A.I. Experiments: Visualizing High-Dimensional Space
HTML-код
- Опубликовано: 2 окт 2024
- Check out g.co/aiexperim... to learn more.
This experiment helps visualize what’s happening in machine learning. It allows coders to see and explore their high-dimensional data. The goal is to eventually make this an open-source tool within TensorFlow, so that any coder can use these visualization techniques to explore their data. g.co/aiexperiments
Built by Daniel Smilkov, Fernanda Viégas, Martin Wattenberg, and the Big Picture team at Google.
More resources:
www.tensorflow.org
I was so happy when they said "Open Source"
Everything is open source now scikit-learn, tensorflow, etc.
I remember the days though when open source was considered crazy.
Ani H. Same. I get so depressed when I see a really cool machine learning application only to find out that it’s not available to the public...
@@zbzb-ic1sr Why is it not crazy? How do they monetize open source?
@@watherby29
You don't directly monetize it.
@@watherby29 for example, they open source android but if you want to use Google service to the android you need to pay their service (ex: Samsung, LG, Motorola, etc)
6 years ago I watched this video and thought to my self, "this is so cool, I hope I get to work with this someday." Fast forward 6 years later, I am a research assistant in machine learning, studying my Masters in data science at a big university. Thank you Google AI team, for not only all the research that you have contributed to the machine learning world, but also for exciting our imaginations :)
I'm doing my bachlors
Well done 😮💨👍🏿
💪💪
10:08 PM
2/28/2023
@@happyjohn1656 are you also planing to learn machine learning 🤔
which university?
RUclips's algorithm took me to a video about RUclips's algorithm
It wants you to understand it so that you can prepare for Skynet.
Kind of a 'meet the folks' moment.
this ruined my life
Weirdly impactful realization
Imagine if the RUclips algorithm was open source
This is a so damn interesting work from a small team. And they explained it so well, someone like me who doesn't know a thing about data visualization could grasp the concept behind all of this. -*And it's open source.*
I think you meant "- And it's free to watch."
"open source" would mean in this context that the video editing files were made openly available for download.
well said, in just 3 minutes I understood what was just a blurry concept in my head; how AI understands meaning.
@@markusheimerl8735 I think they meant that Tensorflow is open source
🌟 Them: *"It's open source..."*
Me: 😵. 😮. 😯.
Like for anything related to OPENSOURCE!
I clicked this video thinking it was new and was amazed to see it was made six and a half years ago. Wow! Fantastic work, fellas, keep it up! I'm proud of you 👍
I didn't realize Nikola Tesla and Nikola Tesla were different people. Mind blown.
Nikola Tesla (the inventor) had a nephew who was named after him. Look it up.
What???
Whaat
Interesting how this is only mentioned now, not back when the video was released haha
Time stamp?
Wow! This visualisation blew my mind. I am practising machine learning right now and this visualisation will be very useful to let us have a peek inside the algorithm.
What an incredible visual. Uploaded 7 years ago…I feel so behind. This helped visualize what’s happening so well
This is revolutionary. Even with just the categorization of the words, we could see the extent of our languange and our possible thought patterns.
As I am currently learning data science,
I will make use of this advanced tool not so long from now. Thanks Google team for this.
Developing more complicated Ai is in some weird way like learning about ourselves
Yes. Agreed. I wanted to add if you will, the world we built is a reflection of the human psyche.
@@superheaton yeah! I was about to say that. and soon they may make it or break it on bots, this is no longer impossible with robots because they have now AI, but what if we are also created the same manner as a reflection from the past of a different being.
@RoiF , agreed. The world will got very fast change in near future. Welcome to watch my channel for more info about data visualization. 🙏
Nikola Tesla is written twice, as a different person (0:48).Was that on purpose? XD
Maybe there are two persons. One scientist as we know. 😀
Nice explanation. It all makes sense why you built a tensor accelerating unit now.
A real time web example is music map website (no commercial intended)
The map uses Global Network of Discovery "GNOD" database to sort information with use of this technique.
You can check the website by your own, GNOD have its own website that have some cool features too, as well you can check music map site and discover some new musical artist with actual example of how this method works.
(Again, no commercial or search-bait intended)
to be honest, Artificial Learning is pretty simple at its core. Just that we have been able to build up fast processors and huge memory capacity that this route has become feasible.
Your comment is confusing. First it's either artificial intelligence or machine learning, not artificial learning. Artificial intelligence can be pretty simple I agree, but machine learning is a hard subfield of AI. It can get extremely complex, especially certain subsets like deep neural network.
