The Strange Loop in Deep Learning | Hacker News Reader

The Strange Loop in Deep Learning

medium.com

66 pointshexrcs24 comments

However, as we are all now beginning to discover, the employment of ‘feedback loops’ are creating one of the most mind-boggling new capabilities for automation. This is not hyperbole, [...]

Always glad to see references to Hofstadter's work, but this sounds like hyperbole. I don't think that mere feedback loops are "strange" enough to invoke his name. I think the idea refers to loops in systems which are heterarchical. I don't think there are any interesting crossings over different layers of abstraction here.

Interestingly, these ideas have been around for quite some time. For example, the work of Warren McCulloch (and Walter Pitts) predates Hofstadter's thought in this space. Relevant passage from his 1945 article, "A Heterarchy of Values Determined by the Topology of Nervous Nets":

"[..] three heterodromic branches link the dromes so as to form a circle in the net which is distinguished from an endrome in that it is not the circuit of any drome but transverse to all dromes, i.e., diadromic. The simplest surface on which this net maps topologically (without a diallel) is a tore. Circularities in preference, instead of indicating inconsistencies, actually demonstrate consistency of a higher order than had been dreamed of in our philosophy."

How do you know whether something is a hyperbole? When the author claims something is not a hyperbole.

Quoting Douglas hofstader to begin a deep learning article is akin to quoting Einstein to begin an article on quantum mechanics...

Douglas Hofstader is against the currrent direction that artifice intelligence is headed .

Especially as this really isn't about "Strange Loops" as coined by Hofstadter, just... regular loops.

Is there somewhere to read more about his objections?

His wikipedia page says

>Hofstadter has said that he feels "uncomfortable with the nerd culture that centers on computers". He admits that "a large fraction [of his audience] seems to be those who are fascinated by technology", but when it was suggested that his work "has inspired many students to begin careers in computing and artificial intelligence" he replied that he was pleased about that, but that he himself has "no interest in computers".[30][31] In that interview he also mentioned a course he has twice given at Indiana University, in which he took a "skeptical look at a number of highly-touted AI projects and overall approaches".[15] For example, upon the defeat of Garry Kasparov by Deep Blue, he commented that "It was a watershed event, but it doesn't have to do with computers becoming intelligent".

Which sort of corroborates what the person you're replying to said.

Out of curiosity, can you elaborate?

I suggest reading "fluid concepts and creative analogies", it's a really great read.

Looks interesting, thanks. Found a fun fact about this book after Googling, turns out it was the very first book to be sold on Amazon.com!

> akin to quoting Einstein to begin an article on quantum mechanics

This analogy doesn't work - Einstein discovered the photoelectric effect.

Yes and hofstader is one of the pioneers of artificial intelligence but he has since been disappointed with where the field has gone

Quantum Mechanics is a well-defined subject. "Artificial Intelligence" certainly is not. You could say that the builders of the Antikythera Mechanism were pioneers of artificial intelligence, long before Babbage, Lovelace, or Turing. You could say that everyone from Grace Hopper to Herbert A. Simon to Noam Chomsky was a pioneer of artificial intelligence. It doesn't really mean anything.

If you take "Artificial Intelligence" to mean "Connectionist Modeling in Silicon" or "Deep Learning" instead, then my point stands.

He was against people calling Deep Blue AI when it was not. I doubt he would say the same thing about AlphaGo.

I've always thought that consciousness seems a lot less mysterious if you consider it as a feedback loop within the neocortex. We know that the brain takes raw sensory input and turns it, via a neural hierarchy of increasing abstraction, into a high-level, integrated model. We also know that there is a lot of additional neural connectivity, both across different sections of the neocortex, and from higher levels back to lower ones.

If you consider this diffuse connectivity not just as a regulation method, but as another form of sensory input, then it stands to reason that the brain would form an additional self-referential hierarchy of abstraction and pattern recognition based on its own internal state, and that this "sense", a literal self-awareness, would be very much like what we experience as consciousness.

