortical column and the rat cortex.

A

So you've basically got a unique separated sense if you think of each whisker as being one sense like hearing or sight or whatever, if you can think of each whisker being one sense that leads to one cortical column of cortex. It's it's a great way to sort of study the cortical column in isolation.

A

Sorry I really like talk reading about the barrel cortex, and he brings a lot of discussion about that. Rhythmic whisker stimulation.

A

There's so many so many papers, I don't have time to read. The multiple frequencies in the cortex seem to be doing different things right.

A

Well, Mark's got theories about X Creek, Formation l23, maybe working at a higher like gamma frequency to do local voting on hex, Reformation or macro file loading yeah, so these different oscillations gotta be doing different things. We they could. They could have to do with scaling. Perhaps you could have to do with I, don't know, there's a lot of things that they could be.

A

Gary, thanks for coming back to the forum, it's nice to see you around again, that's your listening bit!.

A

The papers feel so wide that they're sharing paper isn't excited by the papers. He says when as a thousand brains today, even if you can.

B

A

Its level to each year, so it's the other project where tiny asian learning structure in math.

B

Where did this come from? Oh I,.

A

Haven't gotten here yet.

B

A

And by the end of this summer, the only thing I might only need to add is a moving invisible shocks on marine environment. Those rudimentary brain into that.

A

So the thousand brands, the one thing I, think her and if you look at the 2d object, recognition problem that we're working on I keep waving away the idea of displacement cells, because that adds an extra complication and I, don't quite know how it's gonna work. Because displacement cells allow you to construct objects out of other objects and.

A

Yeah so I I want to do that at some point, but I don't know how to do it yet.

A

Mark wonders if there's a path from lower brain structures through the thalamus through the thalamus to the cortex I've noticed that there are patterns Ursula I assume there are paths from lower brain structures through the thalamus to the cortex. Oh, maybe there's not maybe because I mean all sensory input goes through the thalamus I think, except for some there's. What sense does it for some sense? It's not I can't remember it's it's an olfactory. It's something like that.

A

Oscillation phases of like Kalama, rude I, know factory bulb. That's a no sir penetration.

A

A

Is this just a time-based thing? What is a sniff cycle? Oh.

A

Okay, no Unruh a couple of new things is this: this is one day ago, 13 hours ago, spatial cooler, serialization problem too big I have stored a spatial boy as follows: open temp, write to file I cannot reload it using a third file expected total words less than or equal to. Traversal limit rewards message is too large increase the limit on the receiving end c-captain pea reader options. Ok,.

A

The above error message suggests setting the reader option. I have figured out what needs to be done, set the option call transversal limit in words, which is a default value of 8 x, 1024 x, 10 24, but I cannot figure out how to change it. I figured out that captain fee is a C++ program, but I cannot work out how to set this parameter from Python.

A

Yeah, what is the normal size of a spatial polar when stored by this method? Mine is 83, megabytes yeah I mean it depends. It depends on what you did with it. It depends on how big you made it, because the spatial polar is entirely configurable. So, if you make it a million cells, it's gonna be huge. So how big did you make it do I need to make a smaller by reducing size. My input? No, you need to make it smaller by reducing the number of columns. The number of mini columns. Excuse me.

A

Hi Kathy, P, okay, let's say I might be able to help you, let's see if I can so I know PI, cap and P. Let's see if PI captain P has a way to see it. Doesn't it's missing it.

A

Transversal I think the transversal limit in words and nesting with options are not being passed to flat array. I added the ops argument to the line below.

B

A

A

Okay, wait: maybe I just need to look up reader options, so maybe, let's, instead of Pi, Kappa treat, let's say Phi cap and completer options.

A

Sorry there we go no way of setting getting transversal limits. Let's try no way of setting there's some very large RPV and we're hitting the transversal limit.

A

It's like an order do the same thing on the pike on P side. We we just have to have we just type a reader options from the to party client.

A

So this was closed, that two party clients and servers take reader options. Okay,.

A

So I need to figure out what what happened here, what what was changed!

A

This isn't, where it really where I want to be I, need to find out where the API changed.

A

Is there a link to the spec.

A

I'm not sure, let's look at this one, any transversal limit your options. Now it just comes up in an error, but that's that's the air seeded message. Transversal limit. Wasn't that the same.

B

A

No message is too large: it's not even the same. That's not the same ad wrapping of reader options thanking the same standard. This is I'm. Sorry, maybe I went ahead and wrapped us today with ellipses knit.

A

This is my cup of tea, Oh.

A

The photos mm-hmm.

A

So he added to the helper.

A

Transversal limit and words flatter a message reader that was part of what you will.

B

Perhaps change.

B

A

In it, so whatever is emitted here now can have a transversal limit in words. So what's a called stream FD message reader, that's that looks like it's private or protected. So where is the public API flat or a message? Reader schema hacked a few men, Oh packed I, don't know. Phd is scheme a CPP, PhD I, don't know what a PhD fellas.

A

But it's it's letting you! It's now can send options into this flatter a message, the message feeder, which will allow you to put in traversal limit and words, but how does it help I mean? How does this have this exposed in I.

A

A

Okay, oh maybe I should just look read the docs reader options. Add a reader options, that's it well, what's the third.

B

All it was was a.

A

Api reference, no, it doesn't have any. Is there any easier options there? No, your options.

B

A

Let's go back to the form.

A

That is an time for wants to say, unfortunately, I'm not sure that I can't he.

B

A

Exposes anybody, who's, listening, hi captain Pisa, serialization, sort of a really vast, powerful, flexible serialization protocol in C++ that sort of works on top of protocol buffers or has something to do with protocol buffers I, don't know um anyway hi cap and he is the Python.

A

You know extension of that, the Python interface for that, so that you can save Python objects and just using this protocol, but I, don't think it even exposes reader options, see dots, see, doc, search.

B

A

So I'm gonna say: try.

A

Fewer any columns instead of 2048 dry and 24 it may perform just as well.

A

Okay, one more new thing: a steer classifier will not compute I'm trying to train an STR classifier.

A

But I can't even you.

B

A

Even first training pattern can I get sort of even the first training batter to compute it. Our message.

B

A

Findings, algorithms.

A

So if you're trying to train a classifier, does that mean that your piece Mealing this together you're using directly algorithms API I, guess that's what I need to I need to see your code I need to see your code.

B

A

Able to help, if you can share your code, are you using the algorithms API or something else, yeah.

B

A

I show the algorithms API I really should click on the Audio API QuickStart link it alright I think I have caught up I caught up with the forum, all the posts, all the latest posts. For me, okay, I think. That's like that's good, self-improving intelligence, I read this.

A

All those trying to create a little little organism, basic components, needs control, body form models of itself, its world mind satisfy needs.

A

That's interesting, yeah I, like that idea, like an evolutionary sir thing to judge and algorithm, is only concerned with how. Well, though, it's sort of like swarming yeah, it's something like that is good. Behavior.

A

Self, including intelligence figure out my hair on your mutations.

A

No, oh, this is an alt. This is an old form.

A

This is interesting, I, don't know if these messages are shown for everyone, let's find out I'm gonna open up the name: kognito window, hey Falco, nice, to see you I'm back on the forum. After a week, I'll fly in a sixth grade outdoor science camp.

A

No, okay, so that's interesting! Apparently me I get to see that we haven't seen this user in a while.

A

But could you not put down the evolutionary company if you have to evolve your agents so to speak.

A

It would be great, I mean I, like this idea. It'd be great if we come up with some general intelligence tasks that any intelligent agent could be sort of tested with and and evolved using something like this I like that idea.

A

And this is you Falco.

B

Next, a baby system I.

A

I'm not sure what exactly he was talking about the same thing in there.

A

Yeah yeah, the application would be the dog. The dog show would be like the evolutionary, the the the environment, the applications revolving with them internal states, wouldn't matter. Okay, good well clear that all's good at clearing things up. This is also me just like catching up with the conversation too.

A

Teachers being judged feel to place value on their own internal models based on how well they're meeting their needs. Yes, this of course requires them to know their own internal state. This internal evaluation should self assemble without requiring a third party to view the creatures after the rounds and interpret them. I agree. The judge is a third party which only needs to generate composite intelligence scores based on the results of various tasks. Purely objective function was an acquired knowledge of internal States, yeah, okay, yep good good.

A

The answer I could give Paul a good answer for every one of this post, but I won't I'll. Just give him I'll just give I'll judge edge of it. Judge everyone with a different ruler, Jimmy.

B

A

Yeah yeah labeling happens because I'm pulling out really.

A

Generated prediction: it's a point.

A

You need to coax me a long conversation yeah right.

A

Yes, like a baby starting random gradually becoming purposeful. This is how AGI will be created like a baby, starting with random movements and gradually becoming purposeful I. Believe that wholeheartedly we are going to create a structure for intelligence to emerge. It will explore space through initially random movements as its building up a model of space and a model of itself and the its interactions with space, and at some point it will have learned a model well enough that it can navigate now. I do not know about goals, rewards punishments and I, don't even I.

A

Don't think too much about that. Yet I'm still only thinking about the object, modeling portion of it with the you know, and how motion is required for that and how space is modeled through motion.

A

So much yeah, there's so many things to talk about in this bar. Oh it's hard to keep up with.

A

Everybody's interested in this ideas, because it's good idea, Paul.

A

Sorry about the background noise, my kids are upstairs bouncing off the walls. I, apparently, okay, go distributed. Good, so I have to like compare this to swarming and apply lessons learned some swarming. It doesn't I, don't think it should matter what language it's in. It should be language of agnostic, intelligence, agnostic. It should just be a way to judge a system, an agent or something to perform some general intelligence tasks.

A

Distributed would be awesome.

A

Her face provides a tag, new challenges.

A

Its communication from sensory input to the creature and from the creature to motor commands with Hall Pass, this interface yeah- that's that makes sense so you're, basically creating an environment, an interface for an environment that will be a test test bed and it will decide how well think different agents in that environment cop accomplish whatever goals are set before long.

B

A

New challenges could be added, yeah I, like this idea, I like I, said yeah, I, think I still think and I think Jeff takes us to that. Robotics is going to be probably one of the first major applications of HTM and that's all gonna start in simulated environments like we're all talking about right now. This could be a toy simulated environment of, however many dimensions you decide it will be I'm working on two-dimensional environments.

A

The thing about working in two dimensions is I'm, pretty sure that scaling to higher dimensions is not going to be a huge deal now. Orientation is gonna, be a big deal but I'm, not sure, but I, don't think adding a third I think. Once you go to dimension, that I mean that's many you're in many dimensions. Now you just add another add another dimension to the arrays. You know what I mean.

A

The interface is critical opponent.

A

There's a lot of obviously these put a lot of thought into this frames of mind. A theory of multiple intelligences Howard Gardner interesting I have not read that I'm gonna bookmark that and start it thanks. Paul, that's good stuff.

A

It seems so, let's read about cell death, okay cell death, I didn't read this before I left I, don't understand much about what is cell death in the brain and how the spatial Pooler can implement a cell death. What we currently don't as I understand, if all connections intercept a field of a column are connected, it means that this column is dead. No, no! No! That's not what that means.

A

If all connections intercept a field of a and we're talking mini columns here, except a field of a mini column, if they're all connected this means that the column is death, I, don't think so. Jimmy re answered this death of neurons due to Parkinson's Alzheimer's death by over connection I see some simulation about cell death. I, don't understand this concept! Well, what is alright? We don't have a simulation of cell death, oh yeah, so I know what you're talking about so one of the spatial Pooler papers.

A

Okay, right, one of the spatial one of the one of spatial cooler papers showed.

A

Tolerance to simulated cell death I think this is what where die is referring to. We did not implement. You know how I say this: it is not a part of part of the SP. Let's call this fake cool cooling process. It is just something the that SP handles. Well, yeah cells die all the time. Sometimes it's random switch, pull or simulate stem cells. All good models of the brain should be able to survive. Yes, this is actually a D Mac.

A

There's some discussion of the general SP operation and the part of the paper regarding failures on this thread: okay, okay, so it's relative, related yeah! That's definitely not true.

A

A

Okay, that's well with that! Hey! What's going on here to it I'm still alive! That's good! Three viewers, hello viewers and I'm. Does anybody have any questions or want to want to guide me in some direction? I'm, just I'm, just going through more form messages that I haven't read. This is Mac again, I. Think I already answered one of his questions. Gosh I wish nupoc works better on Windows, so nupoc and captain p, kappa p, does not work on windows. New pic with captain P does not work on windows.

A

Just gonna have to gather face can't be just does not work very well under windows. That's right!

A

Thank you! David I'm, gonna, give you a badge for that. Dancer exercise should read the answer. First, oh he's suggesting a try. The community version, all right, so you're gonna be a community helper. Thank you. David.

