Advancing AI

6 views
Skip to first unread message

Dean Collins

unread,
Jul 17, 2019, 9:19:58 AM7/17/19
to Silicon Beach (silicon-beach-australia@googlegroups.com)

Great 60 minutes article last weekend that might interest some of you - https://www.cbsnews.com/news/60-minutes-ai-facial-and-emotional-recognition-how-kai-fu-lee-is-advancing-artificial-intelligence-2019-07-14

 

I just ordered kai-Fu Lee’s book off Amazon - https://amzn.to/2YgryYl  

 

I wonder how many Australian politicians are thinking about this and how we can use this to get Australia ahead globally eg how do you think Australian politicians/teachers would react to putting a camera into classrooms to monitor attention.

 

 

 

 

Regards,

 

Dean Collins
Cognation Inc
de...@cognation.net
+1-212-203-4357        New York
+61-2-9016-5642        (Sydney in-dial).
+44-20-3129-6001      (London in-dial).

 

 

Xuanyi Chew

unread,
Jul 17, 2019, 9:17:25 PM7/17/19
to silicon-beach-australia
Hi,
Guy who creates a lot of AI here. My take is: we should stop fucking around with fundamental liberties. I shall explain this in entrepreneur terms.

Monitoring attention using AI of school children is the equivalent of entrepreneurs building self-congratulatory dashboards of vanity metrics like page hits. It does nothing to solve the fundamental problem.  I put it to you that people don't really know what the fundamental problem is.

The same can be said about surveillance in the large. It's been found to be largely ineffective (in that it's even hard to understand what "effectiveness" in surveillance even means[0]). Say you want to stop terrorism. Mass surveillance adds little to that conversation. At best surveillance is wankery - see above analogy on vanity metrics dashboard. At worst, you are opening a trove of data as an attack surface.  

You want to get Australia ahead? Let me tell you how. 
  1. First you need a fertile research environment. Where's Australia's OpenAI, DeepMind, FAIR, BAIR? Data61 is peanuts compared to these behemoths. I have no wish to move to America to do AI research because I'm deathly afraid of guns. 
  2. Second, you need a fertile ground for applications of research. For example, here's a well known problem that is fairly pertinent to Australia that is not pertinent to other countries: logistics through unpopulated landscapes. Going coast to coast in USA is different from going coast to coast in Australia because of differing support structures (many cities and towns along the way in America and few in between in Australia). This is  for example, fertile ground for application of self-driving vehicles. Furthermore if you analyze logistics problems in this context you may yield several novel generalizations on topology and geometry that are immediately applicable. 
  3. Third, you need an appetite for risk. I for one do not think VCs in Australia are particularly risk taking. I once pitched a Bell-labs style startup and was given the runaround. Truth is 90% of the AI research will fail. Which is why people keep iterating on the same shitty ideas over and over again. It's safe. Most people dare not even ask the questions. 
  4. Fourth, you need to be less profit motivated to get the next AI unicorn in Australia. Facial recognition has been "solved" for the past 15 years. Almost all applications are negative (mass surveillance) or trivial (faceapp). The following questions should be answered.
Now having said all that allow me to leave some open questions that might massively benefit the Aussie entrepreneur community. IMO if you want to create the next AI based unicorn in Australia, answer these questions. You may think that they have nothing to do with the next killer app, but trust me they are, if you know what questions you are asking.  These are listed in no particular order, and most are articulated to the point of generality:
  • Some gradient optimization algorithms like Adam has a flattening effect on the gradient landscape. The ideal case is when the flattening is in the direction of the landscape of the data. How can we make such a flattening happen?
  • How can we translate the models to a human interpretable manner? 
  • Are there better methods than gradient based methods for optimization? 
  • Should we care about things like VC-dimensions? 
  • Why can't you do scheduling and planning stuff well using deep networks?
  • Can you reduce the computational requirement for these deep nets? (yes for a lot of the models) Why?


Xuanyi Chew
@chewxy on Twitter


--
--
You received this message because you are subscribed to the Silicon Beach Australia mailing list. Vist http://siliconbeachaustralia.org for more
 
Forum rules
1) No lurkers! It is expected that you introduce yourself.
2) No jobs postings. You can use http://siliconbeachaustralia.org/jobs
 
 
To post to this group, send email to
silicon-bea...@googlegroups.com
To unsubscribe from this group, send email to
silicon-beach-aus...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/silicon-beach-australia?hl=en?hl=en

---
You received this message because you are subscribed to the Google Groups "Silicon Beach Australia" group.
To unsubscribe from this group and stop receiving emails from it, send an email to silicon-beach-aus...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/silicon-beach-australia/FA13C27070EBDE4E918C20FC36AB9B8202352E63C4%40COGNATIONSVR1.cognation.local.
For more options, visit https://groups.google.com/d/optout.

Geoffrey Litman

unread,
Jul 17, 2019, 9:49:48 PM7/17/19
to silicon-bea...@googlegroups.com
Great insight !!!

BOSCH is doing some interesting things...


Good friend of mine works for Bosch. 

have a wonderful day,
Geoff 
Optima, Hotel Guest Relationship Oriented, 
This email is strictly private and confidential and only intended to be read by the recipients. 


To view this discussion on the web visit https://groups.google.com/d/msgid/silicon-beach-australia/CAF8rqt4t9WKTtukAk2GDKBjrUwNfyKyaSexSpjd4FibyjnHKiA%40mail.gmail.com.
Reply all
Reply to author
Forward
0 new messages