Page 48 of 50

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: February 17th, 2019, 7:37 pm
by TinMan
You are performing an optimization of an approximation to a non convex function, with the wrong function approximator, based on a finite sample of noisy observations.  The global minimum is overfit by definition.

Lack of robustness isn't cause by stochastic gradient descent.  If the problem is ill posed it's ill posed.

If you're worried about perturbed images then perturb the images in the training set.  If your fit is on a particular sample type why expect it to extrapolate correctly?  How is that different from fitting a linear regression and extrapolating the line of best fit?

Ethics?  Where's the ethics in linear regression software?  There's vast swathes of influential and dishonest propaganda based on regression which affects enormous numbers of people based on spurious regressions.

Where's the ethics in significance testing?  No statistics textbook is complete without examples of doctors not understanding p values or Bayes rule.

Psychology?  The journals and popular psychology books are full of results that don't replicate.  Most of their results are false.  Taleb had a recent argument with some of them on twitter, and none of them actually understand correlation to a first year stats level.

The ethics issue is a red herring.  That's a political battle among the usual technocrats and rent seekers.

 

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: February 17th, 2019, 8:26 pm
by Cuchulainn
Lack of robustness isn't cause by stochastic gradient descent.  If the problem is ill posed it's ill posed.
OK, ill-posed is a good start. BTW what is ill-posed. precisely, ill-posed != well-posed?
I'll have a look at Yuan et al but it seems to me that the issues are caused by perturbations of GD , see eq. (5). You can find adversarial examples by perturbing GD a little. That's my understanding.

There must be something better than (S)GD?

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: February 18th, 2019, 5:41 pm
by Cuchulainn
Ethics?  Where's the ethics in linear regression software? 
Linear regression never did no one no harm. More worrying is Big Brother and the Police State.

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: February 19th, 2019, 10:20 am
by Cuchulainn
So, tons of people are learning how to code now. What happens in a few decades when there are millions of us that can code?

In 5 years time, how many developers  will know and use ML in their applications?

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: February 21st, 2019, 2:13 pm
by Cuchulainn
Facebook’s chief AI scientist: Deep learning may need a new programming language

https://venturebeat.com/2019/02/18/face ... -language/

Well, who wudda thought that?

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: February 21st, 2019, 3:53 pm
by Cuchulainn
Another point: AI has been 'sold' to upper management. I just hope the AI developers can deliver.

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: February 28th, 2019, 1:50 pm
by Cuchulainn
Can this graph predict the future?

Image

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: April 24th, 2019, 12:22 pm
by rmax
Interesting to read page 1 of this thread again.

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: April 26th, 2019, 10:39 am
by Cuchulainn
Interesting to read page 1 of this thread again.
No one forecast the unstoppable rise of Python and AI.

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: April 27th, 2019, 2:29 pm
by Cuchulainn
Coursera raises $103 million to prepare online learners for the ‘fourth industrial revolution’

https://venturebeat.com/2019/04/25/cour ... evolution/


It’s against this backdrop that online education platforms are expected to grow from a $4 billion industry today to a hefty $21 billion market by 2023. 

I wonder which estimation tools they use. BTW 2023 is about in 5 years time. With a relatively small investment such a yuge ROI can be generated.

And all thanks to gradient descent?

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: May 4th, 2019, 12:08 pm
by Cuchulainn

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: June 27th, 2019, 8:00 am
by Cuchulainn
Python code written by mathematicians with scant regard for software design and maintainability.Leading to a big ball of mud.

Instead of using Deign Patterns to customise the application (ever heard of Strategy or Builder?) what happens is that the code is copied and modified., Now you two complete programs to maintains. Each new feature is a new program, and so on by mathematical induction.
People tend to compare languages based on many kinds of metrics, e.g. cute features, easy to use, functionality etc.

A possible bad scenario is Python code becoming unmaintainable because those writing the software (for example, applied mathematicians) have had no exposure to design principles of Design Patterns.

C++ v Python? the latter is (much) more difficult to maintain. And there's an associated price tag.

Corollary: discussions on software maintainability tend to fall on deaf ears.

People don't think like software engineers. It plays second fiddle to cute mathematical formulae.

Some perspectives on writing software:

An engineer, a software engineer, and a computer programmer are driving down the road when the car they are in gets a flat tire. The engineer says that they should buy a new car. The software engineer says they should sell the old tire and buy a new one. The computer programmer says they should drive the car around the block and see if the tire fixes itself.

Last but but not least, who pays the bill?

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: August 17th, 2019, 4:34 pm
by Cuchulainn
Most data scientists/ML engineers have been coding for 2-4 years. Most have master’s degrees

Rookies. Gotta start somewhere.

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: September 3rd, 2019, 2:31 pm
by AndreyShihov
On a deeper level, the mental blocks are 

1. Think about WHAT before HOW
2. Can you explain the DATA FLOW in 5 minutes to the customer?
My two cents to point 1. I would also add WHO before HOW. Might help to understand that parallelism is about Workers which have to be well organised to achieve all benefits of parallelism. Developer becomes Manager who has to find optimal communication and resources sharing strategy between workers. 

Re: FIVE years from now: can you predict the Software Landscape in anno 2023?

Posted: September 3rd, 2019, 4:42 pm
by Cuchulainn
On a deeper level, the mental blocks are 

1. Think about WHAT before HOW
2. Can you explain the DATA FLOW in 5 minutes to the customer?
My two cents to point 1. I would also add WHO before HOW. Might help to understand that parallelism is about Workers which have to be well organised to achieve all benefits of parallelism. Developer becomes Manager who has to find optimal communication and resources sharing strategy between workers. 
Indeed!
I used to be a requirements analyst for a good while and I found Inquiry-based .. to be so invaluable for high-risk (all are :-)) fixed-price projects

https://www.researchgate.net/publicatio ... s_Analysis

Q: what is, what kinds of, who, when, what-if,  how-to, relationship, follow-on etc. and you can interleave them. It's in my 2004 book Domain Architectures (Wiley)

Corollary: some/many developers spend 10 minutes om WHAT (their own internal model,of reality) and an eternity on HOW.
In the legendary words of Luigi Ballabio, I now duck for cover!