Play the puzzle in your browser | The source code is in the description section hunar4321.github.io/top-down-prediction/
@captain_crunk
Жыл бұрын
This channel is criminally under-subscribed. With fantastic content like this, it's only a matter of time until you make it big.
@NightmareCourtPictures
Жыл бұрын
So true. I had a lot of fun playing the testing game too.
@vin2368
8 ай бұрын
This is the best video about consciousness 🤯
@heidolf6002
Жыл бұрын
Every video of you just has such high information quality. I feel like I always leave with some profound new understanding. Truly great work!
@ai_is_a_great_place
Жыл бұрын
I'm subbed but didn't see this until I saw your community post - I know the algorithm will pick you up big time soon!
@newenglandbarbell4647
Жыл бұрын
Amazing, very well articulated 👏 thank you Hunar, and happy new year to you also.
@youchuanwang1738
Жыл бұрын
Tonight, we hunt 😂😂😂😂 just the lion image makes me wanna quote bloodborn. Good video btw, I never thought that point mutation is worse compared to some other operators with memory supported.
@jonrhaider
Жыл бұрын
Another amazing video, thanks Hunar. Can't wait (again) for the next one.
@spagussy
Жыл бұрын
my guy hunar only releases bangers
@bezhanmohamad3881
Жыл бұрын
Proud of u dktor hunar, Supas ka bebashman nakait lawai fery abit
@lizardy2867
Жыл бұрын
Acceptance is one kind of prediction, but for our ego. Ego in this sense is a barrier to entry we apply to our experiences, a high level concept developed in times of high uncertainty and risk. In cases of effective ego, you can imagine it as a curve which tends towards a climax near adolescence, and dips back down to a plateau in adulthood.
@parametric327
Жыл бұрын
I love your videos!! I felt human brain is CPU that learns Why we are trying to reach global maxima is someone calculating something to get answer? I recently feel this universe is someones brain
@decton4461
Жыл бұрын
Your videos are so exciting to me, keep going
@eazypeazy8559
Жыл бұрын
interesting. so consciousness is a mechanism to: 1) analyze information from scanners of the outside world; 2) build enough effective models to predict the outside world patterns? and as soon as these models become effective, this activity becomes "automated" and unconsciousness, freeing the consciousness to concentrate on other problems? and that is how we can solve the problems without brute forcing every possible solution variation like in pure genetic organic evolution world? and actually a live, as subjective feeling of being here right now, is a set of concentration on scanners data moments and building the models? maybe, that is why new experience making to feel you so "alive"... I was ice skating today for the first time, I remember the subjectively high consciousness feeling of being present and remember a lot of details, the learning process itself. From this point of view, I should ice skating more often to be more alive, until I will skate well enough :) but how can you be sure, that information from the outside world is correct enough? maybe your scanners are corrupted and you are getting wrong feedback? and I guess, that human consciousness due to its organic nature has predefined set of problems to solve? because all humans generally want the same from their life, right? each of us can understand a lot of other's motivation for happiness, I think.
@brainxyz
Жыл бұрын
but how can you be sure, that information from the outside world is correct enough? maybe your scanners are corrupted and you are getting wrong feedback? By correct information, If you mean a local truth, with repeated measures our top down process can learn to correct the corrupt sensors like the corrupt text I showed in the video. However, if you where asking whether is an absolute ground truth that everyone agrees on, then this is a tough question and the answer depends on whether the world is deterministic or non-deterministic. If the world is deterministic, then there is only one ground truth (i.e there is only one timeline). All the other alternative timelines are the illusion of our conscious mind. In a deterministic reality the past events can be seen as the ground truth despite the noise of the measurement in the long term one can still find the truthful patterns from the past with repeated measures. In a non-deterministic world, it is hard to figure out what is the ground truth because an objective reality might not exist. The Copenhagen interpretation of quantum mechanics puts too much weight on the observer (the measurement). If a tree falls and no observer (measured) detected the fall, then this event is not registered in reality. In this non-deterministic observer-centric world, the truth could be decided on like a blockchain mechanism where a collective process decide on what timeline should be our next truth. Tiny random events can affect the entire process... It's confusing. If objective reality doesn't exist it's hard to define the absolute ground truth. Elon musk has an alternative Occam's razor, he says the most interesting event is probably most likely be true.
@eazypeazy8559
Жыл бұрын
@@brainxyz thanks
@gabrielschoene5725
Жыл бұрын
Am I the only one hearing the doubletrack at 19:22 Fascinating video btw
@ceilingfun2182
Ай бұрын
I have goosebumps
@JaredFrontman
Жыл бұрын
As an AI researcher myself I truely agree that backprop is one of the most inefficient algorithms. Its not how brain works. Our brain is not a bunch of stupid neurons, but rather an organized set of neurons. Different parts of our brain is specialized for different tasks. I have been on this hunt for First Conscious Algorithm to pass Turing Test for 3 years. I'm really interested to know how your PHUN Algorithm works in detail. I am well accustomed with AI and their several algos. Therefore you may explain it as you like.