Once create sample assignment about five years ago those programs in market helps about input dimensional but I think this way make me understand data better.
...This is amazing. It's amazing because the tool can exhibit words that are worth learning and related words. This can be insanely helpful when learning important terms for some field in particular. I can't wait to get my hands on this.
i just started to get into machine learning and this type of videos give me hype :D
Oh, i thought you had begun to simulate high dimensional space in the physical sense of the word. How disappointing.
Yeah working on some physical problem involving a higher dimensional manifold... the title got me really excited on seeing this.
Yeah was hoping the same but not disappointed
Yes
Well we kinda already did. Let's say you have a room jam packed with books... Well, digitalize them and save them onto a 32gb micro memory card. Boom. Physical space acquired thru the use of a higher dimension.
It's interesting because the use of tensors could be may be direct to use geometric algebra or most often called Clifford Algebra
Remember, once Google refused a project releted to improve accuracy of drones and missiles using AI. Now amazon and microsoft took the project
respect to amazon and microsoft
they still own boston dynamics as well as many war related companies and projects, them refusing one project doesn't suddenly make them angel and microsoft and amazon devils, I'd say Microsoft has the best humanist record so far, especially it's founder Bill Gates
We will have to defeat those two bosses in the future, and Google will be our ally to save the world.
The best visualization on how a machine finds correlation. Thank you for this.
i wonder, could you use T-SNE to partially decode the Voiniech manuscript?
wont help if its unique
Mr.BrYcE What do you mena by Unique
Thomas Lowry
i do think we'd need a similar manuscript to compare it to if we are to interpret anything from this arrangement
Mr.BrYcE perhaps but it could help to understand the nature of the text and its language.
it has more or less been proven to be a very old fake.
Nice, that was a very helpful explanation. I feel like I understand machine learning a bit better now.
How so? This was guff IMHO.
I've never seen machine learning carried out and explained with approaches akin to multivariate statistics. This is great!
Much of many comprehensive ML courses revolve around that kind of approach.
Dan Bolser Maybe you're just an idiot?
Jack Johnson possible, but these grinning cheese nuggets are morons. There are better explanations of the approach elsewhere, and it's not that interesting tbh
Open source are my two favourite words in this video
And I really love AI, machine learning, & data visualization
I would love to use this as a wikipedia for everything. Type in a word then learn about all the associations and keep digging deeper.
Awesome work, being able to better intuit machine learning results will lead to much better ML engineering, as well as conceptual breakthroughs.
what a time to be alive!!
For most purposes, it's probably easier to just use PCA/MDS. Unlike t-SNE, a PCA method will give you consistent results.
imagine having this exact same data visualization but now changed as buttons, for the most used/common chatgpt questions, welcome to the future.
this is a great explanation! although i do have a quick question, so if we treat a pixel as a dimension doesnt that limit the usage of this technique to simple images? since a more complex image would have many more pixels and thus the dimensions would be massive, wouldn't this hinder the processing time and cluster formation, aka it would take forever for it to form clusters, just a question, if anyone with more experience with ML can chime that would be awesome!
I would love to teach machines how to build! What will they one day teach us?
Hey, can you just hang on with this thing? I translate documents for a living and this is pretty scary. Please stop it, Google!
mastersoftoday Sorry bro
mastersoftoday actually they would pay you to make sure the translator is translating correctly and if it can do it's job well enough. expect matinace and double checking document's that was made by a machines.
maybe you should write code instead and automate your job :P
+blind1337nedm that's a good idea probably what will happen lol.
+mastersoftoday, we all have to come to grips that most jobs (including even creative work) will very soon be much more efficiently done by machines.
I suggest to read the following article series to gain a better understanding: waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html
Wow! Data Visualization is super amazing and scary at the same time.
Small Focused Teams are most likely to do the Best Work!
Great work. I was looking for this.
Which software are you using? I downloaded t-sne for python but ended up with a tkinter plot.. How to get that interactive plot?
you need tensorboard (included in tensorflow) and no need to compute t-sne (or pca) yourself.
Abhishek Bera I have no idea what t-sne or tensorflow are lol, I'm learning a bit of Python now, if I wanted to get started with AI and machine learning, what do I need? And what are the common tools and software you need for it?
