What exactly does it do? worth it?
This is OLD news, they talked about this last year,
The ML algo is well known, so it can be placed on asic
What good is it ?
Self driving cars, spy cameras with ability to ID U built inside, red-light cameras that read license plates and email u a speeding ticket,
Big Brother SPYING IOT all on one chip,
All the algos, language translation, image analysis, big-data analysis, its all down now to one small common well understood algorithm
Have they delivered the chip? Is it working? Probably not, Google is also working on the same chip, everybody is
The thing to remember is that BITMAIN is a one trick pony, an the S9 is dead, nothing new is coming as nobody can get 14nm or smaller to work, so in order t stay in biz BITMAIN had to find new markets, so they play the same trick as 'mining' they promise something and hope that the can get the money up front and some day deliver,
The problem is the ML people ( mostly NSA, CIA, FBI MS FB ) google-is-nsa type spying level, they're NOT going to use chinese shit because the US-GOV wants control of the world,
Now I doubt that BITMAIN will give us the $1 chip that does ML, but certainly the people who make the rasberry-pi clone CPU can and will.
***
two most general and popular algo's are ...
Artificial Neural Networks
Deep Neural Networks (Convolutional or Recurrent Networks)
Both of the above can be coded in a page of PYTHON, so its easy to learn this stuff, but it takes a bunch of GPU ( 1080 class ) to 'train', but once you train the system with your data, then you can get 99.9% which is better than a human,
all of it works pretty much the same, huge matrixes are correlated and then a calculus descent function is applied to find the optimum
Y = F (X)
Say you have an 'x' it could be photo of a woman, and you want a Y that indicates on a scale 1-10 is she hot or not, ... the ML algo will generate you a F() function ( matrix ) that will take the image, and return you a number 1-10 of 'Y'.
Now how the F() is built is training, both super-vised and un-supervised learning, supervised is where I pass the algo 1M photos and I tell it 1-10 for each photo, eventually the algo figures it all out, un-supervised training is where the algo all on its own figures out what is a 1, and what is 10,
The best in developing this shit is 10% supervised, followed by 90% un-supervised, once the algo gets general idea, it will always learn better than a human, but initial seed of 10% supervised gets it in the right direction.
***
Now set back and think in the general form of problems
Y = F ( x )
x can be anything, and Y can be what ever u want, about 8 years ago I wrote on ML that took any music 'live' (MP3) and generated 'music script' so that I could play guitar of any song I wish, the transcription sw at the time was expensive and didn't work so I rolled my own.
***
Now specific to MINING, say you have a bitcoin address and you want its PRIVATE-KEY, well there is a database on line of all training data, so you train an ML to learn the relation between all addresses and all keys, ... this algo actually exists
Lot's of interesting stuff to think about
