[Tyler]

In one aspect of vision, computers catch up to primate brain
18 December 2014, 7:10 pm

For decades, neuroscientists have been trying to design computer networks that can mimic visual skills such as recognizing objects, which the human brain does very accurately and quickly. Until now, no computer model has been able to match the primate brain at visual object recognition during a brief glance. Now neuroscientists have found that one of the latest generation of 'deep neural networks' matches the primate brain.

Source: Artificial Intelligence News -- ScienceDaily

[DemonRaven]

That is a excellent article but this statement bothers me "Cadieu says that researchers don't know much about what exactly allows these networks to distinguish different objects." We need to be able to understand this and therefore be able to control it in the future if necessary. You can not control what you do not understand.

[ivan.moony]

sounds like a real science fiction, an algorithm coming from another planet: "we know how to program it, it works, but we don't yet understand it..."

maybe neural nets hide more than simple picture recognition. Does anyone knows what else could be done with neural networks in these days? Maybe speech recognition which is shown to be a big problem?

[Don Patrick]

My knowledge about neural nets is limited to a number of articles, I still don't understand how to code one. Neural nets are very general pattern recognisers, that can apply to any sort of data: It looks for prominent "features" in the information, whether that's bits or pixels or soundwaves, and from this it constructs statistically frequent patterns. If you look at a neural net's visual pattern of a cat faces, it looks like an average of 20 merged pictures of cat faces.
Voice recognition technology is also based on neural nets. That's why they have to be "trained" before they recognise the features/patterns in your voice well. Unfortunately they are a let-loose-and-cross-your-fingers technology that learn purely statistical. Since they have to be shown millions of examples to form a good average pattern, this learning process is beyond human capacity to keep track of and perhaps this is why their trainers don't understand what they do. Because they only learn statistically prominent patterns, they are less accurate at minor differences, which is why Google's vision experiment was only 6% accurate at recognising cat faces last year, and billions more examples wouldn't help increase the accuracy by much. An average never is accurate.

This only changed a few months ago, when neural net scientists decided to give the neural nets some hints manually, and the accuracy of recognising specific subjects skyrocketed to 94%. There are however still surprises and bugs being "discovered" in the way neural nets function, and their accuracy is not at reliable levels. As a statistical bazooka they are great however.

[Art]

Quite an interesting video, Don. It highlights one premise that Ray Kurzweil touted about biological vs non-biological intelligence and why non-biological (digital / electronic / computer) intelligence will one day soon, become dominant.

Somewhat scary, somewhat engaging and enlightening and yet foreboding.

[Don Patrick]

I think the best you can take away from the video is that we're going to have a massive unemployment issue on our hands. But stubborn as I am I can't support the suggestion they make about continued exponential increase of AI, as I'm not quite sure what the graph they show is based on. If for instance you'd mark the 6% accuracy of computer vision last year and the 94% accuracy of this year you would get a spiking graph just like that, but surely it would stop once neural nets can recognise cat faces with 100% perfect accuracy?

[Korrelan]

Pftt... Check the date published... old tech...

[Art]

Many years ago people cried and bemoaned the possibility that robots would "take away" all our jobs!
At that time, robots actually performed labor intensive, dangerous and excessively boring tasks. They
also helped usher in new jobs for quality control inspections, numerical systems control operators,
system analysts, maintenance and repair positions where at one time, none of these existed.

My friends and I have always thought there there would always be positions to sort of coincide with the
robots who's human counterparts were replaced. I'm not so sure any longer, especially after videos like
these and the possibility that such things might be an inevitability in our future lives.

So Don, How's that AI programming working out for you these days?

[Don Patrick]

Truth be told, it's cost me $20000 in life savings and earned me not a penny. However, my resulting programming skills did regain me the use of my right hand (I made some RSI software), and got me several job offers.

I actually operate printing machines for a day job. The printing business used to be fine until the economy drove people to cheaper alternatives like email newsletters and online marketing. Every month we get news that another printing company has gone bankrupt, every few months staff is permanently reduced because there just isn't enough work anymore. Printing greeting cards used to be our biggest business. Now it is rare: email has replaced paper and operating the email servers are just 3 highly educated IT guys as opposed to a dozen bankrupt printing companies employing hundreds of grunt workers. I'm fine with robots taking my job, but only if it means I have to work less. So, governments will have to find a better work/income system eventually.

This decline is not true of all though. This computer vision stuff for instance may be used to replace a handful of human nightwatchers with robots (as are currently on the market), but it is also applied to tasks that didn't exist as human jobs before. Before Google, nobody rang the library every hour to ask a question, and not a whole lot of people were hired to find pictures of cats to laugh at.

[Art]

Oddly enough, it seems that the once dwindling "blue collar" labor force seems to be on the rise. I've yet to see robots that can (aside from a few robotic lawnmowers) do grasscutting, landscaping (timbers, gravel, mulch, leaf cleanup, etc.), construction form building, concrete placement / finishing, overall construction, equipment / heavy equipment operation, carpentry, brick / block laying, electrical work, roofing, house building, siding, windows / doors installations, fire fighting, paramedic / rescue, various inspector positions, testing agencies, etc.

I think there are some positions that a robot simply could not do (ok...at the present) and maybe for the next decade...who knows....

Thanks for your insight Don. Don't tell me they have computers composing verse for the inside of greeting cards now! I do suppose it's possible depending on the category (birthday, anniversary, graduation, birth, death, holiday seasons, etc.). That might make for a nice project for botmasters and other creative sorts.

[Don Patrick]

The messages on the cards are still composed by humans, only the "handwritten" signatures are digital prints nowadays . It would indeed be a nice project, like a twist on AI poetry contests.

I also used to be a building engineer. Here is your house building robot

[Art]

To me, that's a concrete box (with vaulted ceiling). It might serve a few or perhaps many but I'm not sure mainstream urban-dwellers are quite ready for them.

I'm sure robotic building techniques will find their way into our lives in the near future.

We ought to look into organizing a "greeting card" contest for our chatbots! Might prove interesting, fun and thought provoking!

[8pla.net]

If anyone is interested... There are two neural networks, on my site Elizabot.com.
They are both written in PHP.   It is an interesting experience to watch them learn.
Sometimes, they just don't learn... They fail..  More often they're successful learning.
Which makes them more life-like.  Feels sort of like a science experiment.