A

I'm, so glad that this is happening, I, don't if, once we get new pic 2.0. If and when that happens, we'll not have a PP.

A

Your questions, so that's the intention yeah there might be I, can't tell because I'm not keeping track of the new pic, though the community code, but hopefully it'll, be real folks.

A

Community version, they're warnings, that's been strange, which looks like.

A

This, yes, thank you. There is no way that windows can load the Python extension libraries built for Python 2:7. This is a problem with microsoft, visual visual, whatever compiler the c compiler that abilities old life online.

A

Captain p is a captain proto, that's sort for captain proto and it's because it's built on top of protocol buffers and it's and it's a way to serialize any object in any language any time. So you can send it over the wire and something else on the other side can deserialize it and use it and it's very flexible. But it's this huge pain in, but it has been so hard to deal with. I don't want to use it anymore and we're gonna pull it out.

A

We were yeah, I don't think we ever should have put it in, but that's another story.

A

Okay, I am going to move this at this point. I'm going to split this topic. I think I'm gonna split this topic.

A

Because this should be in the community fork, I think let's split it right here. Let's select this one and below and we'll split this off in the community fork move to new topic.

A

They say running.

A

Fourth, on Windows.

A

This isn't a good enough, no yeah for you can. You can use the serialized version to serialize things that captain P creates in any way you want. I mean you. Can we just write them discs and write them on right on the disk, so you can put them in a stream. You can use web services, you can you know whatever you want?

A

This is gonna. Go on the fork.

A

I'm not gonna fit serialization I, don't know what up shell tag, but I'm gonna move that, okay, what other topics running community for Chrome window? That's what I just created regarding multi-step prediction and anomaly: detection, hi everyone, when you guys say the model, will make multi-step predictions. What does it actually mean? Well, the model make predictions next, second or next minutes. If I'm doing a time series and an analysis does the next step prediction depend on the way.

A

I input, my data, okay, so Paul's explaining us, the multi-step prediction is a trick of the classifier new pick or excuse me. Htm' makes prediction, so your brain is constantly making predictions, but it's based on sort of the time frame or the time context of time in which you observe the world, so you're making predictions about what's coming next immediately? Next, you can also think ahead and use your imagination to imagine what you're going to be doing tomorrow and predict what's going to happen tomorrow.

A

That's not the type of prediction that we're talking about right now. At this low level, the the low level type of temporal predictions that are be made that are being made in response to sensory input is, is basically what am I going to feel as I move through space. Those are the types of predictions I'm talking about. What are we going to observe as we're moving through space?

A

That's what HTM is doing, so your brain is not necessarily predicting what am I going to predict one time step next vert or a hundred times steps, and that's when I'm over here you know: that's that's not what we're really those aren't the types of predictions that we're doing when we do a multi-step prediction, it's really something that the classifier is performing. It's saying, okay, based on this, knew this cell state and everything you tell the classifier you want to. You want to learn one step ahead.

A

It'll keep track of all the all the last predictions and how well it did, and it will. It will go back and decode that into the language of the encoding space and tell you the probabilities of what's coming next and when you do a multi-step. It basically does that exact same thing, except it doesn't look one step back. It looks X steps back. So it's not a trick of necessarily of HTM. That's giving you the multi step prediction ability it's how the classifier set up.

A

So so yes, Falco, says: if you get an input, every ten ticks you'll get a prediction and ten tick in Turville. So exactly, but if you get an input. Well, that's as long as you're, not inputting, you know empty empty information, tick, tick, tick, tick and every tenth. You get you get something. It's.

A

If you decide to predict one tick ahead, for example, that's different from 10 ticks ahead. If you get, if you just the hard thing to realize, is your brain is not time, stamping every moment that comes into it. It experiences time relatively based on a lot of internal conditions. Chemical imbalances in your brain or whatever you know when you have a lot of adrenaline running time, seems to slow down.

A

So time is a bit warped in your brain time doesn't mean as much it's not as solid and stable in your brain, as it is I think in reality or how we perceive it in reality, oh yeah. If so, if you have a regular like spike, every X ticks, and otherwise no Z stream, yeah yeah. So if you get a spike every time, you'll it'll it'll predict that if you're one step away from that spike, it'll predict it. If you're two steps away from that spike, it won't predict it because that's not coming next.

