This is amazing. I play drums and while playing I feel like I am moving through space, moving through something, slightly upward and to the right. Although I am sitting still. I wonder if other musicians feel like they are moving. Anyway, this is a really amazing video and I am glad that they are giving the brain its due.
Bravo! One of the best HTM School videos yet. Follow-up questions: 1) What evidence exists that our spatial concepts underlie our temporal concepts? 2) What evidence exists that our egocentric concepts underlie our allocentric concepts?
One other follow-up question: Do you think that it will turn out to be the case that the entities which were depicted as individual grid cells in this video will turn out to be layers (L6b) of individual cortical mini-columns?
Hi! I'd cite a paper by E. Moser (mentioned in the video) to provide arguments on this question: www.researchgate.net/publication/235376121_Memory_navigation_and_theta_rhythm_in_the_hippocampal-entorhinal_system
1) temporal without spatial is nearly meaningless. It's like a 1D array. You MUST process more than that to understand reality. Adding "spatial" to the mix means each data point in the 1D array can contain a wealth of information. It's just how reality is, and it makes sense to me that's how the brain represents it. :) 2) I'm not sure, this might not have experimental evidence. This is theory you know. ;) Good thing is that it can be tested.
Thanks, Matt, this is the most intriguing lecture I have ever had on TH-cam, I happened to come up with similar ideas like the SDRs myself earlier in graduate school and when I meet the HTM theory, immediately I found this is the 'home'. Thanks again for those wonderful episodes!!
That animation did a fantastic job of explaining the concept. While this makes huge steps in terms of understanding the brain, it also has some really important implications in computer science and mathematics. The qualities are truly astounding: - tolerant to inaccurate input - probablistically provide location - accurate measure of change (velocity in particular directions) - accuracy benefiting from high dimensionality - similar features being nearby in space - allowing for feature overlap (does not require probabilisticly independent inputs) From a mathematical view, I think it was unclear whether or not that kind of mapping was even possible with finite discrete input. Cryptography and Search engines could likely benefit from this kind of uniqueness mapping of various different features.
Hey Matt and team. Just binge watched all 15 episodes in two days and appreciated how both the production quality and your hairstyle evolved for the better over the years. Great job well-done. The visualization helps tremendously.
Great video! I have been following both Numenta and the Moser's work for years and I am thrilled to see Numenta incorporate the Nobel price winning work into their models. Exciting times indeed!
Another thing I did not mention was anchoring. You can change where the projections intersect with other projections. I hard coded them all to intersect at 0,0 but it could be anything.
Hi Winfried, thanks for your comment, which I know Matt would've appreciated. Sadly, Matt passed away unexpectedly a few months ago. We miss him terribly. numenta.com/company/newsletter/2020/05/12/matt-taylor
From 12:35 the best part of the video: we see concepts in a spatial perspective! It's just remember that when are learning mathematics in school we make simple operations using visual objects like fruits or hand fingers to understand the results. Furthermore, this new theory explains why is so dificult to us imagine some things like the size of the universe: so many spatial references to draw a big frame of it! Physics, for instance, is all about imagination: you must imagine objects or particles interacting between themselves to understand concepts. This only is possible if your brain manage positions of objects as well their features (compositionality). The new theory (1000s of brains) will clear many lacunes of machine learning not filled by the current approachs.
@HTM School This is extremely interesting. Do we know what the mouse's knowledge of position (the grid cells firing) is based on? Like distance walked, sense of inertia, or the earths magnetic poles? So many questions... (some that probably matter some that don't, and no I don't expect an answer from you for these) does the position of the grid dots change with different iterations of the experiment? if so does it depend on where the mouse is placed in the room and in what orientation? if the grid does not vary when taking the mouse out, would two separate rooms placed side by side, that are not axis aligned, have a grid that mesh with each other or would the grid orient to the walls of the room? what happens when objects are added to the room? does the mouse grid ever drift, shift, rotate or recalculate over time? does grid cell fire for a sleeping mouse that is moved by someone else? do mice use an A* algorithm on a 3d hexgrid? (I could go on but I won't) Really cool stuff (If you know of where to find more information on these experiments a link would be great.)