Self-awareness isn't the fundamental issue with consciousness. It's the colors, sounds, tastes, feels, etc. Why are those a problem? Because the physical world isn't colored, doesn't taste like anything, doesn't feel like anything, etc.

So somehow, a physical system - the brain, creates or is correlated with those conscious sensations. Explaining this in terms of physics, function, computation, math, etc is hard because those fields lack any such sensations. An equation doesn't feel heat or see red, nor do atoms bumping around in the void. So from whence do those conscious sensations come from? Do they spontaneously emerge somehow from the complexity of a nervous system? Is the right sort of flow of electricity result in the smell of a rose? Is it the right algorithm that implements consciousness, if you just arrange the bits just right?

The best I've been able to come up with is that what we think of as "color" or "sound" (and similar) are in fact our brain's model of those aspects of the physical world. Of course you're not experiencing the physical world directly, but your brain does have a representation of the world based on its sensory inputs.

You might be interested by the Sensorimotor Contingencies theory of O'Regan. Here's a summary [0]. I find it really convincing. He also wrote a book if you want to go deeper in the subject (the final draft version is available on his website [1]).

0: http://www.whatfeelingislike.net/tiki-index.php?page=The+Sen... 1: http://nivea.psycho.univ-paris5.fr/

I think I disagree. The hard part is understanding why we have a sense of "self" one that can independently perceive it's place among the solar system and beyond.

The "colors, sounds, tastes, feels, etc" that you refer to can be seen as sensors that perceive narrow views of the various spectrums created by the energy bouncing around in the physical universe.

Examples:

- Hot vs. cold for a human is entirely different that hot vs. cold for a star.

- Color, from our perspective, is the spectrum of light that we need to utilize to effectively interact with our environment (different animals need and perceive different spectrums of light).

- Sound is a vibration that propagates as a [..] mechanical wave of pressure and displacement, through a transmission medium such as air or water. (taken from wikipedia.org/wiki/Sound)

My argument is that they do emerge, though not spontaneously. They are the emergent result of the neural abstraction hierarchy that the brain forms based on the sensory input it receives, and its own sense of that sensation. The "feel" of something is therefore a meta-sensation, a sense of what a sense is like, in comparison with another, within a neural abstraction that models the brain itself.

Like you said, colors, tastes, sounds, etc. aren't "real". They themselves are emergent phenomenon. A colour is arrangement of electromagnetic radiation, a sound is an arrangement of air molecules. And radiation and matter are themselves the emergent result of the interaction of even simpler forces. Often, the nature of the emergent phenomenon seems inexplicably complex in comparison with the basic interactions that give rise to it.

If you examine intelligent life, you can see that it sits at the top of a pyramid of emergent complexity, reaching down through biology, chemistry and physics into quantum physics. It would be very strange indeed if, at the very top of this pyramid, there was another complex phenomenon that somehow did not arise from the interaction of simpler parts beneath it.

Self referential loops in the brain may indeed be the mechanism for our sense of self awareness but to go on and say this makes consciousness any less mysterious is a non-sequitur.

How would any combination of neurons and interconnections between the neurons (feedback loops or not) explain the emergence of a subjective first person experience i.e. consciousness

This might be a confusion of words and not reality. If you define consciousness as self awareness then we are talking about two different things.

What do you imagine as the difference between consciousness and self-awareness?

My point is exactly that a subjective first-person experience is self awareness. The root of all conscious experience is an awareness of own awareness, a sense of our own senses, the difference between them, etc. I would challenge anyone to come up with a definition of conscious experience that does not, ultimately, boil down to a form of self-referential sensory input.

Don't recurrent neural networks (which have been around in several variants for a few decades) incorporate feedback loops in the sense described in this article?

I'm curious how a data flow graph could learn representations of itself and of arbitrary human concepts, but I somehow suspect you'd end up with an unilluminating generator of JSON strings with many of our current architectures.

I think the next revolution in machine learning is when we start connecting specialized neural networks to each other. A neural network of neural networks, basically.

Reply on news.ycombinator.com