Google’s AI program DeepMind learns human navigation skillshttps://buff.ly/2jIwgYH
-
-
Also, I felt that this article skirted many of the most interesting aspects of grid cells, such as the effects of different geometries, compartmentalization, rotation, etc. It jumps almost immediately into AI applications.
-
I was also sadly struck by how the article completely ignored relevant prior work on this topic by Dordek et al (2016), Gustafson & Daw (2011), and Stachenfeld et al (2017; who is thanked in the acknowledgments!)
-
There is also work by Konidaris on RL with a Fourier basis (not cited)
-
Just to be clear, I completely understand that they can't cite everyone. But my worry is that their framing is contributing to a perception (at least in the media) that DeepMind is operating in an intellectual vacuum.
-
Last thing I'll say: bridging neuroscience and AI is great, but engineering demos don't substitute for real models of biology. To claim that your system emulates the brain, you have to grapple with the empirical phenomena in all their richness and complexity.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
To be fair, integrating velocity to track position is not really the same task as starting with position. That said, I find the PCA result -- and especially Stachenfeld's elaboration -- much more informative about what the grid cells are doing.
-
I take it back. I didn't realize the output of the path integration was to produce place cells given as a supervisory input. I agree that although it's set up as a supervised path integration problem, this amounts to the same thing as the PCA result.
-
I think the framing is symptomatic of a larger issue for people who study neural networks. By training a network instead of analyzing the problem mathematically, the authors can say that grids "emerge", which sounds cooler than "grids are the eigenvectors of place cells".
-
Yes and no. There’s an additional empirical question being addressed: how does the brain arrive at a given representation? NN research like this demonstrates that optimizing on cost fns produces solutions like those in the brain, which suggests the brain also optimizes cost fns.
-
But these basis functions are the simplest (irreducible) basis functions that span translations and rotations. They're the only ones that allow for vector like computation of locations.
-
Right, so that's why I agree with
@gershbrain's criticism of this specific paper: the result is a bit obvious given how they set the task up, and LSTMs are not required. But, I think the general study of emergent representations in NNs is an important one. -
When people say something "emerges" they're basically saying "I don't know why this happens." If I run a linear regression, I don't say the coefficients "emerge" from maximum likelihood. They are a mechanistic consequence of an objective function and an optimization procedure.
-
Fascinating thread. I can't speak to whether grid cells emerge, but I don't agree with your statement about emergence in general. Many real-world phenomena cannot be captured or approximated by a few understandable equations, because they span multiple scales...
- 9টি আরও উত্তর
নতুন কথা-বার্তা -
-
-
It's also not particularly impressive to rave about beating people. Wristwatches were better than people at long division when I was 2 years old.
-
I find the entire deepmind-nature relationship a little gross and hype serving to be honest. Are these papers really advancing the field in a major way that the first 2-3 Nature papers didn't? But I'm probably just jealous


-
Yea, I read these papers and don't feel like I've learned anything at all (about the brain, that is). This is not to say that insights will not be forthcoming, and maybe we just need to keep turning that crank.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
ধন্যবাদ। আপনার সময়রেখাকে আরো ভালো করে তুলতে টুইটার এটিকে ব্যবহার করবে। পূর্বাবস্থায়পূর্বাবস্থায়
-
-
-
I was wondering that too
ধন্যবাদ। আপনার সময়রেখাকে আরো ভালো করে তুলতে টুইটার এটিকে ব্যবহার করবে। পূর্বাবস্থায়পূর্বাবস্থায়
-
-
-
Thanks for pointing that out. It’s good to hear the other sides in the interest of balance. They do do a good job of promoting the field to the general public though. Keeps the money rolling in and the AI winters at bay!
ধন্যবাদ। আপনার সময়রেখাকে আরো ভালো করে তুলতে টুইটার এটিকে ব্যবহার করবে। পূর্বাবস্থায়পূর্বাবস্থায়
-
লোড হতে বেশ কিছুক্ষণ সময় নিচ্ছে।
টুইটার তার ক্ষমতার বাইরে চলে গেছে বা কোনো সাময়িক সমস্যার সম্মুখীন হয়েছে আবার চেষ্টা করুন বা আরও তথ্যের জন্য টুইটারের স্থিতি দেখুন।