Forgive me but I want to take a look at two separate scenarios: A) A mouse sits and it fires at any time upon sitting or resting. B) A mouse moves thoroughly (without rest or stop) and it fires at any time. What's the ratio for non-motion to motion? With open interpretation (combining three possibilities with each respective ratio): High non-motion to Low motion -- It needs to sit in its familiar place that is already registered -- so it doesn't need to learn its own place and has the time to think something else entirely - producing new information, accessing to non-related-memory-to-current-event-or-that-it-keeps-adding-attribution-to-current-event. Low non-motion to High motion -- When it's in the "comfort zone", it's already recorded all the previous-knowledge and there is nothing new to fire. But when it's moving around, all small attribution, details, motion, the gravity that throw the mouse off -- it constantly learning new information and if by some chance, some associative memory is familiarized or triggered, and it keeps adding new sequence whenever it trip, small something new, a change of shade...It doesn't consent to his usual rigid routine -- The nature or its reality keeps 'teaching' to the user of his existence.
So presumably one's "map" of a musical work is represented by an analog of grid cells. And perhaps the "internal clock" is related to the entrorhinal-hippocampal system too?
amazing video as always, but I get one question : I see how this system work to represent location when you have already your position in the world, you can translate that position to the gird cell. but how does the cell do to know where it is in the world, and so fire. I can see that work by having a internal position in the brain that is update every time you move for example, but how does the mouse does to know where is position is when we put in in the box at first ?
The whole grid cell thing doesn't explain how a grid neuron knows when it's time to fire. In the case of sight for instance if a neuron is sensitive to vertical lines we can easily see how that neuron knows when to fire: if neurons arranged in a straight vertical line on the retina fire, then that particular neuron will know a vertical line has been detected. In the case of space, how does a neuron know that I have moved to the location it's supposed to represent? There must be a whole lot of processing of movement and vision feeding into the grid cells, much like there is a lot of processing done in the retina, before other neurons can know anything about the world.
Excellent video. The 3D grid seems to naturally be tetrahedral rather than strictly orthogonal: A packing of spheres. This extends to hyperspaces as represented in Pascal's Triangle. Is there any way to test this experimentally? Has it been so tested? See Close packing of equal spheres: en.wikipedia.org/wiki/Close-packing_of_equal_spheres
Information is expressed spacially by default for a few reasons. Reuse of course. But on deeper level concepts have a relational space. Or we think they do. Chicken or egg?
This video deserves a lot more views.
You can help by sharing it with your friends over social media or email. :)
the views will grow exponentially for sure
The whole series deserves a lot more views! It's cool to be early. 😎
Someone just posted this on our piazza and I couldn't stop laughing. Can't wait to watch all the other videos!!
What an incredible explanation of such complex ideas -- well done and thank you!
This is amazing. I play drums and while playing I feel like I am moving through space, moving through something, slightly upward and to the right. Although I am sitting still. I wonder if other musicians feel like they are moving. Anyway, this is a really amazing video and I am glad that they are giving the brain its due.
Bravo! One of the best HTM School videos yet.
Follow-up questions:
1) What evidence exists that our spatial concepts underlie our temporal concepts?
2) What evidence exists that our egocentric concepts underlie our allocentric concepts?
One other follow-up question:
Do you think that it will turn out to be the case that the entities which were depicted as individual grid cells in this video will turn out to be layers (L6b) of individual cortical mini-columns?
Hi! I'd cite a paper by E. Moser (mentioned in the video) to provide arguments on this question: www.researchgate.net/publication/235376121_Memory_navigation_and_theta_rhythm_in_the_hippocampal-entorhinal_system
@HTM School perhaps it's a good addition to your list of papers?
I don't think so (that's me Matt speaking). Jeff might think otherwise? I don't know.
1) temporal without spatial is nearly meaningless. It's like a 1D array. You MUST process more than that to understand reality. Adding "spatial" to the mix means each data point in the 1D array can contain a wealth of information. It's just how reality is, and it makes sense to me that's how the brain represents it. :)
2) I'm not sure, this might not have experimental evidence. This is theory you know. ;) Good thing is that it can be tested.
this gives me chills.. no doubt that this'll contribute to the future of HTM a lot!
Thanks, Matt, this is the most intriguing lecture I have ever had on TH-cam, I happened to come up with similar ideas like the SDRs myself earlier in graduate school and when I meet the HTM theory, immediately I found this is the 'home'. Thanks again for those wonderful episodes!!
did you follow up on the SDRs representing concepts ? Will you?
Im new to programming and an old fan of Jeff and i will promote this channel as much as humanly possible .
