Yeah, this is something I've consistently noticed + also been very weirded out by, with "What a neural network sees"/"Neural network image synthesis" pictures.
Many of them are eerily similar to the visual phenomena I've experienced on various psychedelics. To the point where you don't really have a choice but to recognize that there's some mathematic/scientific explanation of consciousness and the way psychedelics interact with it that relate to how neural networks function.
I'm far too dumb to put those pieces together but I'd believe whatever someone more intelligent put in front of me.
Wish it were possible to show other people what these kinds of experiences look/feel like, our language and vocabulary obviously don't have terms for most of it.
Seeing impossible colors or shapes, or being/existing in multiple points in space + time simultaneously, as different creatures or people, in different places. It can shake you to your core by challenging what you always assumed were fundamental givens of the universe and "the way things work" (TM).
> To the point where you don't really have a choice but to recognize that there's some mathematic/scientific explanation of consciousness and the way psychedelics interact with it that relate to how neural networks function.
I don't get any of these types of visuals (or any visuals whatsoever other than wobbling/throbbing edges/borders) on most of the common psychedelics on which most normal people get visuals, even at 2-10x standard doses. I'm aware of the stereotypical wooshy GAN-style trailer visuals people report; I've never seen them.
Either I'm an alien or your hypothesis has a hole.
I think an issue here is that I know what NN outputs look like and have some folk intuition as to what psychedelics to do my brain, so I end up seeing what I want to see.
And a bunch of others will do this. Psychedelics are all unique feeling but the further you push the doses the more similar they become.
At the furthest edge of the spectrum is the "grand psychedelic experience" which, in my incredibly unscientific & anecdotal opinion, is an (nearly) identical experience that you will reach every time, on any psychedelic, given the right circumstances and a high enough dosage.
But this experience is just the "Can't see my hand in front of my face because of the visuals" level.
You pass this level on the way to the last stop, if you take the ride all the way to end (whether you meant to take it there or not!)
You’re gonna love learning about optical illusions – bugs in our visual processing. We already see the world through approximation.
Obvious example: saccadic masking. Your brain shuts off visual processing during rapid eye movements[1]. You don’t notice the movement, the blur, nor the dropped frames. Brain fills it back in.
Another cool one is how we deal with the blind spot in our eyes. Brain fills it in with a continuation of what’s around the blind spot. The area is big enough to hide a whole car or even a large truck at certain angles/distances.
That’s why you have to turn your head when driving, just your eyes isn’t enough. You’ll get stuff flying at you out of your blind spot.
The world models paper directly addresses seeing the world as an approximation [1], though in the other direction of compressing the world into a smaller but workable latent space by autoencoding.
On a tangent, I've experimented a lot with lucid dreaming and dream clarity as well as our recollection of it varies a lot depending on how deep asleep you are. When I'm calm and the dream is stable the visuals are incredibly sharp and detailed. I'm often amazed how lifelike it is.
One issue seems to be that dreams get fuzzy as we start to wake up and our memory of dreams is often seen through a fuzzy lens even if things were clear during the dream itself.
I've especially noticed this last part when being amazed at the detail and describing it to myself as I inspect it. Then I wake up and the memories are hazy and lacking the details I described.
The textures seem glued to screen space rather than moving with the blocks. Would the "alias-free GAN" network (https://news.ycombinator.com/item?id=27606347) allow implementing a similar network without these artifacts?
The music is really sweet too. Its all fuckered up but still on beat and even has a pretty nice pocket to it. Those kicks that happen about every two measure slap
It looks like they need to tweak their Neural network some.
From the video it looks like the reconstruction network is confusing background and walls, which should not happen if they are distinct labels.
I suspect they are passing the pseudocolor labelled images between the two networks instead of the one-hot labeled data that a vision system would normally deal with internally.
Me watching this video makes me think i'll see full, conscious hard AI in my lifetime.
You can never really point on the evolutionary timelines and say, 'That's it! that's when consciousness arose. It's obviously going to be the same with AI.
When AI rises to the point where it is obviously conscious, there's never going to be one key technological breakthrough that put it over the edge. Like the fall of rome or anything complex really, it will be a collection of things and a backwards pointing measure of when it actually occurred.
Jeff hawkins has some extremely interesting thoughts on AI. His book 'On intelligence' is one of the best books i've ever read.
These aren't entirely his thoughts butThree things; some ability to move around your world (e.g. cyberspace, robots), perceive the world (Vision, data,dumps, training data, supervised learning, deep learning) Some mechanism with interacting with the world based on motives and desires.
What happens if a non-sudo'd trading bot, with a wallet starts trading on the blockchain? Seed it with capital, have it access some smart contract, and delete sudo on both the smart contract and wallet.
it's not doing anything too 'sophisticated' but it is a) able to 'Die' by it's wallets balance to hit zero. b.) has a 'motive' to increase it's total coins.
pre-programmed, yes but also is ingesting data from oracles and acting on that information.
there is only one problem, most of these "AI" algorithms don't have an actual understanding of what's going on. These visual pattern matchers can't generalize beyond very narrow meaning of the word.
You can train it to associate certain patterns with certain labels or responses, but it's basically a mechanical puppet that always needs a puppeteer to "train" it. As long as the prevalent view of AI is based on machine learning, using some form of optimization to match patterns, it has no hope of reaching AGI. It's fundamentally backwards looking and bigger models won't make any less "dumb", for a lack of better word.
DL can probably represent lower levels of sensory processing, akin to dimensionality reduction in visual cortex, but it can't generate new abstract concepts from lower level ones, form clusters of such concepts and facilitate navigation in "concept space", in other words go from low level associations to "understanding".
In natural cognition learning & memory are highly intertwined, and acquisition of basic knowledge is not based on some teaching signal or "delta rule" reinforcement, it's more akin to self-organizing hierarchical networks operating via very limited set of computational primitives such as JOIN and LINK, temporal synchronization and innate system of coordinates in abstract concept spaces via something like grid cells https://en.wikipedia.org/wiki/Grid_cell
The constraints nature puts on living organisms, such as limited energy and time define the computational architecture of cognition. As long as these constraints are ignored AI will always be "5 years away".
Why not? This is NOT my area, and I'm most likely VERY wrong. But we get input from our eyes, it goes through all the layers of the brain and comes out as a word in speech.
Could that not be how our mind evolved before getting this 'consciousness layer' on top of it?
Maybe building on this adding more and more abstract layers, training them to return useful info, could nudge things in the right way?
Someone correct me if this is not true, but the current AIs using NNs are "static", in that the model is created offline using tonnes of data, and then at runtime it's simply doing pattern matching / prediction. There's no "plasticity" as seen in human-like intelligence. It seems to me that we need a slightly different approach that will change the NNs during runtime.
Probably not. NNs right now are just glorified pattern matching/curve fitting. Getting to a generalized AI is some serious "draw the rest of the owl" work.
HN discussion: https://news.ycombinator.com/item?id=27559106
You are a Tube: https://www.youtube.com/watch?v=udPY5rQVoW0
GitHub: https://github.com/sentdex/GANTheftAuto/