This is a continuation of Part 1.
Machine Learning (ML)
AI includes a wide variety of subfields, but recently the terms ML and/or deep learning (DL) have been used interchangeably with AI. In our interview, Dr. Kunio Matsui mentioned that a boom for ML comes every 15 years. Back around the year 2000, ML got a lot of attention. Then in 2015, there was another boom. I used the past tense because it is 2016 now. Matsui referred to Gartner’s Hype Cycle 2015. He pointed out that ML was no longer at the peak of inflated expectations. I checked it carefully and found ML is slightly off the peak. Matsui said that depending upon the use/area/application, the boom could last another 12 months or longer. But he said that it was inevitable to have another AI winter.
Suitable Areas of Applications
Is ML a panacea for each and every area of problems? How does ML work?
Because I asked Matsui to keep his answers at a layman’s level, they were easy to follow intuitively. He described how ML worked, using a simple example. Suppose we have many beads and are to put each bead into one of a certain number of boxes according to some predefined criteria. If the number of boxes is small, it is easy to put each bead into the appropriate box. But if there are a huge number of boxes, it is not that easy, and sometimes impossible, to do so. Classifying emails into spam and not spam is relatively easy because there are only two boxes (spam or not spam). Matsui said that for NLP, there are far many more boxes than we have spam detection for, and it is usually hard to apply ML to that.
In the case of spam detection, performance has improved dramatically in the past 15 years. He said this was because of the availability of a large number of examples, on the order of hundreds of millions. Of course, the increase of computing power is another factor in better performance.
ML Bottom Line
He summarized three conditions in which ML works well:
Previous experience data exist
- Some well-defined rules exist
- The scope is well defined
I draw the conclusion that ML is not necessarily a panacea for every problem. Learning in ML still needs to be programmed prior to execution. On the other hand, humans can tackle something they have never encountered before with innovation and intuition, although previous education and experience may help. In other words, if we prepare everything (including the scope of the data sources to use and specific models of how to learn), ML can do very well and may actually do much better than humans. It is very important to understand which areas ML can and cannot be effectively applied to.
I asked Matsui about DL, which some consider to be ML on steroids. He said DL is a subfield of ML and has more power than normal ML. But even so, the bottom line is the same. Matsui said that NLP usually has far too many boxes, and it is very hard to apply ML effectively to NLP. It is very important to combine both rules and ML/DL for NLP.
Some researchers and business leaders have expressed their concerns over the danger that AI could take over the world and harm humans. This may be attributed to the fear that strong AI/artificial general intelligence and the 2045 Singularity may become reality.
I asked Matsui if he worries about this problem. He was very clear on this subject and did not think AI would ever surpass human intelligence. I asked him twice if he really meant it. The answer was a resounding NO. In any event, I do not have to worry about this in my lifetime. In 2045, I may be around, but I am not sure that my mind will be sharp enough to understand what’s going on around me.
In the end, he listed three researchers (Japanese nationals) in the area of NLP who will lead this research area:
- Professor Kentaro Inui of Tohoku University
- Professor Sadao Kurohashi of Kyoto University
- Professor Satoshi Sekine of New York University
Dr. Matsui was very nice to answer my layman’s questions, which may have sounded silly to an expert like him. I often think that those who know the most can explain hard concepts most simply. That was exactly the case.
AI has been researched for many years and its scope is very wide. Anyone can write an article after reading five to ten articles on the subject. But it is very hard to grasp a succinct idea of what it is, where it stands now, and where it is going. I was very fortunate to be able to speak with Dr. Matsui. Based on our conversation, I conclude the following:
- ML can be applied to many areas, as long as the aforementioned three conditions hold. We have not exhaustively looked into its applicability yet. There are many vertical areas and subareas for consideration. The current AI/ML/DL boom will continue for some time until this exercise is completed.
- Those who successfully apply this will be strong proponents of AI, while those who cannot will downplay AI, leading to the third AI winter.
- The combination of well-defined rules and “brute force” (ML/DL) will allow further progress in NLP and other areas. When this is pushed until no more progress can be made, a new approach to AI will be considered and further advances can be made.
- I am not sure if the Biological Neural Network advocated by Jeff Hawkins will be the next stage of research, but it is certainly a good read, and I highly recommend your spending your time on it. The blog is not highly technical. Rather, it is written for the average Joe. Hawkins listed traditional AI (expert system), neural network (NN, including ML/DL), and biological neural network, which he advocates.
Finally, I would like to thank Mr. Hitoshi Matsumoto, Executive Advisor, Fujitsu Laboratories of America, Inc., for introducing me to Dr. Matsui.