One of the principal reasons I'm getting into python and programming is HTM theory and Numenta
Triangular lattice, triangular logo, triangulation in Euclidean space; all made possible by hexagons. LMAO!
That animation did a fantastic job of explaining the concept. While this makes huge steps in terms of understanding the brain, it also has some really important implications in computer science and mathematics.
The qualities are truly astounding:
- tolerant to inaccurate input
- probablistically provide location
- accurate measure of change (velocity in particular directions)
- accuracy benefiting from high dimensionality
- similar features being nearby in space
- allowing for feature overlap (does not require probabilisticly independent inputs)
From a mathematical view, I think it was unclear whether or not that kind of mapping was even possible with finite discrete input.
Cryptography and Search engines could likely benefit from this kind of uniqueness mapping of various different features.
Thanks so much! :D
Hey Matt and team. Just binge watched all 15 episodes in two days and appreciated how both the production quality and your hairstyle evolved for the better over the years. Great job well-done. The visualization helps tremendously.
Great video! I have been following both Numenta and the Moser's work for years and I am thrilled to see Numenta incorporate the Nobel price winning work into their models. Exciting times indeed!
Thank you for this incredible video. Made all the light bulbs light up!
This one was really great. It helped bring some of the concepts together. Please keep going.
I really like the visualization. I hadn't been thinking of rotating the grids before.
Another thing I did not mention was anchoring. You can change where the projections intersect with other projections. I hard coded them all to intersect at 0,0 but it could be anything.
This is where I really learned something new. About the brain and nn both. Awesome....
That was quick. The half life of comments.
you guys never cease to inspire :D I can't wait to hear all about the tricks of grid cells xD have a good one, and thanks again
Perfect explication. Very clear and dynamic. Thanks :)
Superb, just superb... This is gonna be big someday... Jeff, Matt, all of Numenta keep it up! U may change the world someday!
Great stuff, you guys have got it. I am convinced that from your work will come the machine intelligence that I dream of.
Wow this is awesome.
Watched the entire series in two days, excited for what's to come!
This is some awe-inspiring stuff. I'm ready to dive into the open-source. Can't wait for the next video!
Matt, this was wonderful!
Hi Winfried, thanks for your comment, which I know Matt would've appreciated. Sadly, Matt passed away unexpectedly a few months ago. We miss him terribly.
numenta.com/company/newsletter/2020/05/12/matt-taylor
Wow, amazing stuff! I'll definitely watch this video again.
After 14 of these dad jokes openings, I kinda learned to appreciate it.
awesome video, Matt! Great work
this made my day
I really enjoyed the excellent grid animations. Is that available anywhere to play around with?
simply, amazing
From 12:35 the best part of the video: we see concepts in a spatial perspective! It's just remember that when are learning mathematics in school we make simple operations using visual objects like fruits or hand fingers to understand the results.
Furthermore, this new theory explains why is so dificult to us imagine some things like the size of the universe: so many spatial references to draw a big frame of it!
Physics, for instance, is all about imagination: you must imagine objects or particles interacting between themselves to understand concepts. This only is possible if your brain manage positions of objects as well their features (compositionality).
The new theory (1000s of brains) will clear many lacunes of machine learning not filled by the current approachs.
You get it, David!
Amazing video
As always, excellent content. And the whole grid cell idea might have implications beyond the human brain. I won't say any more.
Cells ko train kaise kre
Thank you form making this video holy moly
This was fantastic!!!
@HTM School
This is extremely interesting.
Do we know what the mouse's knowledge of position (the grid cells firing) is based on? Like distance walked, sense of inertia, or the earths magnetic poles?
So many questions... (some that probably matter some that don't, and no I don't expect an answer from you for these) does the position of the grid dots change with different iterations of the experiment? if so does it depend on where the mouse is placed in the room and in what orientation? if the grid does not vary when taking the mouse out, would two separate rooms placed side by side, that are not axis aligned, have a grid that mesh with each other or would the grid orient to the walls of the room? what happens when objects are added to the room? does the mouse grid ever drift, shift, rotate or recalculate over time? does grid cell fire for a sleeping mouse that is moved by someone else? do mice use an A* algorithm on a 3d hexgrid? (I could go on but I won't)
Really cool stuff (If you know of where to find more information on these experiments a link would be great.)
Hopefully you saw the papers I linked in the show description. That is where I got most of this info.
