Saturday, 9 April 2016

The Link between Codil and the Evolution of Human Intelligence

The following notes have been prepared for the Leo Computer Society Reunion on April 10th. It is a brief summary describing the link between the original observations made of how salesmen understood complex sales contracts at Shell Mex and BP and later research in CODIL. It also outlines the recent reassessment which relates the earlier work to neural nets and the relevant to research into human intelligence. If you want to know more do not hesitate to contact me.

The Surprising Connection between the Leo III and Research into the Evolution of Human Intelligence
by Chris Reynolds
10th April, 2016

In 1967 I was asked to look in detail at the Shell Mex & BP sales accounting programs which ran on their Leo 3 computers at Hemel Hempstead. The aim was to see how they might be moved to the next generation of computers – which would have computer terminals. The result was a proposal for a computer with a user-friendly symbolic assembly language called CODIL (COntext Dependent Information Language). Nearly 60 years later it is possible to link my original observations on how salesmen though about contracts with a model of how human intelligence may have evolved. This note briefly explains the link and suggests that some way should be found to re-vitalize the research.

The 1960's Shell Mex & BP sales accounts were processed by a batch processing computer system using magnetic tape as storage and paper tape as input. Its customer contract files were held on tape but were printed on about one million record cards and there were about 5000 different brand codes used for sales of a wide range of products from liquid propane gas to tar for roads. Customers varied from individual households to major organisations such as British Rail. The program I worked on was written in Intercode [The symbolic assembly language for the LEO III computers] and was probably one of the most complex commercial sales accounting programs then in existence.
I was a comparative newcomer to computers, having previously worked as a human cog in a complex international management information system. In addition I spent a year learning to program in Cleo [A high level language combining features of COBOL and Algol used on the LEO III computers]. On the day I arrived they had upgraded the system – reprinting all the customer record cards – and things had gone seriously wrong.  As a result I became very interested in computer "bugs" and how to program to minimize the dangers.
In considering the plan to move the application online there were no clear guidelines as no-one had every tried to do anything on this scale before. I proposed a system that the sales staff could understand and directly control – which meant that the system should be able to explain, in salesman language, what it was doing. I honestly had no idea that anyone would imagine this was difficult! Using my extensive background in complex manual systems I suggested a way in which the system could use the same concepts the salesmen used. A concept could be the name of a customer or brand, a quantity (goods or money), a date, the name of a standard contract, etc. etc., I found that all contracts could all be reorganized into one or more short easily understood lists and a contract-understanding interpreter could be written to work in a way that the salesmen would understand.
In retrospect what I had accidentally “discovered” was that the human short term memory cannot handle more than about seven concepts simultaneously and used this fact as the basis of a two-way language which reflected how a salesman's short term memory would handle a sales contract. When I outlined my ideas I was told that salesmen were not able to program computers so it was impossible. Despite this objection the new accounting system used “variants” based in part on my proposal.
I was not involved in this as I was “head-hunted” to join John Aris's group within English Electric Leo Marconi, working with George Stearn on market research into the next generation large commercial systems. A few months later I suggested that my proposal could be mapped onto CPU hardware targeted at non-mathematical open-ended data processing tasks which needed good mutual human-computer understanding. I was supported by David Caminer and John Pinkerton (who I understand consulted Professor Wilkes) and was given a generous budget to program and test a simulation. This worked better than expected. Unfortunately the project was chopped because ICL was formed and such innovative ideas were incompatible with the plans for the 2900 computer.
In retrospect a more serious problem was that commercial secrecy had limited discussion about the theoretical foundations of what turns out to be a significant paradigm shift and too much emphasis was placed on getting the rather crude initial idea working – without properly understanding why it worked.
