Live data from Hacker News

Consciousness is a recurrent neural network

news.ycombinator.com

1–10 of 50 posts

Consciousness is a recurrent neural network

#1
One of the problems with consciousness is that you know that you are conscious but you can't know if others are. Consciousness is like seeing a dog in front of you but only you can see it.

Let us begin with an example of seeing a dog: when you see a dog, photons from the dog come to your eyes and converted to neural signal, then, a neural network will recognize that what you're seeing is a dog.

How do you know that you're seeing a dog?

I believe we can explain this in the same way as seeing a dog. But in this case, what you see is not photons but signals from within brain itself. In other words, brain takes its current states (signals) as its input (this is possibly no different from brain takes signals created by your eyes as its input.) And as you may already known, this is similar to recurrent neural network.

In this way, consciousness is a concept learned by brain about the inner processes of brain itself which is similar to: dog is a concept we learned when seeing many dogs!

Stream of consciousness is the result when brain trying to model itself and continuously receiving signals from itself!

We can put this further and ask: is any brain of any kind (dog, cat, whale) conscious?

My idea is that there is a threshold when brain becomes complicated enough to model its own signal!

Re: Consciousness is a recurrent neural network

#2
When i saw the title, i expected this post to be somebody's 'stoner philosophy' of what conciousness is. But i really like this idea. If true, it means conciousness is naturally emergent as soon as networks can be efficiently trained to represent complex enough functions. And we might not be too far away.

How do you define conciousness? Like the voice in your head? Are you saying that voice is essentially the predicted copy of you I.e. what you predict you would do in the current situation? I think this could explain why we have a concious and unconscious mind, the unconscious is the actual brain, the conscious is just our prediction of what we would do in the current situation.

Re: Consciousness is a recurrent neural network

#3
Consciousness is the searchlight of attention in an AI Mind.

                                         __________             
     /^^^^^^^^\     | |                 /          \     /^^^^^^^\
    /   EYE    \    | |                /  FreeWill  \   /  EAR    \
   /  SENSORY   \   | |               /   VOLITION   \ | SENSORY  |
  |    ORGAN     |  | |   _________   \    MODULE    / |  ORGAN   |
  |              |  | |  /         \  /\____________/  |          |
  |   visual     |  | | /  EMOTION  \/   |       |     | auditory |
  |   memory     |  | | \  MODULE   /    |   ____V___  | memory   |
  |   channel    |  | |  \_________/     |  / ROBOT  \ | channel  |
  |   _______    |  | |    ________      | ( MOTORIUM )|          |
  |  / BIRD  \   |  | |   /        \     |  \________/ |          |
  | / image   \--|----+->( VisRecog )  __V__________   |          |
  | \ engram  /  | a| |   \________/  /             \  |          |
  |  \_______/   | b|c|       |      ( DEEP THINKING ) |          |
  |              | s|o|f      |       \   MODULE    /  |          |
  |              | t|n|i      |       /\___________/   |          |
  |              | r|c|b   ___V______/        |        |          |
  |              | a|e|e  /          \        |        |          |
  |              | c|p|r |            |-------|------->|---"I"    |
  |              | t|t|s |            |   ____V_____   |          |
  |              |  | |  |            |  /          \  |          |
  |              |  | |  | NounPhrase | ( VerbPhrase )-|--"SEE"   |
  |              |  | |  |   MODULE   | /\  MODULE  /  |          |
  |              |  | |  |            |/  \________/   |          |
  |              |  | |  |            |       |        |          |
  |    ______    |  | |  |            |-------|------->|-"BIRDS"  |
  |   /      \   |  | |   \__________/        |        |          |
  |  /  new   \  |  | |        |              |        |          |
  | /  percept \ |  |_|________V______________V____    |          |
  | \   of     / | /                               \   |          |
  |  \  BIRD  /--|-\ Psy CONCEPT ARRAY             /   |          |
  |   \______/   |  \_____________________________/    |          |
http://mind.sourceforge.net/conscius.html

Re: Consciousness is a recurrent neural network

#5
This explanation feels a little tautological to me. We knew this already because physics doesn't care what is a brain and a dog. All physics knows is that there are atoms, and whether they belong to a brain or a dog doesn't really matter. It's all the same.

And this theory fails to explain how we can feel pleasure or pain, and, it fails to predict whether an artificial neural network can feel pleasure or pain.