Forgive me but I want to take a look at two separate scenarios: A) A mouse sits and it fires at any time upon sitting or resting. B) A mouse moves thoroughly (without rest or stop) and it fires at any time. What's the ratio for non-motion to motion? With open interpretation (combining three possibilities with each respective ratio): High non-motion to Low motion -- It needs to sit in its familiar place that is already registered -- so it doesn't need to learn its own place and has the time to think something else entirely - producing new information, accessing to non-related-memory-to-current-event-or-that-it-keeps-adding-attribution-to-current-event. Low non-motion to High motion -- When it's in the "comfort zone", it's already recorded all the previous-knowledge and there is nothing new to fire. But when it's moving around, all small attribution, details, motion, the gravity that throw the mouse off -- it constantly learning new information and if by some chance, some associative memory is familiarized or triggered, and it keeps adding new sequence whenever it trip, small something new, a change of shade...It doesn't consent to his usual rigid routine -- The nature or its reality keeps 'teaching' to the user of his existence.
There are other cells in the brain called "speed cells" that respond to an animals speed.
Smartest stuff I've ever whitnessed
Where can I find the website where you show the demonstration with the mouse?
numenta.com/blog/2018/05/25/how-grid-cells-map-space/
You all should get the Nobel prize!
Incredible
Fantastic :)
So presumably one's "map" of a musical work is represented by an analog of grid cells. And perhaps the "internal clock" is related to the entrorhinal-hippocampal system too?
amazing video as always, but I get one question :
I see how this system work to represent location when you have already your position in the world, you can translate that position to the gird cell.
but how does the cell do to know where it is in the world, and so fire.
I can see that work by having a internal position in the brain that is update every time you move for example, but how does the mouse does to know where is position is when we put in in the box at first ?
See paper “Computational models of grids cells” in video description
Brains re-map whenever one enters a new space; Grids are (re)established by boundaries one senses or recalls.
Curious why you liken it to hexagons, not triangles? Also, why would the third dimension have a different form than the rest?
hexagons are more fun
@@NumentaTheory Well, they _are_ a bunch of triangles having a party...
are distances same for human and mice, or are they proportional to their body weight?
Bravo!
Grid cells
- deals with location in the brain
Place cells
Paper - evidence for grid cells in a human memory network
Is the demo used in this video available anywhere?
discourse.numenta.org/t/how-to-run-htm-school-visualizations/2346
incredible explanation . RIP Matt You will be missing Numenta s team.
Very interesting. I wonder if there's been any implementation of that into AI systems.
Lots of folks working on it!
He just turned the dad jokes notch into 11 around 13:40 and they are actually no jokes, but the truth.
Looking snazzy today.
The whole grid cell thing doesn't explain how a grid neuron knows when it's time to fire. In the case of sight for instance if a neuron is sensitive to vertical lines we can easily see how that neuron knows when to fire: if neurons arranged in a straight vertical line on the retina fire, then that particular neuron will know a vertical line has been detected. In the case of space, how does a neuron know that I have moved to the location it's supposed to represent? There must be a whole lot of processing of movement and vision feeding into the grid cells, much like there is a lot of processing done in the retina, before other neurons can know anything about the world.
Right, we are not explaining how grid cell behavior emerges, but there is tons of research on that in neuroscience today.
is the demonstration tool you use starting from 2:25 available online?
numenta.com/blog/2018/05/25/how-grid-cells-map-space/
Excellent video. The 3D grid seems to naturally be tetrahedral rather than strictly orthogonal: A packing of spheres. This extends to hyperspaces as represented in Pascal's Triangle. Is there any way to test this experimentally? Has it been so tested? See Close packing of equal spheres: en.wikipedia.org/wiki/Close-packing_of_equal_spheres
Haven’t tested it ourselves but that’s how I imagine it.
This guy is good…
Deep man. Oh wait
Information is expressed spacially by default for a few reasons. Reuse of course.
But on deeper level concepts have a relational space. Or we think they do.
Chicken or egg?
I thought this was taking too long to come out because you were 'gridlocked'.
Main raste bhul jata hu iska koi teatment hai
can you link the website you used for the visualization ?
www.numenta.com/blog/2018/05/25/how-grid-cells-map-space/
thanks, man great stuff. You should check out this 3d grid paper www.biorxiv.org/content/early/2018/03/14/282327
Yes, i read that one too. :)