Saturday, October 13, 2018

Overview of AI

The gist of this post is a recap, somewhat, that is being done briefly. I have been working in Quora, for the most part, but intend to get back to this blog.

The recent CACM of the ACM published an overview of AI by the head of CompSci at UCLA. His view resonated real strongly with me. So, I want to provide a quick look at some of the key items.

First, the article: Human-level intelligence or animal-like abilities?  Author: Adnan Darwiche.

Adnan starts by summarizing some of the recent successes (if you can call them that) that got so much attention in the press. The computer (AI driven) winning in Go. The computer learning how to play a game and beating expert humans. The list goes on. And, the basis is the neural net approach that is not new. Marvin Minsky wrote of these one-half a century ago.

But, too, people got upset. Some are believing in the coming singularity. I say, worry about the current crapularity. Oh, the robots are coming. Just like there was a warning 200 years ago about the invasion of the Brits who wanted to spank the free Americans.

Adnan mentions some of the mainstream areas of AI that have been ignored in the rush after the numeric paradise that is being given us by machine learning. He mentions "knowledge representation, symbolic reasoning, and planning" in short. Adnan was there when the AI winter came about from the mania of the '80s. He was student.

I would tell him that knowledge based engineering came out of that thrust and went fine for decades doing great things. I can (and will) point to writings about this. However, let's continue with Adnan's look from his lofty position.

He reminds us that AI had a reasoning approach, from the beginning. The recent work? No, it's numeric processing, pure and simple. Now, he mentions that particular neural nets (NNs, of which there are many variety) are doing some type of representation of a model. Given that, we can understand these supposed black-boxes with the proper approach. He uses two concepts: model-based and function-based. The former was (is) a huge part of AI, in general. You know, the latter was seen in engineering, too.

So, what we are talking is that the tried-and-true of science and engineering somehow got lost in the rush after the search for the magic function (oh yes, there ain't none). The main argument for the NNs is the advancements that we see in "sophisticated statistical and optimization techniques" that do seem to have amazing results.

But, as Adnan said, it seems to be that we're chasing what animals already do well. And, not doing as well as they. The article provides sufficient examples. But, consider this: no animal can make a computer.

Too, there seems to be a huge divide between the true believers here (chasing after money, power, fame, what have you) and those more scientifically oriented. Besides, the whole of the atmosphere has been polluted with out-of-control hype. Adnan uses 'bullied by success' (let's talk some type of innoculation, shall we?).

What is real here? Well, Adnan has the answer. He says that we're dealing with 'function engineering' where one goal is to approximate cognitive functions. There are more. In fact, Adnan's position is strongly influenced by Judea Pearl who recently wrote on the subject (The Book of Why: The New Science of Cause and Effect).

To me, Adnan is talking the old middle-out technique that has been the reality of complicated systems all along. We could think of the NNs as supporting bottom up. But, be aware that there are more areas of technology awaiting attention. The top-down would be somewhat assisted by the traditional approaches. Albeit, we are really talking the human expert rising to the challenge. Per usual.

However, NNs might just be universal as some have claimed from the beginning (not Marvin as he was a critic to his death) in some simulation sense. Guess what? Physics and reality might be closely mimicked by our computationalist's approaches, however being is there irregardless. The 'crap' from extensive computer work (and there are whole bunches to denote) grows. Unfortunately, that whole bit of accumulating residue gets no (or little) attention.

Throughout the paper, Adnan quotes anonymous sources. There is good reason to not get into the cross-hairs at this time. What is really at stake? Well, truth engineering has been dealing with these issues from its inception (my work of the past two decades). So, our pace has (might have) been that of a snail, yet it's tied into other than gaming. Oh yes, we have a whole lot of work to do in reviewing the just past decade and one-half.

Remarks: Modified: 10/16/2018

10/16/2018 -- As Adnan suggested, since there is little scientific thinking going on, I see that we have raging imagination causing the 'running amok' atmosphere. And, we forget the past. This little study has major importance: Carnegie Mellon is Saving Old Software from Oblivion. Those who know have been concerned that experiments (heavily reliant upon computers) cannot be repeated due to technology due to several factors. One is the flash push (wow factor) versus sustainable science. 

No comments: