Sunday, 29 July 2018

Comments on the "A DARPA Perspective on Artificial Intelligence" video

 My attention has been drawn to the following recent video by the web site Conscious Entities which is useful in explaining some of the background to my research.

 This video discusses the history of AI into three stages
  1. Wave 1 AI is based on handcrafted knowledge where experts took knowledge that they had about a particular domain and they described in in rules that they could fit in the computer. The computer was then used to explore the implications of those rules. Applications included ches s. This approach can be very successful when the rules are well known in advance. Such systems have no learning capability and poor handling of uncertainty.
  2. Wave 2 AI is based on statistical learning and has been very successful in areas such as voice recognition and analysing photographs. However it is very dependent of specially engineered statistical techniques appropriate to the particular domain, often using large data sets. Such approaches, while successful in doing the predefined task lack reasoning capabilities. A mathematical techniques, referred to as neural nets, is used to distinguish between different patterns.
  3. We now need a third wave of AI technology linked to contextual adaption to produce systems than build explanatory models which allow them to characterize real-world phenomena because it is clear humans are doing things in a different way.
This is very relevant in explaining the problems I had with my own research.

In 1967 I was asked to come up with ideas as to how one of the most complex commercial applications (a contract procing system dealing with about 250,000 customers and 5000 products) could be moved to the next generation of computers. I saw the market place as posing complex real-world problems and that what was needed was a system that could dynamically interact with the sales staff. I concluded that the top priority was to have a system that the sales staff could understand and control - and this required the system to be able to explain what it was doing in terms that the sales staff could understand - which meant using a common language and having a communication "window" which did not overload the human short term memory. Within the year I have discovered that my "specialist" proposal could be generalized to become task independent, and potentially able to cope with poorly defined and fuzzy information. CODIL (COntext Dependent Information Language) was created. By the early 1970s I found a model which had started as a solution to a complex commercial problem could handle a wide range of tasks including a powerful heuristic problem solver called TANTALIZER.
The problem was I was attempting to model how people processed a wide range of information processing tasks on a computer - so what I was doing must be "Artificial Intelligence." Unfortunately what I was doing was conceptually very different to what the video called "Wave 1 AI" and this made getting finance very difficult. The work finished when a new boss was appointed who was an enthusiastic supporter of 1970s style AI.
Once I had retired I could look at the problem again - and discovered that there are no predictive models of how the neurons of the human brain support our more intelligent human activities. However I then came up with a problem with "Wave 2 AI" as I was interested in networks of neurons and how they work - but this is very different to how A.I. uses "Neural nets" - which is a very sophisticated mathematical technique - needing much human engineering support. The difference between the approaches is such that there appears to be very little similarity between a biological network of neurons and a the mathematically complex neural nets of A.I.
So on to "Wave 3 AI" which will be concerned with reasoning and context - which is what I was proposing in 1967. Perhaps after 50 years my idea will prove to have some relevance.

No comments:

Post a comment