Page 28 of 34

Re: If you are bored with Deep Networks

Posted: April 30th, 2019, 1:23 pm
by ISayMoo
Both sources recommend the learning rate of 3E-4. Looks like we have a winner.

Re: If you are bored with Deep Networks

Posted: May 5th, 2019, 9:22 pm
by Cuchulainn
  • "By 2030, PWC estimates the AI market to be worth the current GDP of China and India combined."

Re: If you are bored with Deep Networks

Posted: May 5th, 2019, 9:25 pm
by MarkBrezina
^Jesus!

Re: If you are bored with Deep Networks

Posted: May 6th, 2019, 7:55 am
by Cuchulainn
In 1989 they predicted that the AI market in 1993 would be 3 billion.

BTW when can we expect patented AI products on a massive scale?

Re: If you are bored with Deep Networks

Posted: May 6th, 2019, 9:12 am
by MarkBrezina
What do you mean me making AI products?

If I had to suspect in general the patented AI products production becoming massive I would say it is already or has already happened?

Re: If you are bored with Deep Networks

Posted: May 6th, 2019, 11:06 am
by Cuchulainn
Software + hardware products.

massive aka mainstream

Re: If you are bored with Deep Networks

Posted: May 6th, 2019, 11:17 am
by ISayMoo
"Narrow AI" (aka Machine Learning) technologies adopted at massive scale:
- machine translation
- video captioning
- object & face recognition
- customer service chatbots
- driver assistance software in cars? (I don't know how much ML they use - certainly it's not end-to-end but modular)

Entering mainstream use:
- ML algorithms in "computational photography" (first on smartphones, later in digital cameras)
- Deep Learning raytracing algos in new NVidia video cards (RTX)
- DL autofocus algos in new Olympus and Sony cameras

AGI (Artificial General Intelligence) doesn't exist yet.

Re: If you are bored with Deep Networks

Posted: May 6th, 2019, 11:23 am
by ISayMoo
Not mainstream use but with large economic or societal impact (entering use or already there):
- crime prediction
- drug discovery
- automated data labelling in life sciences and astronomy
- genetic data analysis
- fraud detection
- autonomous weaponry
- optimising compilers

Re: If you are bored with Deep Networks

Posted: May 6th, 2019, 2:49 pm
by Cuchulainn
Nice summary. When will the 'laggards' jump on board?

Re: If you are bored with Deep Networks

Posted: May 6th, 2019, 7:59 pm
by ISayMoo
Who knows? Sometimes corporations fail to notice the opportunity (like Big Pharma with protein folding). Sometimes the opportunity isn't there (e.g. there is little progress in automated theorem proving).

Re: If you are bored with Deep Networks

Posted: May 6th, 2019, 9:04 pm
by ISayMoo
You should like this line of work, Cuch: https://arxiv.org/abs/1902.09592 (presented at ICLR this year)

Re: If you are bored with Deep Networks

Posted: May 10th, 2019, 4:27 pm
by ISayMoo
It's not a bug, it's a feature

They want to solve the adversarial robustness problem by claiming that humans are wrong and computers are right.

Image

So now you know: the animal on the right is not a panda. It's a gibbon. We're just too limited computationally to notice the truth.

Re: If you are bored with Deep Networks

Posted: May 10th, 2019, 5:30 pm
by katastrofa
Like the race of New Men from PK Dicks's Our Friends from Frolix 8. The ML gnomes are losing touch with reality.

Re: If you are bored with Deep Networks

Posted: May 10th, 2019, 7:21 pm
by Cuchulainn
Will we see GD doing a wiggle dance in our lifetime, Jim?

https://www.rte.ie/brainstorm/2019/0425 ... or-pollen/

My bees have access to a nature reserve. Many recruiters, less foragers?

Re: If you are bored with Deep Networks

Posted: May 11th, 2019, 10:29 pm
by katastrofa
Cognitive approach of spatial and temporal information processing in insects

One of the schools of AI assumes that it can't be realised without embodying it in a physical form interacting with the environment. I'm not convinced - a ping-pong ball or a thermometer come under this definition.