A

It's not coming next. However, if you have a two step prediction, a multi step prediction, you say, don'ts predict what's coming next predict what you think is coming two steps: it's just you're telling the classifier to keep track of the last two steps and to do a reverse classification of it. It's it's not as I think Paul's probably explains just fine.

A

My stream produces the same timestamp for two different value for two seconds. My my stream produces the same timestamp for two different values for two set doesn't make any sense. That was a key sense.

A

Yeah yeah sample data give us the sample data. What do you mean? How do you set that prediction request the multi step? It's a parameter in the model params for temporal memory, I think it's steps ahead. It's something like prediction: steps ahead! It's something like that! I, don't know what that is. Something's notifying me. Somebody else is live on. Twitch I've been like perusing twitch lately, so I like signed up to watch a whole bunch of people's streams.

A

Okay, stupid specific questions about coding, hm I think, oh great, alright, someone's trying to write an HTM and C++ cool of you. Are you new host esos ISOs.

A

And spin around a little bit join March 1 not to too much okay I just like to see how much people have been reading. How does the brain do it? How do you set that prediction request? I, don't know how the brain does it? It's a good question. I, don't know, but when you start talking about predictions in five minutes, it's not the same type of prediction. This is it's not a sensorimotor prediction: you're imagining you're like bringing in the whole. It's not just your immediate state of things.

A

It's the state of things outside your immediate sensory input like when you, when you think about what's gonna happen in five minutes. You have to think about things that you're not immediately observed. You know what I mean it's not the same type of prediction: I, don't know, and I really don't know how it works. I mean we are AGI is not right around the corner. Okay, so about the pseudo-code.

A

Segments for column right give it a set up. You know I hope forever. His answer is given a set of segments return. The segment Center breath is working on his own implementation. Rebbe used to work for the mint that he did all of our web stuff for a long time. It's great engineering, it looks like he is letting them know right when it says. On a cell, like the synapse went from the cell, it's burning stem cells, burrow snaps,.

A

Neck sound out, dendrites dendrite segments upon so synapses in.

A

Not straight get it quite understand that graphic in a set of segments. Okay, so we're talking about distal dendritic segments segments on a mini column that this function should. If I were writing this function, I would call it distal segments for mini calls.

A

Given a set of segments return the segments that are on a cell in this column when it says on a cell yeah, it's coming from the cell, so so the tricky thing here is segments. We don't in a real neuron.

A

It's like dendritic segments fork and fork and fork and fork and fork and I, create this great big tree so outside of procs the proximal sort of bubble around the soma of a cell, all the distal segments we're treating as just like a list of segments where we're not treating them together in any way, but we're saying that each one of these segments could potentially create a dendritic spike, an NMDA spike that will cause the soma to become depolarized and go into a predictor state.

A

All of these distal segments we're just putting into a big array we're not saying. Oh some are close and some are further away and some over here and some over there, it's all like sort of random and they're, just in a big array, so on the cell just means that somehow attached to the cell, it's not coming and and we're simulating that the mini columns are basically we're sort of simulating that they're. Just one neuron. You know we're saying we're they're all going to share the same input, but that's for proximal input, not distal. So.

A

Yeah, so the distal stuff will spread each segments for column C that may be proximal segments for column. Now that I think about it. I don't have to take a look at the code. Let me just see where this conversation before I go, spend a bunch of time on it.

A

It helps a lot okay, okay, I'm, just gonna, see if this guy gets his it is making progress or not.

A

Come from a hierarchy.

A

Okay, I, don't know if this is a newbie conversation. I think I might move this getting quite theoretical here, but I'm gonna, let it go until the Oh Pete comes back with some more detailed question because I'm not sure where exactly it's going with this but I'm. Not. um This is a newbie question. This is gonna, be in hackers.

A

Alright, what's going on here, so a couple viewers thanks for watching I, know: Falco is probably still there. Who else is online who's, love to hear you pipe up, and let me know what your what you're working on you're coming from twitch or from someone else from the forum. Oh man, this is a deep detailed.

A

This is really detailed. I love that Casey does this, but I honestly cannot keep up with it.

