Serving the Quantitative Finance Community

 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Machine Learning: Frequency asked Questions

December 13th, 2019, 8:23 pm

Anything goes.
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

February 5th, 2020, 9:47 pm

Using Julia for ML by Neural ODEs
https://julialang.org/blog/2019/01/fluxdiffeq/

There is no a priori reason why it can't be done this way.

So as our machine learning models grow and are hungry for larger and larger amounts of data, differential equations have become an attractive option for specifying nonlinearities in a learnable (via the parameters) but constrained form. They are essentially a way of incorporating prior domain-specific knowledge of the structural relations between the inputs and outputs.
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

May 2nd, 2020, 10:18 am

A.I. can’t solve this: The coronavirus could be highlighting just how overhyped the industry is

https://www.cnbc.com/2020/04/29/ai-has-limited-role-coronavirus-pandemic.html?__source=sharebar|linkedin&par=sharebar

“It’s fascinating how quiet it is,” said Neil Lawrence, the former director of machine learning at Amazon Cambridge.

“This (pandemic) is showing what bulls--t most AI hype is. It’s great and it will be useful one day but it’s not surprising in a pandemic that we fall back on tried and tested techniques.”
Those techniques include good, old-fashioned statistical techniques and mathematical models. The latter is used to create epidemiological models, which predict how a disease will spread through a population. Right now, these are far more useful than fields of AI like reinforcement learning and natural-language processing.

// In fairness, it is not a law of gravity that AI should be good at everything. Maybe stick to statistics?
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

May 2nd, 2020, 6:30 pm

Stupid question: I want to do ML application but I have no (or not enough) training data; is this possible?
 
User avatar
JohnLeM
Posts: 379
Joined: September 16th, 2008, 7:15 pm

Re: Machine Learning: Frequency asked Questions

May 4th, 2020, 10:22 pm

A.I. can’t solve this: The coronavirus could be highlighting just how overhyped the industry is

https://www.cnbc.com/2020/04/29/ai-has-limited-role-coronavirus-pandemic.html?__source=sharebar|linkedin&par=sharebar

“It’s fascinating how quiet it is,” said Neil Lawrence, the former director of machine learning at Amazon Cambridge.

“This (pandemic) is showing what bulls--t most AI hype is. It’s great and it will be useful one day but it’s not surprising in a pandemic that we fall back on tried and tested techniques.”
Those techniques include good, old-fashioned statistical techniques and mathematical models. The latter is used to create epidemiological models, which predict how a disease will spread through a population. Right now, these are far more useful than fields of AI like reinforcement learning and natural-language processing.

// In fairness, it is not a law of gravity that AI should be good at everything. Maybe stick to statistics?
What is truly fascinating is that Neil Laurence needed the Covid crisis to make this observation.
 
User avatar
JohnLeM
Posts: 379
Joined: September 16th, 2008, 7:15 pm

Re: Machine Learning: Frequency asked Questions

May 4th, 2020, 10:23 pm

Stupid question: I want to do ML application but I have no (or not enough) training data; is this possible?
Maybe you should consider the same approach than the one used for the curse of dimensionality : just do the best possible with your training data and see if it is enough.
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

May 5th, 2020, 4:58 pm

A.I. can’t solve this: The coronavirus could be highlighting just how overhyped the industry is

https://www.cnbc.com/2020/04/29/ai-has-limited-role-coronavirus-pandemic.html?__source=sharebar|linkedin&par=sharebar

“It’s fascinating how quiet it is,” said Neil Lawrence, the former director of machine learning at Amazon Cambridge.

“This (pandemic) is showing what bulls--t most AI hype is. It’s great and it will be useful one day but it’s not surprising in a pandemic that we fall back on tried and tested techniques.”
Those techniques include good, old-fashioned statistical techniques and mathematical models. The latter is used to create epidemiological models, which predict how a disease will spread through a population. Right now, these are far more useful than fields of AI like reinforcement learning and natural-language processing.

// In fairness, it is not a law of gravity that AI should be good at everything. Maybe stick to statistics?
What is truly fascinating is that Neil Laurence needed the Covid crisis to make this observation.
Of course, AI wasn't around in 2001 when the ODE model were used to predict [$]2e^6[$] deaths during the foot-and-mouth disease.
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

May 5th, 2020, 5:50 pm

Stupid question: I want to do ML application but I have no (or not enough) training data; is this possible?
Maybe you should consider the same approach than the one used for the curse of dimensionality : just do the best possible with your training data and see if it is enough.
I think that classic AI is not built for these kinds of problems.
BTW have you already published your article?
 
User avatar
JohnLeM
Posts: 379
Joined: September 16th, 2008, 7:15 pm

Re: Machine Learning: Frequency asked Questions

May 7th, 2020, 5:52 pm

Stupid question: I want to do ML application but I have no (or not enough) training data; is this possible?
Maybe you should consider the same approach than the one used for the curse of dimensionality : just do the best possible with your training data and see if it is enough.
I think that classic AI is not built for these kinds of problems.
BTW have you already published your article?
I don't think so. I'll write to Paul to have news
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

June 1st, 2020, 6:58 am

I have nothing against hypes in general, they are inherent to every field of human life that was infected with progress. My only problem with the AI hype is that its inspirators claim that these methods "work" and we can start to deploy them. Since using them requires very little brain and effort, it's tempting to believe them.

Not me.
And it was not so long ago that we were being told - almost on a daily basis here - that AI was the greatest thing since sliced bread. And dissenters were shouted down and insulted that they did not understand. 
The funniest was computer scientists lecturing mathematicians on what mathematics is. Hubris.
I have lost count, but since the 1960s AI has risen and dramatically fallen at least three times. It's on record.
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

June 1st, 2020, 7:21 am

The term "Machine Learning" was coined by Frank Rosenblatt (the inventor of the ill-fated Perceptron which cannot learn XOR).A better name might be "Pattern Recognition" as show in this mathematically precise article on Hyperplane Separation

http://www-isl.stanford.edu/~cover/papers/paper2.pdf

An exercise: use Cover to show why Perceptron XOR has no solution.

Furthermore. these problems are addressed in Functional Analysis.

http://math.uchicago.edu/~may/REU2014/R ... s/Peng.pdf

What is emerging IMO is a new generation of Python-based Data Science plumbers without adequate mathematical and programming background. Of course, I could be wrong. 
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

July 4th, 2020, 12:22 pm

Image
 
User avatar
ISayMoo
Posts: 2332
Joined: September 30th, 2015, 8:30 pm

Re: Machine Learning: Frequency asked Questions

July 4th, 2020, 6:55 pm

Image
Excellent!
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

July 18th, 2020, 3:56 pm

What about basinhopping? Global minimum?

https://docs.scipy.org/doc/scipy/refere ... sinhopping
 
User avatar
Cuchulainn
Topic Author
Posts: 20255
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Machine Learning: Frequency asked Questions

September 18th, 2020, 9:49 am

All of a sudden 

Grady Booch on Software Architecture for AI (under videos)

https://thalesians.com/videos/


I think we are getting to the stage when AI becomes a set of algorithms to be integrated into 'mainstream' software design. Looking into my crystal ball, this is the next major challenge for data science developers wishing to write production code.

I discuss this very topic for computational finance in my recent Thalesians' video.

As UML et alia expert from the beginning, I don't think traditonal OO is the way to go, but architectural patterns as Grady Booch briefly alluded to. But a small step before that is System Decompositomn and Data Flow Diagram (Structured Analysis), parallel data dependency graphs.