http://www.i-programmer.info/news/105-artificial-intelligence/7352-the-flaw-lurking-in-every-deep-neural-net.html

A recent paper “Intriguing properties of neural networks” by Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow and Rob Fergus, a team that includes authors from Google’s deep learning research project outlines two pieces of news about the way neural networks behave that run counter to what we believed - and one of them is frankly astonishing.

This could prove to be really bad news for anyone with a vested interest in neural networks unless a solution is found.

Digital processes will never achieve intelligence. NN have their place, but they will never aspire/achieve “intelligence” in an analog world.

NOTHING in nature that is intelligent is also digital, and nature has had millions or billions of years to work on the issue.

For intelligence, we need something new. Something analog. See where I’m going with this? (And if you don’t, that’s ok. <grins>)

Digital systems are logical systems, therefore you are saying that NOTHING in nature that is intelligent is also logical.

It is not difficult to come up with definitions of “nature” and “intelligence” where this proposition is true.

That is remarkable . I had read of the 6% “cat recognition” success rate that Google achieved in their wanton experiment last year (i.e. 6% of the time, it recognised a cat as a cat, 94% of the time, it did not). At the time I said such a rate of recognition might as well be chance. It seems that’s closer to the truth than I thought. How very odd. Perhaps neural nets are far-sighted, in that they will not consistently recognise something that is within 2% similarity of their learned pattern, but will recognise things further deviant from the pattern. That is the only area in which I could imagine a solution to be found.

Andrew Smith - May 27, 2014:

Digital systems are logical systems, therefore you are saying that NOTHING in nature that is intelligent is also logical.

I don’t think he’s saying that at all, Andrew. Just because digital systems are logical doesn’t mean that logical systems have to be digital. Analog systems don’t exclude 0 or 1, after all. To my way of thinking, “digital” is just a subset of “analog”.

If a finger is a digit then fingers are digital.

I don’t see any problem with digital and intelligence : It’s possible to treat 10 base numbers with binary, it’s possible to listen do compact disc which is a digital record out of analogic datas.

DISCLAIMER: I personally don’t like NN solutions.

I think there is a wider issue with NN solutions, that being that you can’t really be sure what they’ve learned.

E.g. You may think you’ve trained your NN to differentiate cat and dog pictures.

It’s done really well on the training set - but for all you know, all of it’s answers could be based on the colour of the pixel in the top left hand corner!

This basic problem applies even to the most sophisticated NNs we know of. Humans are notorious for this and many problems in human culture such as racism can be traced in part to using the wrong data from a training set and making invalid associations.

The problem is not really analog vs digital. Analog is after all only an analogy of the world. The problem is that humans are really incapable of self observation - we ourselves are products of genetic and social programming and are rarely able to break out of these boxes.

Humans are generally capable of self-observation, though that capability is not uniformly distributed. It is related to the concept of a highly “self-monitoring” personality. Combining high self-monitoring with some level of understanding of artificial intelligence and the computational and algorithmic underpinnings of human intelligence and the process of human evolution allows one to begin drawing conclusions about the deeper nature of what is observed about human nature.