Zero-shot. I just used to think of this as useful for class discovery using a certainty metric or for soft-decision coding as in a comms milieu. Interesting to consider auxiliary info as a descriptor - never did that. Very well-organized and presented. Thanks.
Is it correct to consider speech recordings or their corresponding transcripts (text derived from audio files) as time series data? Do you think they could benefit from time series analysis after tokenization?
The "time" or ordering dimension in a text transformer is typically added to the semantic embedding using sinusoidal encoding of the token position. So basically packaging up the position information in the token. But it would be interesting to see what additional information and relationships that can be gained from embedding the actual chronological time (breaks, pauses, stutters, hesitations)
This is actually just in time for me, I've been building a stock trading bot and have been exploring alternative transformer architectures that are more applicable to time series. 😁
Wont work, it is a stochastic process. Algotraders were the death knell of trading on technicals for retail. Unless you institutional, you simply don't have access to the signals the big boys do. Best you can do is pattern detection to try sniff out the smart money and close follow that.
The market is completely rigged. Literally the Market makers get all the trades and can move the market any way they want against any moron who puts in a large bet. The house ALWAYS wins. Be prepared to lose everything in the long run. You cannot be them when they have full control. The prices and volumes are fake as is the news to justify the movements. The house is not going to go bankrupt. If there was a way to beat the house then they would eventually have to shut down. Literally there are no consequences for cheating. The SEC fines the cheaters a small fraction of what they steal. At best you can skim a little off the top as the algorithm works to undermine larger pots. E.g., if you throw in a big order thinking you got a winning hand 100% of the time the price will move against you to make you lose. The options all dictate how the price will move against you. It's designed literally to maximize profits. Their algorithm will always beat yours 100% of the time in the long run. The market exists not as an investment opportunity but as a rigged game to fleece the morons that play it. If you want, you can learn the hard way. All you have to do is go look at all the pump and dumps in the stock market to see how rigged it is. Crypto is 100x worse but basically same people. As they steal all the $$$ it becomes more condensed and chaotic.
@SlappyZA idk if it won't work, I've been reading several papers of people trying different architectures for stock predictions and I was getting decent results when I was training my earlier models for price prediction. Either way it's a side project of mine and whether it works or not I'm having fun and learning a lot Thanks for the input!
Would be nice to give an intro to VQVAE. Would also be nice to see some real examples including failures, rather than just theory. Where do the tokens for sensors come from? Is an FFT an efficient encoder? How do the code books scale over different time scales or do they only work well over amplitude scales? Nice, well rehearsed presentation!
![Felix "Unfair" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/Oswujxm2Ag0/mqdefault.jpg)
Great explanation and visuals. Sabera you are a terrific communicator!
congratulations on your presentation! was great to see this all layed out! good luck!
😁👍👍💪💪🥳🍌
Zero-shot. I just used to think of this as useful for class discovery using a certainty metric or for soft-decision coding as in a comms milieu. Interesting to consider auxiliary info as a descriptor - never did that. Very well-organized and presented. Thanks.
very cool talk, thanks for making
Justo lo que buscaba! Thank you so much ;)
this is so cool, I hope I have to time to try this out
PERFECT!!
Nice presentation. I'm just wondering how it deals with shifted and stretched signals.
Is it correct to consider speech recordings or their corresponding transcripts (text derived from audio files) as time series data? Do you think they could benefit from time series analysis after tokenization?
The "time" or ordering dimension in a text transformer is typically added to the semantic embedding using sinusoidal encoding of the token position. So basically packaging up the position information in the token. But it would be interesting to see what additional information and relationships that can be gained from embedding the actual chronological time (breaks, pauses, stutters, hesitations)
@ You are right, that would be interesting to investigate. Thanks!!
This is actually just in time for me, I've been building a stock trading bot and have been exploring alternative transformer architectures that are more applicable to time series. 😁
Wont work, it is a stochastic process. Algotraders were the death knell of trading on technicals for retail. Unless you institutional, you simply don't have access to the signals the big boys do. Best you can do is pattern detection to try sniff out the smart money and close follow that.
The market is completely rigged. Literally the Market makers get all the trades and can move the market any way they want against any moron who puts in a large bet. The house ALWAYS wins. Be prepared to lose everything in the long run. You cannot be them when they have full control. The prices and volumes are fake as is the news to justify the movements. The house is not going to go bankrupt. If there was a way to beat the house then they would eventually have to shut down. Literally there are no consequences for cheating. The SEC fines the cheaters a small fraction of what they steal.
At best you can skim a little off the top as the algorithm works to undermine larger pots. E.g., if you throw in a big order thinking you got a winning hand 100% of the time the price will move against you to make you lose. The options all dictate how the price will move against you. It's designed literally to maximize profits. Their algorithm will always beat yours 100% of the time in the long run. The market exists not as an investment opportunity but as a rigged game to fleece the morons that play it. If you want, you can learn the hard way. All you have to do is go look at all the pump and dumps in the stock market to see how rigged it is. Crypto is 100x worse but basically same people. As they steal all the $$$ it becomes more condensed and chaotic.
@SlappyZA idk if it won't work, I've been reading several papers of people trying different architectures for stock predictions and I was getting decent results when I was training my earlier models for price prediction.
Either way it's a side project of mine and whether it works or not I'm having fun and learning a lot
Thanks for the input!
😢 1:44 😢 1:46
1) how does your method compare to Chronos-t5-large?
2) will you make your model available on HuggingFace?
Would be nice to give an intro to VQVAE. Would also be nice to see some real examples including failures, rather than just theory. Where do the tokens for sensors come from? Is an FFT an efficient encoder? How do the code books scale over different time scales or do they only work well over amplitude scales?
Nice, well rehearsed presentation!
Can you guys come up with something related to deep generative models ?
Overuse of the word "even" in the beginning
Lol
Doesn't matter