A

I'm, just not good enough of the neuroscience, so if you want this is so packed with info like I said: Casey works or has a lot of knowledge about barrel, cortex and rats, and so a lot of this I think is informed by that. But it's really a great place to study cortical columns so I. Imagine all of this stuff was really useful, um but wow. That's a lot notes. He's been working on this for a while. It gets links to sources.

A

Yeah, so the functional properties of Baral, cortex neurons and then look at all these references yeah in case you're, really on point here, yeah yeah mark, says I see I have several evenings of review in the near future. Absolutely this is insane this.

A

Is I have to I, know I've given Casey, so many of these awesome badges but I'm gonna keep doing it. This is when you can put something like this together. I haven't even read this, but I can tell that he is put sauce together in a really interesting in a way that is trying to communicate like it's it's this type of stuff.

A

That's that can help you with with a further theory like if you can, if someone summarizes a bunch of papers and says this is connected to that, and this leads that listen to that it's it's super helpful and and I don't usually I'll. Let the research guys read this as they want to cuz. They I think most of them see the neuroscience stuff, but most of them, whenever I bring stuff up, they usually are like yeah I know they they prove. They have a pretty good idea of this. The connections it is yeah.

A

It's super daunting and, like I said you saw me skim it I, didn't you can do a quick, read and and look at all these papers. Geez Casey it's crazy, but the REIT I mean this is a huge, really awesome resource.

A

Yeah another functional imaging of visual cortical layers and sub plate and awake mice with optimized 3 photon microscopy.

A

So sub plate remnant cells and probably modulate higher-order thalamus from primary cortex higher-order thalamus. Oh.

A

It's like a hierarchy.

B

A

Getting really good at monitoring there on some thinkers, we're gonna discover a lot about the hierarchy. Soon enough, oh I missed the depth issues. We thought it was feedback down the Purple Heart l6 cortical thalamic projections from vortex to thalamus coming out of l6 or usually thought of as feedback down the cortical hierarchy. Size.

A

A

Good brain I'm gonna skip this one. Oh.

B

A

The UM well, if you look at the thalamus and how it projects to cortex, it's like a little like if you look at the map of the areas of thalamus and and what areas of neocortex it projects to it's like a mini neocortex like the visual areas here and the cortex and middle areas here, and they they have like I've got up here.

A

This is my I drew I colored in all this stuff. There's a big section on earth, I'll miss again, but I was trying to study here. Yeah, that's the thalamus, and and it's got all these projections.

A

Where is it here? We go so here's I, don't know if you can see this or not check this out. I think I can.

A

There it is the thalamus and just ejects up right. Here's that here's the right picture, but but this thalamus is all broken up into pieces here and and each piece projects to a different part of neocortex here and and there they matched like this one goes to the same side and in the same location. So the little football is like a sort of a mini version of half of your neocortex and and the projections project right into this same area of the neocortex down up and down.

A

So you know, we think that they all the all the thalamus researchers like to call the thalamus a blackboard they call it. You know they use the metaphor blackboard, but you have to think of it. That's like a really really fast blackboard. It's like I think it's like projecting what you think is gonna happen and then overlaying what actually happens and then erasing it and like doing extremely quickly in a way that you can compare immediate sensory input with with predicted sensory input and respond.

A

Somehow I, don't know how it works, but they call it a blackboard. It's not something that I. Don't think I'm not sure that it has a huge learning capacity. That's other parts of your brain are learning, but it is like the place where everything where reality sort of plays out it's the first place. Sensory input, hits and I.

A

Think that's where you perceive reality is in the thalamus, but, like I said I'm, not even actually neuroscientist as you're a classifier and JavaScript okay, Jimmy looks like Paul's already taking this one I have not written an SDR classifier in JavaScript either so.

A

What's the biggest concern with maybe I honestly, don't even know how registers really even work, but if it's, if it's the difference between RAM and disk, potentially that might be some way to do. I, don't even think Ram makes sense, because you can. You can keep a ton of memory and RAM and leave it there for a long time and and I, don't think that's happening. The thalamus I think it's like an instant sort of like the now like what you perceive as now.

A

The present is happening in the thalamus and I think that it's constantly rewriting it's constantly comparing since we input two two predictions of there.

A

With point neural network, so you get an input mapping to some output. There is no good, one-shot learning and the system generally has to learn offline right at each time. You get input, mapping, nothing changed or something happened.

A

Good input, mapping so I think that the hex grid coding is such a big deal with hex creed coding. You get infant mapping to some stable hex grid pattern and all the other advantages about HTM offers for fast online training at C concern.

