SERVING THE QUANTITATIVE FINANCE COMMUNITY

Cuchulainn
Posts: 58952
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

### Re: If you are bored with Deep Networks

the first link s broken.

katastrofa
Posts: 7565
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

### Re: If you are bored with Deep Networks

Oh deer!
Attachments

tw
Posts: 1009
Joined: May 10th, 2002, 3:30 pm

### Re: If you are bored with Deep Networks

Has anyone had a look at rekognition?

Unsurprisingly, it seems not without controversy.

Cuchulainn
Posts: 58952
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

### Re: If you are bored with Deep Networks

Has anyone had a look at rekognition?

Unsurprisingly, it seems not without controversy.
In the old days the sales team fixes it in the next revision of the product.

Cuchulainn
Posts: 58952
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

### Re: If you are bored with Deep Networks

Damiano Brigo's views on ML.

He seems to be saying that regulators don't trust black boxes.

tw
Posts: 1009
Joined: May 10th, 2002, 3:30 pm

### Re: If you are bored with Deep Networks

Has anyone had a look at rekognition?

Unsurprisingly, it seems not without controversy.
In the old days the sales team fixes it in the next revision of the product.
It is not so much the bugginess, more the scope for abuse.

From a recent write up:
Last month, journalists at the New York Times began collecting images from a publicly accessible camera on the roof of a restaurant in Midtown Manhattan. After gathering one day’s footage, the team employed Amazon Rekognition, an online service that uses artificial intelligence (AI) to match faces from different images, to see if they could put names to the 2,750 pictures they had collected. To the concern of local workers, they found that it worked, even from a partial image or a steep angle. For around $60 (£46), the system was able to identify large numbers of people without their knowledge or consent . since it was so cheap, I tried getting an account in order to see who it could ID from a few (online) photos. In order to pay you had to give credit card details in advance and, a very few seconds after entering some, I had a call from my bank warning about about suspicious transactions, so I stopped the experiment there! Cuchulainn Posts: 58952 Joined: July 16th, 2004, 7:38 am Location: Amsterdam Contact: ### Re: If you are bored with Deep Networks This sounds illegal.And too good to be true. BTW which company did you want to pay? Last edited by Cuchulainn on May 20th, 2019, 9:33 pm, edited 1 time in total. tw Posts: 1009 Joined: May 10th, 2002, 3:30 pm ### Re: If you are bored with Deep Networks This sounds illegal.And too good to be true. BTW which company did you want to pay? Amazon Web Services Cuchulainn Posts: 58952 Joined: July 16th, 2004, 7:38 am Location: Amsterdam Contact: ### Re: If you are bored with Deep Networks Hmm. And the bank started to react? tw Posts: 1009 Joined: May 10th, 2002, 3:30 pm ### Re: If you are bored with Deep Networks Hmm. And the bank started to react? Indeed. I don't read too much into it but it is nonetheless revealing the "Amazon" name didn't reassure them. ISayMoo Topic Author Posts: 1645 Joined: September 30th, 2015, 8:30 pm ### Re: If you are bored with Deep Networks They were worried that someone is running jobs on AWS using a credit card number stolen from you, either for their own purposes or just to test whether the card is active. Cuchulainn Posts: 58952 Joined: July 16th, 2004, 7:38 am Location: Amsterdam Contact: ### Re: If you are bored with Deep Networks Has anyone had a look at rekognition? Unsurprisingly, it seems not without controversy. In the old days the sales team fixes it in the next revision of the product. It is not so much the bugginess, more the scope for abuse. From a recent write up: Last month, journalists at the New York Times began collecting images from a publicly accessible camera on the roof of a restaurant in Midtown Manhattan. After gathering one day’s footage, the team employed Amazon Rekognition, an online service that uses artificial intelligence (AI) to match faces from different images, to see if they could put names to the 2,750 pictures they had collected. To the concern of local workers, they found that it worked, even from a partial image or a steep angle. For around$60 (£46), the system was able to identify large numbers of people without their knowledge or consent .
since it was so cheap, I tried getting an account in order to see who it could ID from a few (online) photos.
In order to pay you had to give credit card details in advance and, a very few seconds after entering some, I had a call from my bank warning about
about suspicious transactions, so I stopped the experiment there!
But it is not so with these AI combinatorial-based methods is that we can't check for bugginess because these methods are black boxes and thus non-interpretable? The underlying 'model' is absent.

katastrofa
Posts: 7565
Joined: August 16th, 2007, 5:36 am
Location: Alpha Centauri

### Re: If you are bored with Deep Networks

I'm getting bored with that argument:
1. The problem is neither new nor specific to the AI field. Any system simulations - in climate science, biology, medicine, ... have it, but nobody's picking at those.
2. Codes are often published with sufficient information to repeat the simulation (even stochastic). Obviously, more people complain about non-interpretability than actually check them (because the codes are standardly tested by more experienced developers?).
3. If the method was possible to implement, it's possible to re-implement, repeat and reproduce the result. In some cases, you may require access to Google's Borg, though

non-interpretability - what do you mean by interpreting the method? Methods (or models) are way of interpreting data or representing theories. I assume it's non-reproducibility.

Still, I think that sooner or later every field will have to embrace the idea of results which are impracticable to reproduce.

Cuchulainn
Posts: 58952
Joined: July 16th, 2004, 7:38 am
Location: Amsterdam
Contact:

### Re: If you are bored with Deep Networks

what do you mean by interpreting the method?

aka not a black box.

Or more accurate;y, I suppose interpreting the I/O as in nuerical methods based on models. AI seems to use data with no models.

tw
Posts: 1009
Joined: May 10th, 2002, 3:30 pm

### Re: If you are bored with Deep Networks

[snip]
But it is not so with these AI combinatorial-based methods is that we can't check for bugginess because these methods are black boxes and thus non-interpretable? The underlying 'model' is absent.
I leave it for the experts to debate the technical details.
I imagine it cannot conceivably work 100% of the time, but will probably increase in its efficacy continually to get to levels which will surprise and terrify the lay person (such as me).

The interesting aspect for this to me is the pricing.

if you are going to run the gauntlet and supply an application that many with consider illegal and many more will consider unethical why do it for seemingly paltry sums of money?
Is it to get first movers advantage and use the brand to quickly squeeze out others before they can get to market?
Or is it to democratize the subject to remove spooks/Cambridge Analytica feel and make it a novelty, normalizing it and taking the sting out of the (as yet nonexistent) regulator?
Or both?
Given that the (UK) Police National Database is trialing some German provided software, it is interesting the price/quality drivers on who will win out.