+nishad nadkarni Do yourself a favor and start with tf learn python library. It's just a tensorflow library but provides more higher level apis on top of it. So its very beginner friendly anfter that try to use Keras a little deeper and after that go for tensorflow/CNTK
Look up Saraj Raval
siraj is click bait
I really want to learn more about it!!!
If each data point is 200 dimensions how do we represent each data point in 3d space as shown in the demonstrations?
I don't think the placement in 3D space is based on the individual "dimensions" more so just a way to visualize it. It's just grouping similar things in what looks like just a sphere. Think of it more like a 3D Venn diagram, the video is more a representation of how AI can find the similarities in an input, not necessarily the visualization of them, just makes it easier for a human to see what's going on, you could as easily export the output to a document, just harder to extrapolate data from a human perspective this way.
Thanks for making it open source, It will very useful for my project which I had a similar idea.
Nice video. How do you visualize a high dimensional space in the three dimensions that are shown here? Is a type of compression taking place?
Thanks I needed this to transfer my big data
So that's how text recognition works.
Well next gen text recognition... its still not out to public
+Vignesh What part of this isn't public? tSNE has been around for a decade or more.
Vignesh, the first perceptron was written in the 50's and I learnt how to code a text recognizer in my undergraduate in the 80's... the difference today, are that AI and ML is becoming mainstream... so there's more people producing algos, replacing 'old' ways of doing things, such as the activation function 'sigmoid'. There's 'no secrets', behind AI there's just maths , specially tensor (multivariate) calculus, linear and dynamic programming (which are areas of operations research, aka maths) etc...and loads of programming frameworks in python, java, js...
This video explained one tiny component in a very simple way as far as I can tell.
Would be awesome if you had a sort of "3D" search engine where you could search the words and you could surf around them or something.
I'm born in 1995, and Im sure that in my lifetime there will be A.I that can be compared to human.
Gytis321 S Me too, and I feel the same way
go check out exurb1a. He literally makes amazing videos discussing different possible futures involving A.I. and touches on different ideas around life and consciousness. Also, in one video, he says how in the future people might become one giant conscious computer that conquers worlds. It's pretty cool.
I recommend you to read this book: "Superintelligence: Paths, Dangers, Strategies" by Nick Bostrom.
Do you think AI can be emotional?
See if "Future" and "Half-Life 3" are grouped together.
The SkyNet.Exe. TSNE
I always wondered introducing randomness into a series of potential images or patterns in the way you have could be used as a means of thinking ahead.
Some basic stuff. Well presented, though
it's a great tool, as there were no satisfactory higher dimension feature validation techniques before. this is more intuitive and gives a clear visualization of how good your features are!!
Can we use this on our dataset?
these features are open source, so yes
donot call them multi dimensional. because those clusters are nothing to do with dimensions. instead define this multi layered op as multi brane app.
Keeper Raziel Everyone of those datapoints is a multidimensional vector. For example the numbers are probably 784 (or more) dimensional vectors with every dimension representing the alpha value of a pixel. In my opinion it's pretty accurate to call them multidimensional
But the alpha of a pixel just a scalar value...
Im a physicist and this is not right. The 4th dimension (Quaternion: ±i,±j,±k & w0) is all what you need to explain a system like this. The color value of a pixel is not a dimension by itself, but can described as 2 dimensional scalar field. The pixel together with its color value is quantumphysical a mixed state and nothing else. So, see the pixel position & color position vector like a clock. You got one positionvector for the location of the pixel and one positionvector for the color value on the scalar field, together they represent the whole data. Thats possible, because the second dimension can be embedded as topological layer inside the fifth dimension.
en.wikipedia.org/wiki/Bloch_sphere
en.wikipedia.org/wiki/Quantum_state#Mixed_states
en.wikipedia.org/wiki/Quadrature_amplitude_modulation
I can definetly see my computer exploding when plotting this
Hmm interesting Google and thank you for sharing.
Great approach!! Congratulations.
That's amazing! But what about sentence structures? They are quite different for languages around the world.
Thank you and congrats!
So no ones talking about Jarvis here, its trippy...
Jarvis: yes
It's so much exciting and so amazing....I love it...and I'm pursuing it..
Thank you so much
August 29th, 2019 - Skynet becomes self-aware.
Oh boiii thats gonna be fun
Let's play a high dimension game, what's 13% and also 50%!
Hehe
What's 4% and also 67%?
thanks for the video, guys you mentioned 200 d space, but you show words as points in 3d space.. how you reduced the 197d ?, any info will be appreciated,
But how do you project this high dimensions down to a 3d Visual model?