Re: Consciousness is a recurrent neural network

#6
You need to study some actual neuroscience.

some things to remember:

- neurons are not linear.

- neurons are not time-invariant.

- neurons are not causal.

- neurons and actual anatomical connectivity do not map well to electronics-inspired wiring diagrams. (see: dendritic arbor)

- anatomical structure does not imply functional connectivity.

- functional connectivity does not imply anatomy!

- synaptic junctions respond more or less well to different neurotransmitters, all of which are continuously present, at different times. We don't know why.

- So far, what we have observed about anatomy and physiology of the brain does not look like a RNN.

- The relationship between the meat architecture and the phenomena it hosts may not be as 1:1 as you'd like. Opinions vary.

My favorite analogy about relating brains to consciousness is this: "if you believe that brains are like computers, (which you shouldn't, but just for the sake of argument, let's), then even if you really and truly produce a full map of the brain, what you've _got_ is the spec sheet for an x86 processor. What you _want_ is the user manual for Mac OS."

Re: Consciousness is a recurrent neural network

#7
post #5

This explanation feels a little tautological to me. We knew this already because physics doesn't care what is a brain and a dog. All physics knows is that there are atoms, and whether they belong to a brain or a dog doesn't really matter. It's all the same. And this theory fails to explain how we can feel pleasure or pain, and, it fails to predict whether an artificial neural network can feel pleasure or pain.

It's just survival - when a species gets good enough at survival, it learns to adapt to the environment. First, it's just simple reflex, trigger and response, but with evolution, species have become more capable to adapt. We, humans, need to adapt not just to nature, but also to the human society and technology, and human consciousness is just that function. Without it we would be dead of hunger and thirst in 3 days. What is special about conscious systems is that they adapt and preserve themselves in the face of adversity.

Pleasure and pain are related to reward signalling in reinforcement learning. Humans have inborn reward signals which have been selected by evolution. Artificial agents have reward signals too, but they are designed.

So I think the path to consciousness is to create artificial agents that learn to survive and adapt. When such a complex behavioral system is designed, it naturally learns about its relation to the world. It has consciousness by extension of being capable to protect its life.

Re: Consciousness is a recurrent neural network

#8
Great explanation, makes perfect sense to me! My own heory is pretty much the same. Just like you can experience other internal sensations in your body when nerves send signals that are recognized by the brain, the brain can recognize it's own signals.

I highly recommend to read Godel Escher Bach(if you dont have time - just read the introduction to get the general idea). In this book author explains how the meaning arises when things(like language, or math equations, or neurons in the brain) "mirror" things in the real world, when their structure can be "mapped" onto some other structure(so called "isomorphism").

He says that brain "mirrors" the world around it as it builds a world model. But the brain itself is a part of the world, so it builds the model of itself as well. Neurons recognizing/observing/experiencing other neurons.

Just like you can see a dog you can "see" your own brain state.

I've also heard a cool quote somewhere - "Consciousness is simply just what it feels like to have a brain". You can close your eyes and feel the position of your body, you can feel your stomach being full, and you can feel your brain thinking.

Re: Consciousness is a recurrent neural network

#9
Can't the same argument be used to show that consciousness is a whole bunch of things that it clearly is not. Just arguing that consciousness has certain attributes and it shares those attributes with something else, doesn't make consciousness an example of that something else.

I think there is a general trend towards thinking of consciousness as an emergent phenomenon which exists when a system is complex enough and has certain attributes, such as memory, information processing capability, etc.

But none of these definitions enlighten me as to when or why such a system might become self-aware.

Re: Consciousness is a recurrent neural network

#10
The problem with this line of thinking is that any data model for anything has absolutely no intrinsic meaning. It's just an encoding.

To give the simplest example: you can model a color in many different encodings, 'red', 'hsl(0, 50%, 50%)', 'rgb(255, 0, 0)', etc.

Any "thing" can be encoded in an infinitely different number of ways. The same applies to the state of your mind. Why should one encoding give rise to consciousness?

Arguably, the manner in which water is flowing within a sewage pipe network could be interpreted to be encoding some kind of information. Would you argue that some particular arrangement of water flow within a pipe network could give rise to consciousness?

I think there's a difference between a system modelling itself and becoming conscious of itself.

Any system can model itself. We can do it easily with computers. Arguably the linux kernel has a model of itself, its hardware, its inputs, etc. That doesn't make it conscious.

Post reply on HN