A

When you with HTM, you get input, mapping, nothing changed or or something happened. I.

A

Don't understand that, do you feel the same way if you once you think about incorporating motion and it had the role HTM plays and in sensorimotor did.

B

A

Put my face back yeah I'm stuck, um let's see, ok, thank you, Paul for splitting that topic.

A

Okay, good job I.

A

Think I'll ask this: what do you mean by this.

A

What do you mean.

A

What do you mean? What do I, what do white when a team gets an input and it and it creates a prediction of what it thinks is gonna happen, so he but he's trying to I'm trying to talk about neural networks, neural networks, you get an input, some outputs, like you, get some input and it's like classifies it to some output, but with HTM you get input and yeah. There is no classification, I don't get it. It's not nothing happened or something changed, I guess I, just don't staying with this way.

A

What do you mean by this.

A

Yeah that I mean classification. It seems like neural networks, mostly what they do is classify now. There's generative neural networks that do more than that, but you know all of I.

A

Don't know enough about generative I, don't know enough about Ganz to talk about them. All so I probably shouldn't say anything, that's something that I need to look into, but but all the examples I see all the like powerful things, especially pattern. Recognition is a spatial pattern. Recognition seems to be it's some type of classification and and and sometimes it's just a classification of what feature exists in the set. You know.

A

You can sort of think of no I'm talking out of my yeah.

A

Okay, crashing on serialisation SP model to file spatial color, proto error.

A

This is Mac, I, think you're on Windows and that's the problem. I think this is the same person. Who's yeah running.

B

A

The problem is that serials, a captain P realization when serialization does not work on.

B

A

Okay, nab codependent metrics from a single system.

A

Alright, does any of the nab datasets provide different metrics from the same physical system. That would therefore provide different views into the same behavior them know, for example. Well, there may be some of the data sets might internally, but.

B

Yeah I, don't think so.

A

I, don't think so well, I'm, just gonna answer that question.

A

I, don't think so.

A

Non-Redundant contributions of Mars three level of analysis for explaining information processing mechanisms, I've read a little bit of Marr favorite hobby horse.

A

It's another paper.

A

You know my therapist.

A

Ignore swarming, because I have two issues here, once I swarm over and set for the a the next time, I wanted, the prediction of B Y SH Y should not it went up. I should order a beat.

A

It makes me hurt, takes much time, which sometimes makes me hurt yeah it's awful.

B

A

Do you want your application to do if you just once what kind of input data do you have available.

A

If you need to swarm or if you can use standard starting model grams and.

A

What's the word commute and.

B

A

B

A

Why there should be like I thought, I had you know, I'm just talking to myself question about sparsity and local inhibition. So a lot of messages on this form while is gone, which is a good thing. I have I, also have problems sparsity and local inhibition, because I've read the code and oh I, think I saw this who flag this for moderation,.

B

A

Okay, that was just to make it in some thread: good job.

A

You can't strictly control the global spar series near that you need an alternative approach such as this very good. Yes, we talked about this I I'm, pretty sure. Okay, thank you. Community is monitoring itself max started a lot of topics while I'm gone.

A

How can I know that the performance of SPO is improved during the learning process.

B

A

You send and similar input to the DSP get good news. Similar output, the more different the input is, the more different the output start with that.

A

Why is this still not marked as read Oh.

B

Oh, my gosh I think I made it through them all I made it through them all: okay cool. How long did that take me that.

A

Took me I've no idea Oh. What's my stream uptime.

A

One hour and 35 minutes all right, so it took me 1 hour and 35 minutes to get through all of the HTM forum topics or all of our discussions that happened. While I was gone at 6th grade science camp there we are okay, I am going to log out thanks for watching if you're watching tune in on Monday I'm gonna have the live stream Monday at 1 p.m. check. My channel it'll be about I'm, going to chat so hopefully interactive about artificial intelligence.

A

Just in general, I'm gonna talk about weak AI, strong, AI, deep learning, a classic AI, biologically inspired a eye on the bunch of stuff and I'd be more than happy to take questions. I hope people join in and see me then. So I will be back in my main office, Monday 1 p.m. Pacific Time, that's UTC, minus 8, so hope to see you there thanks for joining. Thank you file, Co, for always being there see on the forums.