Hacker News Reader

Comment by hctaw | Hacker News Reader

I don't know enough about biology or ML to know if what I'm posting below is totally wrong, but here goes.

"Backprop" == "Feedback" of a non-linear dynamical system. Feedback is mathematical description of the behavior of systems, not a literal one.

I don't know of BNNs are incapable of backprop anymore than an RLC filter is incapable of "feedback" when analyzing the ODE of the latter tells you that there's a feedback path (which is what, physically? The return path for charge?)

So what makes BNN incapable of feedback? Are they mechanically and electrically insulated from eachother? How do they share information, and what is the return path?

Other than that I wish more unification was done on ML algorithms and dynamical systems, just in general. There's too much crossover to ignore.

View on HN

> Other than that I wish more unification was done on ML algorithms and dynamical systems, just in general. There's too much crossover to ignore.

Check out this work, "Deep relaxation: partial differential equations for optimizing deep neural networks" by Pratik Chaudhari, Adam Oberman, Stanley Osher, Stefano Soatto & Guillaume Carlier.

https://link.springer.com/article/10.1007/s40687-018-0148-y

The back-prop learning algorithm requires information non-local to the synapse to be propagated from output of the network backwards to affect neurons deep in the network.

There is simply no evidence for this global feedback loop, or global error correction, or delta rule training in neurophysiological data collected in the last 80 years of intensive research. [1]

As for "why", biological learning it is primarily shaped by evolution driven by energy expenditures constraints and survival of the most efficient adaptation engines. One can speculate that iterative optimization akin to the one run by GPUs in ANNs is way too energy inefficient to be sustainable in a living organism.

Good discussion on biological constraints of learning (from CompSci perspective) can be found in Leslie Valiant book [2]. Prof. Valiant is the author of PAC [3] one of the few theoretically sound models of modern ML, so he's worth listening to.

[1] https://news.ycombinator.com/item?id=26700536

[2] https://www.amazon.com/Circuits-Mind-Leslie-G-Valiant/dp/019...

[3] https://en.wikipedia.org/wiki/Probably_approximately_correct...

I think there's a significant difference worth illustrating that "there is no feedback path in the brain" is not at all equivalent to "learning by feedback is not possible in the brain."

It's well known in dynamics that feed-forward networks are no longer feed-forward when outputs are coupled to inputs, an example of which would be a hypothetically feed-forward network of neurons in an animal and environmental conditioning teaching it the consequences of actions.

I'm very curious on the biological constraints, but I'd reiterate my point above that feedback is a mathematical or logical abstraction for analyzing the behavior of the things we call networks - which are also abstractions. There's a distinction between the physical behavior of the things we see and the mathematical models we construct to describe them, like electromechanical systems where physically no such coupling from output-to-input appears to exist, yet its existence is crucially important analytically.

The article says this:

> The backpropagation algorithm requires information to flow forward and backward along the network. But biological neurons are one-directional. An action potential goes from the cell body down the axon to the axon terminals to another cell's dendrites. An axon potential never travels backward from a cell's terminals to its body.

The point of the research here is that backpropagation turns out not to be necessary to fit a neural network, and that it can be approximated with predictive coding, which does not require end-to-end backwards information flow.

So... I don't understand.

>An action potential goes from the cell body down the axon to the axon terminals to another cell's dendrites.

How do you figure that doesn't allow backprop?

A neuronal bit is a loop of neurons. Information absolutely can back- propagate. If it couldn't, how does anyone think it'd be at all possible to learn how to get better at anything?

Neuron fires dendrite to axon, secondary neuron fires dendrite to Axon, Axon branches back to previous neuron's dendrites, rinse, repeat, or add more intervening neurons... Trying to disinclude backprop based on the morphology of a single neuron is... Kinda missing the point.

It's all about the level of connection between neurons and how long or whether a signal returns unmodified to the progenitor that effects the stability of the encoded information or behavior. At least to the best I've been able to plausibly model it. Haven't exactly figured out how to shove a bunch of measuring sticks in there to confirm or deny, but I just can't how a uniderectional action potential forwarding element implies lack of backprop in a graph of connections fully capable of developing cycles.

> How do you figure that doesn't allow backprop?

When you backprop through a linear layer (a matrix W), you need to multiply with W.transpose which is impossible if connections are not bidirectional.

> Information absolutely can back- propagate. If it couldn't, how does anyone think it'd be at all possible to learn how to get better at anything?

Local error aggregation can have a similar effect with backprop but you can run each layer in parallel, you don't need to wait for the signal to reach the loss function and then gradients to come all the way back.

An interesting read: Decoupled Neural Interfaces using Synthetic Gradients (DeepMind, 2017) http://proceedings.mlr.press/v70/jaderberg17a/jaderberg17a.p...

You're conflating the implementation with the principle. There is no matrix math with neurons; quite the opposite, we posit the existence of the matrix math from the behavior we've observed with neural systems governed by a sigmoid function. The equations we've derived are secondary to the initial implementation. Just as you tweak the error factor in backprop, so too do weights between intersecting neuron networks adjust until thought and intention falls into line with eventual perception and execution.

The map is not the territory.

> There is no matrix math with neurons

Are we talking about the brain or neural nets biological plausible neural nets?

> If it couldn't, how does anyone think it'd be at all possible to learn how to get better at anything?

Presumably part of the feedback loop (at least for things like motor skills and rote memorisation) is external to the brain. Our brain causes us to act, which alters our perceptions, which causes the brain to adjust.

Annnd... What exactly encodes that perception thing? The brain! There's nothing magicabout sense data except that it's how we're used to tweaking things. It is still, in the end, just more neurons telling other neurons what's off. Sure, some are hooked up to photoreceptors. I could swap those out with about anything else and still get learning. Unfortunately, the practical benefit of mastery of stuff you imagine tends to be less useful than it feels like it should be IRL.

>> If it couldn't, how does anyone think it'd be at all possible to learn how to get better at anything?

Personally, when I've come to the point where I'm thinking to myself "that must be it, what else can it be?", I am at the point where I need to do more work to answer the latter part of the question.

Yeah, but then you run into the problem of computation speed. Any given neuron in the middle of your brain does 1 computation per second absolute maximum, and 1 per 10 seconds is more realistic. More to the outside (the vast majority of your brain) 1 per 100 seconds is a lot. And it slows down when you age.

This means brains must have a bloody good update rule. You just can't update a neural network in 1 billion operations per second, or 4e17 operations until you're 12, about 2 million training steps per neuron, or about half that assuming you sleep. You cannot get to the level of a 12 year old in 4e17 operations, because GPT-3 does more and while it's impressive, it doesn't have anything on a 12 year old.

Reply on news.ycombinator.com