Exactly my thoughts.
Can you explain to what degree or level of
dimension we are viewing this at?
Não tem como não reconhecer esse sotaque. Kkkkkkkk
hi we are building an app using Google AR Core.. as our potential users are mostly middle class and lower middle class.. we are kind of in dilemma about the initiative.. as in Google AR Core is still not supported in several smart phone models.. I would like to know that whether you guys keep on adding the AR Core feature to all models.. irrespective of the origin of the brand's.. please let me know the interval of adding Google AR Core in maximum number of devices..
And im here . Still learn how to create CRUD with no reload :(
BuBadiBaKa hahaha. What lil Wayne say? Repetition(and the study of the theory/math)..added that extra part lol. You will be there one day. I believe in you. Stay blessed.
use jquery ajax and axios bro. hahahahaha
have fun
Angular 7 is here for you, don't worry
Gyus, you are awesome! Develope it! This is great!!!
and eventually,....chatGPT
wanted to comment my super excitment after watching this video, but... NO WORDS!!!!! AMAZING!
SkyNet is coming soon, pals...
Fandyus CZ Stop it, it's old, give it a rest.
But it is kinda true.
Whenever I have seen a video on AI it is there at least once. It would be more correct to say Watson and Sophia are coming. The military is not going to take the human out of the loop when it comes to killing.
Fandyus Overused
*So, yeah i'm 2 years late but, SO, you made a brain of things*
NIIICEEEE
I was like "ah glad to live in the future" but then it said open source..
well.
What is this supposed to mean?
What are you trying to say here? Open source is good.
Thanks I needed this to transfer my big data
Well... This was clickbait
MNIST: The "Hello, world!" of machine learning.
I see a big risk in the following topic: us humans have spread so much data about ourselves on the internet. What if there's a bunch of neuronal networks that can someday categorize people into "psychologically stable people" and "psychologically unstable people" or so? Why should we trust algorithms that can make mistakes?
Truly Outrageous That's exactly what IBM Watson can do now, analysing big data to make sense of it, what you just postulated is entirely possible, but it's doesn't mean anyone is going to be culled for it.
If there is something you can abuse, it will be abused. That's how politics works. That's why it is so dangerous.
I mean, that's more of an issue with education, and teaching people to know their sources, etc.so that when one person abuses something like this, the masses will ignore it.
right now *people* already categorize people into "psychologically stable " or "unstable". so even if a machine is not perfect and can make mistakes, if it makes less mistakes than the people that are currently doing it then that would still be an upgrade, right?
We have been performing such categorization for a very long time, this is merely another tool in the shed for intelligence agencies. The status quo will remain regardless of AI's success, so don't worry (but feel free to be mad at said status quo).
The question of mistakes is an interesting one however. If you take imperfect test, even one with a very low false positive and false negative rate, and use it on a large group of samples (people in this case) with a low incidence rate (psychological instability, which is rarely detrimental to society), the false positives will flood the true positives in the test results, making the chance of a true positive lower than expected in the group of people who tested positive. en.wikipedia.org/wiki/False_positive_paradox
Imagine this but with all the data collected from Google thru android sensors.
In minute 0.35 and 0.49 Nikola Tesla is repeated but his "dimentions" are different 😅😅
Good to remember Nikola Tesla, but it would be more great when People out there will understand the work of Nikola Tesla. He has the answer for the Quantumproblems in Physics, but look at our age. A whole planet full of naked apes and their goal is just to spread over the complete living room like a virus.
Who ever this understand, understands what Nikola Tesla means with the sentence: "If you only knew the magnificence of the 3, 6 and 9, then you would have the key to the universe."
www.intmath.com/blog/wp-content/images/2016/06/tesla-map-to-multiplication.jpg
When Anton Zeilinger would use another number range would he told us the same numbers. :)
this is the tech version of how gypsies used to read people's palm and tell them about the future.....science is so far behind
sne and t-sne are not the only methods ( and have their own drawbacks, as every method)...
a useful tool in a tool box
noch much, not less
And each time we run t-sne it shows different results
@@anushkachathuranga8943 right, and that's NOT surprising! it's stochastic neighbouring, so there is a kind of part of ' randomness' ( which you don't have with a pca for instance, cause it's only the computation of the eigenvalues and eigenvectors)
6 years ago...why didn't I watched this before
Google can do this, but they still can't fix the RUclips algorithm.
SuperStriker7US Team Name:The Last Bosses they sure can, but they won't. Everything they do is in favour for money.
