Serving the Quantitative Finance Community

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

Re: Hacking ML for fun & profit

July 16th, 2018, 11:26 pm

Thus, another solution is to train the NN to detect scenes that contain potential adversarial inputs.
No, it's not a solution: such attacks cannot be detected because both original and perturbed image are perfectly valid inputs.

The problem lies with the architecture itself. A feed-forward convolutional network is not enough. I am hopeful about this approach, waiting for the authors to test the robustness of their network against adversarial attacks.
 
User avatar
Cuchulainn
Posts: 20254
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Hacking ML for fun & profit

July 17th, 2018, 7:23 pm

Thus, another solution is to train the NN to detect scenes that contain potential adversarial inputs.
No, it's not a solution: such attacks cannot be detected because both original and perturbed image are perfectly valid inputs.
Meshuggah
That figure 1 has done it for met. The models and theory are wrong, totally.
And soon all this stuff will be used in hospitals etc.

In  CAD/CAM a rotated compressor is still a compressor. Now AI transforms a revolver to a mousetrap just by flicking it. Perceptron XOR all over again.
Probably means it won't work for moving part simulation"??? Just guessing of courses

Image
 
User avatar
Cuchulainn
Posts: 20254
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Hacking ML for fun & profit

July 18th, 2018, 11:10 am

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

Re: Hacking ML for fun & profit

July 19th, 2018, 7:31 am

For the avoidance of doubt: they won't deploy this algorithms in hospitals unless they pass the same clinical trials as other methods of treatment or diagnosis.

"The research has been submitted to a peer-reviewed medical journal amid hopes that the technology could enter clinical trials within a few years."

They're not going to dump this on the patients tomorrow.
 
User avatar
katastrofa
Posts: 7440
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

Re: Hacking ML for fun & profit

July 19th, 2018, 9:50 am

Like Thalidomide? There's no one on those commissions who understands those models. All competent experts have their money in the  AI business.
 
User avatar
Cuchulainn
Posts: 20254
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Hacking ML for fun & profit

August 18th, 2018, 12:09 pm

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

Re: Hacking ML for fun & profit

August 30th, 2018, 1:50 pm

Like Thalidomide? There's no one on those commissions who understands those models. All competent experts have their money in the  AI business.
There are some ML people who make a career out of bashing the Deep Learning fad, most notably Gary Marcus and Filip Piekniewski. And I'm sure they'll find a suitably jaded statistician to take the models under scrutiny.
 
User avatar
katastrofa
Posts: 7440
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

Re: Hacking ML for fun & profit

August 30th, 2018, 9:02 pm

Only if ML gents will be wiling to roll their sleeves up and do the testing themselves, because no one else has so much computational power to waste - I'm definitely against it as a climate change believer.
 
User avatar
ISayMoo
Topic Author
Posts: 2332
Joined: September 30th, 2015, 8:30 pm

Re: Hacking ML for fun & profit

August 30th, 2018, 9:41 pm

Once you train it, you don't need that much power. I think the model used for Moorfields diagnostics can easily run on a single PC.
 
User avatar
katastrofa
Posts: 7440
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

Re: Hacking ML for fun & profit

September 1st, 2018, 11:03 am

OK, but we are not discussing the rare good ML work, but the whole sewage of poor papers which ruin my science experience.
 
User avatar
Cuchulainn
Posts: 20254
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Hacking ML for fun & profit

February 11th, 2019, 11:03 am

Here, we have considered three, four, and five degree polynomial for regression fitting. One may consider higher degree polynomial in the simulation but it has been seen that by increasing the degree of the polynomials, the accuracy does not usually increase. In the future, it needs to develop a methodology about what degree polynomial one should use to get a result with acceptable  accuracy. This is however not of the scope of this paper and the authors are working in this direction and hope to communicate the findings in the future.

What does this mean?
 
User avatar
Cuchulainn
Posts: 20254
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Hacking ML for fun & profit

February 11th, 2019, 1:19 pm

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

Re: Hacking ML for fun & profit

February 11th, 2019, 8:47 pm

An apple a day keeps the robots away!
 
User avatar
Cuchulainn
Posts: 20254
Joined: July 16th, 2004, 7:38 am
Location: 20, 000

Re: Hacking ML for fun & profit

February 12th, 2019, 3:48 pm

An apple a day keeps the robots away!
I went to an AI doctor, told him that I broke my wrist in 2 places. Guess what he told me? Don't go to those places.Still more accurate than my human GP.
 
User avatar
ISayMoo
Topic Author
Posts: 2332
Joined: September 30th, 2015, 8:30 pm

Re: Hacking ML for fun & profit

February 19th, 2019, 10:16 pm

That would be awesome if AI was capable of such common sense advice.