After a short break the project continued (basically unfunded) in 1971 at the newly founded Brunel University. In retrospect an environment dedicated to teaching existing technology and with no experience of supporting “outside the box” thinking was the wrong home for the project. Despite this a range of applications were explored and some interesting work was done in the Artificial Intelligence field. From 1980 the CODIL simulator was supporting significant online teaching packages and other applications. In addition an A.I. demonstration package [MicroCODIL], written for a BBC computer (in part to show that the key parts were small enough to fit on a single chip) attracted very favourable reviews in publications as different as the New Scientist, the Times Educational Supplement and The Psychologist.
Despite this success, active support for the project was withdrawn as part of Maggie Thatcher's plan to streamline universities (see article in ITNow,June 2015, for background). Research was abandoned and I retired from academic life to do voluntary work to help the mentally ill.
Nearly thirty years later retirement gave me the freedom to do the missing “blue shy” thinking about the underlying model and the following notes very briefly summarize my provisional findings.
  1. CODIL was a bottom up system which dynamically “learnt” from the human users in marked contrast to the stored program approach which requires someone to create a comprehensive predefinition of the complete task.
  2. CODIL can be considered to model a neural network where the concepts (user supplied names) are nodes and a table was used to record the links between the concepts. (The neural net aspect was not recognized at the time.)
  3. The research demonstrated that CODIL could support a wide variety of potentially open-ended tasks – including many in the artificial intelligence field. It therefore provides a working model of how a neural network can morph from simple pattern recognition, via set processing, to handling complex rule based tasks, possibly including supporting natural language. Further research may show that Turing's Universal machine model is a proper subset of an even more general neural net model of information processing.
  4. It is useful to think of CODIL as a language designed to transfer information in both directions between one neural net (the human brain) and another (the CODIL computer). While it is not "natural language" it provides an approximate working model for how information is transferred from one human to another.
  5. The recursive nature of CODIL provides a staircase by which human intelligence could have evolved from lowly beginnings. It suggests that normal animal intelligence is limited by what an animal can learn in a lifetime. Once a primitive (CODIL-like?) language made it possible for significant amounts of information to be passed verbally between brains, there was a critical tipping point leading to accelerated learning – meaning that much more could be learnt in a lifetime. This model suggests that most of human intelligence resides in the culture passed from one generation to the next and this has expanded at a steadily increasing rate as the centuries go by.
  6. The simple CODIL-like neural model I am developing reflects neuroscience and psychology research such as concept cells and the role of dreaming as a learning aid. It also explains some of the well-known weaknesses of human thought – such as confirmation bias, an unreliable long term memory, and a tendency to follow charismatic leaders (often religious) without checking the soundness of their ideas.
This all looks very exciting. There is a very obvious black hole in current brain research as there is no published model that explains how the brain's neural network supports observed human intelligence. This weeks New Scientist highlighted the problem: “The mystery is, and remains, how matter manages to give rise to thinking, imaging, dreaming and the whole smorgasbord of mentality, emotion, and intelligent action.” What my assessment of the CODIL project suggests is that the unfortunately abandoned research project provides the starting framework of a brain model which not shows how a neural net can support intelligent actions but also suggest a pathway by which human intelligence has evolved.
We come to the Big Question. Could a 50 year old study of how salesmen thought about sales contracts really be the answer to the mystery of how the brain works? Could government inspired plans to streamline the computer industry (by creating ICL) and by making university research more commercial have led to a promising counter-intuitive blue sky research project being junked? Might some counter-intuitive ideas of a 78 year old pensioner possibly have relevance today?

If you think my ideas deserve more attention please contact me via where you will find copies of the original CODIL papers and my more recent ideas. At my age it is not practical for me to restart the project on my own, but I would be happy answer any detailed questions and to advise anyone who would like to explore these ideas further.
Over the next month or so I propose to add a number of short reports looking at different aspects of CODIL and its possible relevance to modelling human intelligence and its evolution. I would welcome and comments, suggestions (with references where possible) and questions, as this will allow me to concentrate on those aspects of the research which people feel need addressing. 

No comments:

Post a Comment