@brainxyz
Жыл бұрын
Thanks! I agree there are many more differences between our brain and the current artificial neural networks. I didn't have time to mention all of them in this video. Regarding PHUN Algorithm, It's an incomplete algorithm and a work in progress. It's like a glorified N-gram system with a more sophisticated back and forth connections. It's fast at learning but bad at generalization. Transformers are much better at generalization. I'm still researching what is missing and I'll talk about it as soon as I get a better picture. Regards
@JaredFrontman
Жыл бұрын
@@brainxyz Yes! Some years back I had developed an android app(i may share the apk if u want) that uses modified N Gram algorithm to do some text generation(i didn't knew n gram algo at that time, i had to literally invent it with my limited knowledge at that time). And when I tried your Braifun app, I had a feeling it's using some clever N Gram techniques. And now that you mentioned the N Gram, I can provide some additional modification that you may like. Now the thing is, N Gram with N=3 works best. It is because it's able to capture phrases and makes sentences more meaningful. Moreover, we can use reinforcement. N Gram can learn any sequential data. But it needs a sense of which data to come next. I implemented Phrase Vector Prediction( i made it up). What Phase Vector Prediction does is that it assigns a random number to each phrase. This number represents the meaning of the phrase to the computer. Now whenever a sentence is generated, the user can give the computer -ve or +ve feedback. Upon -ve(for ex: -6) feedback, random vectors of respective tokens from the generated sentence are updated based on the feedback( i.e. vector[i]-=6). This creates a dictionary for the computer to use, which can be later used for creative algorithms. Now, as you said, Transformers are good at generalization. I agree. But transformers still are not the best. But, transformers are good only because they are trained with lots of data. You see, ChatGPT can do a lot of things and can even solve math problems. But ChatGPT is basically A highly modified ensembled N Gram algorithm. Basically a probabilistic Markhov Chain, that can quantize relations between tokens in sentences. If you ask it what's 2+3. It will answer 5. But it cannot calculate. The answer 5 is from the dataset. A true NLG should be able to calculate as well as understand sentences. The problem is, our brain do not represent the data the way computer does. For example, image is represented as pixels. Our brain sees the as pixels but the optic nerves converts then into the language of brain. Every sentence, sound, touch is also converted into this langauge. This allows more flexibility in generalization. Moreover, electrons travel 1.5 million times faster than neuron signals. Therefore, a real AI algorithm should be able to learn and think much much faster. Your top down and bottom up approach is very plausible. And I warn you, please do not use mere neuroscience or psychology to unravel how brain works. I have tried to do that and always ended up with the neural network algo. The entire neuroscience revolves around neural network. Wasted almost a year on that. Moreover, the functioning of neurons do not reflect how we think. Take this short example: A computer can do ray tracing. But you don't know how ray tracing is done. So, you look at the architecture of microprocessor to understand how ray tracing is done. Do you think you'll ever be able to learn ray tracing that way? Absolutely not. "It's Not the Architecture that matters. It's the code that matters" Similarly, neurons work in strange counter-intuitive ways. But simulating a bunch of neurons can never achieve intelligence. We need to know the "collective code" behind these 1000s of neurons before we can find the true AGI algorithm. But we don't the "single code" behind 1 neuron, it's meaningless. (thank you for giving time reading this. i understand you're a busy person)
@brainxyz
Жыл бұрын
@@JaredFrontman Thanks for the detailed comment and for sharing your ideas. I agree with most of what you said but N-grams are not the answer to be quite honest. Higher N gives more meaningful sentences but requires a very large memory and less capable to generalize than an equivalent neural network (in terms of the number of parameters). The good thing about N-grams is that it's easy and fast to train them. I guess the better solution if one combines the best of both worlds. There are still some easy toy problems that both N-grams and Neural nets fail to learn while our brain can learn them easily. I'll try to explain them in future videos once I have a more coherent thought process about these topics. Keep in touch.
@kusaedonai
Жыл бұрын
i just spent a while playing with the puzzle game, i tend to keep getting a score somewhere between 12 and 17. I wish i could optimize it better lol
@Gigasharik5
Жыл бұрын
Those periodic sine waves patterns seems to be fundamental core of processing information in brains, even in v1 orientational columns behave like that (youtube deletes links) so heres the name of the video that clearly shows it "Relationships between orientation-preference pinwheels"
@samir7838
Жыл бұрын
I am confused when you were talking about how machine learning algorithms only have bottom up thinking methods. How so? What implications does the content in this video have for AI?
@brainxyz
Жыл бұрын
Current machine learning methods are only feed-forward systems (especially during inference). This makes them inefficient. Most machine learning experts agree that today's methods are inefficient sample wise compared to the human brain. From neuroscience we do have a crude idea on how and why our brain is more efficient (some points explained in the video). However, the intricate details on how do we integrate top down with the bottom up during learning novel things is still under research.
@samir7838
Жыл бұрын
Truly a beautiful video. Keep this up and you will explode in popularity soon.
@asdf8asdf8asdf8asdf
Жыл бұрын
Is this closely aligned with Tversky and Kahneman’s book “Think fast, Think slow”?
