Predictive coding has been unified with backpropagation • Hacker News Board

Re: Predictive coding has been unified with backpropagation

#81

Post by andyxor » Tue, Apr 06, 2021, 9:32 PM UTC

eli_gottlieb wrote: post #80

andyxor wrote: post #57

Earlier quoted context omitted.

I suggest starting with the works by Howard Eichenbaum on memory and Edvard & May-Britt Moser (and John O'Keefe and Lynn Nadel) on place & grid cells. For the latest and greatest see https://twitter.com/doellerlab https://twitter.com/KordingLab https://twitter.com/preston_lab https://twitter.com/memorylab https://twitter.com/ptoncompmemlab https://twitter.com/MillerLabMIT https://twitter.com/hugospiers Once you start…

We know a fair bit about how cognitive maps work in 2D and 3D Euclidean environments. We know damn little about how nontrivial manifold structure can be learned, particularly in spaces with more than three dimensions.

spatial cognitive maps used for navigation are extendable to arbitrarily high dimensional spaces for abstract concept representation, using pretty much the same machinery.

There is a ton of work on this, both theory and empirical evidence, here are just a few:

"Navigating cognition: Spatial codes for human thinking" https://science.sciencemag.org/content/362/6415/eaat6766.abs...

"Organizing conceptual knowledge in humans with a gridlike code" https://science.sciencemag.org/content/352/6292/1464

"The Hippocampus Encodes Distances in Multidimensional Feature Space" https://www.sciencedirect.com/science/article/pii/S096098221...

"A non-spatial account of place and grid cells based on clustering models of concept learning" https://www.nature.com/articles/s41467-019-13760-8

"A learned map for places and concepts in the human MTL" https://www.biorxiv.org/content/10.1101/2020.06.15.152504v1....

"What Is a Cognitive Map? Organizing Knowledge for Flexible Behavior" https://www.sciencedirect.com/science/article/pii/S089662731...

"A map of abstract relational knowledge in the human hippocampal–entorhinal cortex" https://elifesciences.org/articles/17086

"Map-Like Representations of an Abstract Conceptual Space in the Human Brain" https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7884611/

"Knowledge Across Reference Frames: Cognitive Maps and Image Spaces" https://www.sciencedirect.com/science/article/pii/S136466132...

"Concept formation as a computational cognitive process" https://www.sciencedirect.com/science/article/pii/S235215462...

"Efficient and flexible representation of higher-dimensional cognitive variables with grid cells" https://journals.plos.org/ploscompbiol/article?id=10.1371/jo...

"The cognitive map in humans: spatial navigation and beyond" https://www.nature.com/articles/nn.4656

"A general model of hippocampal and dorsal striatal learning and decision making" https://www.pnas.org/content/117/49/31427.short

"On the Integration of Space, Time, and Memory" https://www.sciencedirect.com/science/article/pii/S089662731...

Re: Predictive coding has been unified with backpropagation

#82

Post by nl » Tue, Apr 06, 2021, 11:07 PM UTC

cs702 wrote: post #77

nl wrote: post #61

Earlier quoted context omitted.

I don’t think anyone familiar with the field is in anyway surprised by this results. The breakthrough seems really limited to showing it holds for graphs. We already knew this was practically true though anyway.

Agree, no one is surprised. But the authors successfully show how to train CNNs, RNNs, and LSTM RNNs without backpropagation , i.e., every layer learning only via local rules , without having to wait for gradients to be backpropagated to all layers before the entire model can move on to the next sample. As I understand it, this work has paved a path for training very large networks in massively parallel, fully distri…

> But the authors successfully show how to train CNNs, RNNs, and LSTM RNNs without backpropagation, i.e., every layer learning only via local rules

The basic version of this was shown in [1], as mentioned by the ICLR review:

"Specifically, the original paper by Whittington & Bogacz (2017) demonstrated that for MLPs, predictive coding converges to backpropagation using local learning rules."

That Whittington & Bogacz didn't extend to complex ANN architectures, but it would have been very surprising if what they showed didn't extend to other ANNs.

OTOH, while local-only updates are great it doesn't help much if the overall algorithm needs vastly more iterations. Again, from the ICLR review: "The increase in computational cost (of 100x) is mentioned quite late and seems to be glossed over a bit."

[1] https://pubmed.ncbi.nlm.nih.gov/28333583/

Re: Predictive coding has been unified with backpropagation

#83

Post by klmadfejno » Wed, Apr 07, 2021, 5:50 PM UTC

TheRealPomax wrote: post #76

klmadfejno wrote: post #72

Earlier quoted context omitted.

That's, at best, pedantically true. You can determine the function of individual components of a network, and they will correspond to concrete thing. It's just that the utility of doing this is low in the scheme of things and the function of individual components is going to be fuzzier than nice constructs that humans like to think in. If you wanted to take painstaking steps to align the functionality of nodes to ide…

That's literally what we cannot do in a neural net. It's why they're so problematic: you can determine the propagation functions of individual nodes perfectly, and that knowledge tells you exactly nothing about all of the many things it's values contribute to. There is no "concrete thing" at the node level: a single node fundamentally can't see a wing, or a color, or anything else, that's only emergent behaviour of n…

That's the machine learning 101 worldview. Initial level nodes are likely just minor inscrutable transformations. Later layers will be coding for features that, at least to some degree if it's a tractable problem, humans can understand and agree with as useful features. They'll be fuzzy and not as clearly defined as a human would frame the problem, but their purpose can be explored and generally identified.

In any case, speaking of them as representing singular features for simplification is appropriate. Maybe it's not one node that codes for legs, but two nodes that codes for legs like this and legs like that, but that's not relevant to the point.

Re: Predictive coding has been unified with backpropagation

#84

Post by cs702 » Thu, Apr 08, 2021, 2:54 PM UTC

nl wrote: post #82

cs702 wrote: post #77

Earlier quoted context omitted.

Agree, no one is surprised. But the authors successfully show how to train CNNs, RNNs, and LSTM RNNs without backpropagation , i.e., every layer learning only via local rules , without having to wait for gradients to be backpropagated to all layers before the entire model can move on to the next sample. As I understand it, this work has paved a path for training very large networks in massively parallel, fully distri…

> But the authors successfully show how to train CNNs, RNNs, and LSTM RNNs without backpropagation, i.e., every layer learning only via local rules The basic version of this was shown in [1], as mentioned by the ICLR review: "Specifically, the original paper by Whittington & Bogacz (2017) demonstrated that for MLPs, predictive coding converges to backpropagation using local learning rules." That Whittington & Bogacz…

In my view, there's a big difference between successfully training, say, LSTM RNNs, versus successfully training "vanilla" MLPs.

This work opens the door for using new kinds of massively parallel "neuromorphic" hardware to implement orders of magnitude more layers and units, without requiring greater communications bandwidth between layers, because the model no longer needs to wait until gradients have back-propagated from the last to the first layer before moving on to the next sample.

Scaling backpropagation to GPT-3 levels and beyond (think trillions of dense connections) is very hard -- it requires a lot of complicated plumbing and bookkeeping.

Wouldn't you want to be able to throw 100x, 1000x, or even 1Mx more fully distributed computing power at problems? This work has paved a path pointing in that direction :-)

Re: Predictive coding has been unified with backpropagation

#85

Post by cs702 » Thu, Apr 08, 2021, 2:58 PM UTC

YeGoblynQueenne wrote: post #78

cs702 wrote: post #75

Earlier quoted context omitted.

Yes, I linked to that same page at the top of my comment :-)

Oops, sorry- I noticed the link but I thought it was a HN url, like the one before... and I wondered why it was greyed-out (visited). But still I didn't check it out. My very bad.

No worries :-)