This is basically how the related videos system works on RUclips.
@@Ownage4lif31 without money, your fav youtube won't get paid or must paid to use the platform.
@Chawza Dark There's millions of ways they can monetize it.
What's wrong with it?
Mind blowing actually my mind is blower away,this is really complicated.
Really nice video! btw Marie Skłodowska-Curie...
this would be good for google searches because when you look for something it would find it more accurately
That is an abuse of the words and the concept. This title and description suggest that the AI/Machine Learning system is visualizing higher dimensional space, which also implicitly suggest a type of understanding. Even the narrators state that it is finding a meaning of the words based on their context. This couldn't be further from the truth, as the word "meaning" implies understanding. When speaking in scientific contexts, especially in a subject as Artificial Intelligence, we cannot throw the word "meaning" haphazardly or metaphorically.
There is no meaning understood or discerned. What we are seeing is a graphing of data points, nothing more. It is a crude, though respectable, beginning, at trying to relating data to one another. Let us not misuse the language because it will lead to false perception and expectation on part of the lay audience at large when that's the last thing we need.
Nour Douchi However, the machine learning is acting upon a set of defined drives to identify and assess value in information. The directed activities find meaning according to the laid out requirements of the code (contextualizing the machine as a proactive actor, as it can act independently) and by the people directing it (contextualizing the machine as a passive tool). I've studied linguistics and I honestly don't have an issue with their use of the word, "meaning". Regardless, "meaning" already has a broader definition than what you prescribe, so unfortunately for you, your language is already tainted. Then again, English is a hodgepodge of taint.
visualization in ML means something different than you think.
This is amazing
High dimension space??? 😂😂😂😂in India we call it jatakam
powerful open source software is the future.
Pi: 3,14156....... Really Beautiful Video !!! I hope it will help a lot of people to OPEN their MIND and SOUL !!!
Pascal Gula And Soul?
you don't have any? :)
that's just 6.283 18 53 07 17958 64 72 halved. no big deal
of course not !!! :D
but 2xPi ~== 6,283185307179586477
What if they reverse those dimensional spaces to words...like converting a neural network to sentence.. galaxy or constellation to words..
Would I be correct that the dimensional space I'm seeing here -- the closer they are altogether, like venn diagram -- the chance of "erroneous" or "correctness" for likeness? I suppose, the further away they are, the most contrast they are to each other.
these are represented by feature vectors, and they are in the "feature space"
actually you can see this just like 2D xy space or 3D xyz space, in 2D, the feature is p=[x,y], and you use the distance between points to see their "similarity"
while in high dim, the core idea is the same. the feature may be f=[x1,x2,x3,x4,...]. you also need to see which point is closer to which point
It they are closer, it maybe they are similar in the feature space. (classification)
so based on the above
if the samples are indeed in the same category, what we want is that they should also be closer to each other in the feature space.
So how to design the feature, and use what compute rules for distance are the most important things here
e.g.
apple = [price=3, color=red, shape=round, weight=3, sweet=3].
orange = [price=4, color=orange, shape=round, weight=4,sweet=4].
ball = [price=3,color=red,shape=round, weight=3, sweet=0].
these are 5 dimension of the feature of apple, orange and ball
but how to classify the 3 things? by computer?
actually we could put these 3 things in the feature space (price, color, shape, weight, sweet)
if they are closer by distance, they maybe in the same category
In this case, we want apple and orange to be closer, at least they are fruits
And if we want apple and ball to be closer, we are doing another classification.
One example is that "what we could buy if we spend 3 dollar" or whatever
So the distance is mainly decided by what your categories are.
This is fantastic work!!!! I was brainstorming some concepts today and I came across this video and it just nailed it for me!!!! Amazing work Google guys and gal!!!! Kudos to you.
0:13
wow what is that
my eyes are hurting now
partly because of that
but mostly coz i play games too much
beyond interesting. I wonder what would happen if we fed all existing data of literally everything into one of these...
How about visualizing Quranic Verses or Arabic language using this technology ?
Cause Muslims generally tend to claim that arabic has quite deeper meanings and links within words and sentences.
@Billy Casper: Hmmm -- good application to consider. I'm trying to think how well that would work . . . and can't decide without actually seeing it at work. For example, does it only move the problem of meaning from one level to another (?) That's probably an improvement, but is it enough (?)
Erm this is incredible, how does it do that? Look the words surrounding other words and group them if the concept seems similar?