goupaz.com
goupaz.com copied to clipboard
goupaz
[06062021] - Chapter04 Session02 - LDA Topic Modeling: Techniques and AI Models!
SignUp: https://forms.gle/aeiZ9G13NoW2x3tH7
To join Slack: https://join.slack.com/t/goupaz/shared_invite/zt-oe4q61w4-qR8JTxAsWYUk4H2fq8G5pA
Twitter thread: https://twitter.com/sakows/status/1401600780646313984
Linkedin threads: https://www.linkedin.com/feed/update/urn%3Ali%3Aactivity%3A6807354558667345920/ https://www.linkedin.com/feed/update/urn:li:activity:6807354945965305856
Hi What are your views on the statement "Machines can take over humans." Is it completely true or partially?
Is there any ways/best practices to decrease some training time of models? If yes then how?And what algo should be used for the music generator. some resources for it
what is the most promising field in deep learning for beginners? NLP, computer vision, multi-agent systems, reinforcement learning or etc
Have you tried LDA with the combination of word embedding logic to get the semantic and syntactic relationship between words?
What are the key indicators that tell us LDA is the technique to be used ?
What recommendations would you give Students who haven’t started looking into AI yet? ☺
What are your views on the statement "Machines can take over humans." Is it completely true or partially?
Sorry I was busy last week, but your question is particualarly interesting and I have a strong opinion. I don't like movies that show AI acting human, and changing their minds to defeat humans. They don't have a mind. But our definition of "mind" is not clear. So it comes down to 2 questions: (1) will a computer make key decisions instead of a human, and (2) can a computer pass the Turing Test?
The Turing Test is easier to answer because it's well defined. I bet they will soon because in 2018 they became close.
Do they make key decisions "instead of" a human? Yes. The first one I studied was using DNA evidence in the courts in the US. They use HMM as a modeling technique, same as we used for speech recognition in 1997. It often makes mistakes. If a human says one thing but the DNA test says the other, then the jury usually decides with the DNA test because it's a computer. I don't like that.
But now for autonomous vehicles, AI does make many of the decisions instead of the driver. And it has been shown to have a lower error rate, so lives are saved, overall. And, the kinds of errors it makes are different from what a human driver would make, so combining the human and the AI gives a good result. So I do like that.
A lot of work is now being done on detecting mental illness via NN-based systems. I did that in 2018. Error rate is high. But human judgement of mental illness is also very high. So here is the question: if the AI can do a little bit better than a human (higher accuracy) should it? Or not?
So to answer your question: it's partially true, but getting more and more.
Is there any ways/best practices to decrease some training time of models?
Hardware acceleration by GPU or, better, TPU. Intel made a very convenient hardware accelerator that took the form of a USB stick, but looks like they discontinued selling it. On the algorithm level, try using a simpler activation function like RELU instead of Sigmoid. If you have a lot of data, then throughput becomes a speed bottleneck, so get a faster disk
And what algo should be used for the music generator. some resources for it
Anna Huang is the expert. See her paper and software at https://research.google/people/105787/
How can we start contributing to projects, as a beginner?
Sounds like you're especially interested in NLP, so start with tomotopy. (Topic Modelling Toolkit for Python). I'm using it, and it has some annoying bugs. It has several different kinds of Topic Models. Look at the issues list in [1] and see if you can recreate the problem and fix it. A simple one that annoys me is that it crashes if the input token list is empty.
[1] https://github.com/bab2min/tomotopy/issues
Start with a really simple one like #120 which is an error in the documentation.
Thanks alot for such a detailed answer Sir. The facts you mentioned made my thinking more clear.
On Sat, Jun 12, 2021, 11:38 PM Ben Reaves @.***> wrote:
What are your views on the statement "Machines can take over humans." Is it completely true or partially?
Sorry I was busy last week, but your question is particualarly interesting and I have a strong opinion. I don't like movies that show AI acting human, and changing their minds to defeat humans. They don't have a mind. But our definition of "mind" is not clear. So it comes down to 2 questions: (1) will a computer make key decisions instead of a human, and (2) can a computer pass the Turing Test?
The Turing Test is easier to answer because it's well defined. I bet they will soon because in 2018 they became close.
Do they make key decisions "instead of" a human? Yes. The first one I studied was using DNA evidence in the courts in the US. They use HMM as a modeling technique, same as we used for speech recognition in 1997. It often makes mistakes. If a human says one thing but the DNA test says the other, then the jury usually decides with the DNA test because it's a computer. I don't like that.
But now for autonomous vehicles, AI does make many of the decisions instead of the driver. And it has been shown to have a lower error rate, so lives are saved, overall. And, the kinds of errors it makes are different from what a human driver would make, so combining the human and the AI gives a good result. So I do like that.
A lot of work is now being done on detecting mental illness via NN-based systems. I did that in 2018. Error rate is high. But human judgement of mental illness is also very high. So here is the question: if the AI can do a little bit better than a human (higher accuracy) should it? Or not?
So to answer your question: it's partially true, but getting more and more.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/goupaz/goupaz.com/issues/217#issuecomment-860089032, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKOYXNFDMNU4A5WBAX2GQO3TSOPB3ANCNFSM46DBG7GA .