@brainxyz
Жыл бұрын
Yes, it's indirectly related. This video is more about the Predictive Coding theories in Neuroscience.
@FrancuzMinecraft
Жыл бұрын
Nice watch
@karakson
Жыл бұрын
Thank you very much for this essential perspective too! Awesome!
@coolsai
Жыл бұрын
great video 🎉
@vantahku7211
Жыл бұрын
Fascinating content. Thanks and stay safe.
@Game_Lab_Germany
Жыл бұрын
I have to report a bug to KZitem there are 2 or 3 missing zeros at the MUST HAVE subscriber number. Now i doubt my existence, again, after every of your videos.
@genericname2284
Жыл бұрын
Lmao I did this in school, just kept failing until I knew the right answers based off of my initial score. Im glad somebody actually did the math on this
@dante7228
Жыл бұрын
Awesome content, unfortunately I 'll have to rewatch it a couple of times to really understand it. Allow me to suggest you to make some pauses after you make your points so the information can sink in. The speed of your explanation is to high for some non native speakers like me. Also I' m not a programmer so I 'm not familiar with this kind of topics.
@smbdtexter1044
Жыл бұрын
I'm leaving this comment in 2023.05.20 and predicting that once your video will blow up to millions of views.
@Enju-Aihara
Жыл бұрын
more scary concepts that opens the brain pls
@01k
Жыл бұрын
Thank you for sharing
@HemnAhmedhm
Жыл бұрын
Well done ,Brilliant!!
@Leao_da_Montanha
Жыл бұрын
Always bringing great insights
@SebastianSchwank
Жыл бұрын
Is this the same problem as reaching a point in higher Dimensions ?
@SebastianSchwank
Жыл бұрын
I think this is easier to think about for solving efficently.
@brainxyz
Жыл бұрын
Possibly. But this one is more related to finding a hidden pattern or finding a needle in a haystack
@SebastianSchwank
Жыл бұрын
If you have a continues room it's another Problem i think, if the room is discret it's the same because the pattern we search could be seen as a point in this space when every part of the pattern is seen as axis.
@SebastianSchwank
Жыл бұрын
In fact i've recognized it's the inverse of the classification problem, i guess.
@ProgressScience888
4 ай бұрын
super underratted
@freepythoncode
Жыл бұрын
We can solve something like captcha by generate random captcha images and make compare between generated captcha and captcha image but this is not good idea this need super computers
@parametric327
Жыл бұрын
can you make a video about psychedelics?
@brainxyz
Жыл бұрын
I thought about putting psychedelics in the context of this video but the mode of action depends on the type of the psychedelic and also many aspects are still unknown. I haven't experienced any psychedelics myself but I think the hallucinations are due its effects on the top down prediction. The common repetitive patterns start to appear in your mind and this is why you experience the view of fractal like shapes.
@parametric327
Жыл бұрын
@@brainxyz that's very interesting! I want to research about psychoactive substances. I feel like psychedelics makes mind more liberal and feels more limitless do you think there's global maxima and local maxima in brain? everyday thinking is maybe just tip of iceberg of conciousness and it's local and other than that is global like our brain cells and galaxy looks similar. tell me your perspective!
@brainxyz
Жыл бұрын
@@parametric327 That is plausible but I'm not sure about the broader effects of psychedelics. You have to take their effects as a whole package. Their addictive & other side effects may bring more harm than use. And yes there is local and global maxima in the brain. Our mid-brain (also called lizard brain) stores our most basic instincts so it usually prefers to move toward the short gains and tend to get stuck on local maxima. On the other hand, our frontal brain tends more to move toward the longer gain directions and has less tendency to get stuck on local maxima. By the way, that doesn't mean local maxima is always bad. In fact, in dangerous or jungle like environments our most basic instincts are probably most useful for survival. Think about the example I gave in this video and also my liberal v conservative video, if there is a fast raising flood, if you move toward the far away global maxima you are probably dead before you get there. I mean seeking longer gains are only fruitful in safer environments. That is why at the population level we usually see people with different risk taking tendencies some more liberal than other. This can ensure the survival of the species and at the same time finding better directions in the long run due to having risk taking people. At individual level, we have the two parts in our brain too.
@mynameusedtobelong
Жыл бұрын
The first test remeber me of hamming code. That we can do in 6 tries.
@LaurentLaborde
Жыл бұрын
nice video, thx :)
@brainxyz
Жыл бұрын
glad you liked it 🙂
@flor.7797
Жыл бұрын
My adhd brain can maximum remember 3 things at a time 😅
@karokamal
Жыл бұрын
Subtitle for the videos are necessary
@brainxyz
Жыл бұрын
added
@karokamal
Жыл бұрын
@@brainxyz we are proud of you doctor Thank you for your service
@Artista1010
Жыл бұрын
Who are u man? Are u human? Damn how do you figure out all this thing? Make video on now you think? ... It is abnormal..
@settingfeedback321
Ай бұрын
You didn't even address cconciousness once
@ceilingfun2182
27 күн бұрын
Consciousness is like poetry, it only makes sense how you wanted to make sense to you.
Пікірлер: 64