Where did all this come from?

Hi! I’m Jeff Wilson and I believe it will be helpful if I provide some personal background. 

It is probably easier to explain how my technical skills developed if I start in the middle of my story, at what, one might call, the pivotal point of my life.

It was the summer of 1977 and I was driving through France back to England. I had just taken three months unpaid leave from my UK employer Marconi, in order to work for Merlin Gerin, in Grenoble, in France – so that I could improve my knowledge of French and in particular my conversational skills.

Now over 40 years later I still read French novels but my spoken french has somewhat faded.

I realised then that I now had three main technical interests:

• Speech and language
• Electronics
• and Computers

and as I thought about my future I made the decision to focus my energies, insofar as possible, on the technologies of speech and language.

When I got back to work at Marconi, my then boss Paul Dent, later of patent fame at Ericsson, told me that the company had received an enquiry from the U.K.’s National Air Traffic Services (NATS), for an automated Volmet weather announcement system.

Let me quickly describe Volmet: every commercial airport takes weather measurements every half hour and these measurements are sent as coded messages, using a Metar encoding format, to central locations, where they were, at that time, read by human operators onto a tape loop – and then this was broadcast. This approach was both costly and the audio quality was variable given the number of different speakers. All such announcements worldwide are in English – the language of aviation,  although some countries also broadcast in their own language.

I still remember the moment when I reached over and took the enquiry out of Paul’s hands. This was to be my first high-tech project:- I designed a solution, bid it, won the bid, and then build it on time – finally installing the system in the control tower at Heathrow airport where it went live on the 5th of May 1981.

I used a technique called Synthesis by Word Concatenation, although for clarity the term Vocable was used to distinguish between the different spoken forms of each word. With a careful ear and a lot of attention to detail a system was delivered that was truly of human-quality; to the extent that a visitor once asked when our automated system was going live – some months after it had done so!

The system was dual redundant, with an ingenious voice detector that automatically switched systems in the event of any failure.

Perhaps two points are worthy of note about the project:- firstly, given the system’s need for memory and the technology then available, the Vocables were stored in ferrite core memory cards, and secondly, there was at the time, i.e. the late 70s, a belief that Synthesis by Word Concatenation could never achieve human-quality speech. Not only did the Marconi Automated Volmet system prove this to be untrue, but there was also similar work going on at British Telecom, where the same technique was being used for the System X telephone exchange network announcements.

The UK Central Office of Information made a film about Automated Volmet and I and my customer wrote an article on it for the International Civil Aviation Organisation bulletin – published in English, French and Spanish.

Marconi then sold several further systems overseas.

I also gave a talk on Volmet at the U.K.’s Institute of Electrical Engineers in London and one of the other speakers that day was from British Telecom talking about the System X announcements.

By a strange quirk of fate, following an end of seminar conversation at the end of that day, I was subsequently to become involved in the design of start at the beginning network announcers in the UK and other overseas networks. This would eventually lead me to setting up Telsis, with funding from 3i Ventures in June 1987, and Telsis would go on to become a leading international supplier of innovative text and voice services equipment – with many key innovations to its credit.

However, perhaps now is the time to step back to June 1963, when I was at Darlington Queen Elizabeth grammar school in what one would call a fast stream – so O-levels in four years rather than five, and A-levels in six years rather than seven.

Although as my A-levels were double maths and physics, I took and passed single maths at A-level when I was 15.

However to return to June 1963, I failed one of my O-levels … French! This was an unfortunate failure as at that time it was an accepted requirement that an O-level in a foreign language was required for admission to university.

And so I took the exam again in December 1963 and again failed!

At that point my parents took pity on me, paid for a few private lessons and in summer 1964 I finally passed O-level French.

Of course in hindsight there is a funny side to all of this, as I did both of my degrees at Imperial College London and to the best of my knowledge that was the only English university, at that time, that didn’t require an O-level in a foreign language!

There is also the strange fact, that having started to work for the Marconi company as a design engineer, I decided to go to evening classes to study French – more successfully this time as I took and passed A-level French and discovered a taste for French literature. Of course it was my inability to converse easily in French that led me to Grenoble and then to the decision as to my future career focus.

I left school in the summer of 1965 and before going up to Imperial College the following year I spent a year as a telecoms apprentice at AEI Woolwich in London.

As part of the training I was taught machine shop skills and we all did day release at Woolwich Polytechnic. Two memories stand out from this time, firstly the college had a very primitive valve computer, and secondly we did all of our calculations on hand powered mechanical calculators.      

So my career has spanned mechanical calculators, telephony viewed as a natural monopoly, a single MNOS transistor and onwards to today’s smartphones.

And yet in all of that time the problem of Natural Language Understanding was not to be solved until my own key breakthrough in 2016 – following on from my earlier discovery as to the essentiality of Cognitive Memory.

At school I had had three main hobbies, fell walking, caving and recreational mathematics. At University I joined the sub Aqua club for three years and was the diving officer for my last two years. Subsequent recreational pursuits were to be sailing and for a few years, iceskating.

However throughout my adult life I have had a love of languages and I have had lessons in French, German, Spanish and the very basics of Arabic. I have also looked at Chinese and on one occasion used a Chinese dictionary – which is a very interesting experience.

However, I have now started to learn Korean, as it uses a non-Latin alphabet, but still an alphabet – so that a dictionary is easy to use, and I find its use of particles to be fascinating. So fascinating in fact that I was actually able to discover some particles in English – a very unexpected discovery.

In my early years at Marconi prior to my 3 months in France, I worked in three different divisions of the company. These experiences helped me develop my design skills and in particular I learnt to produce tolerant designs that worked in difficult environments, designs that were producible and testable and when it was appropriate – failed safely.

This focus on the security of designs has remained with me throughout my engineering career, and one instance stands out from this period.

I needed a toleranced design of an oscillator, and please remember that this was the era of the transistor and the transistor’s base-emitter nonlinearity. Fortunately I was by then aware of the modified Bessel function and I used this mathematics to produce a toleranced design.

One other transistor fact from this period is that when the MNOS effect was discovered, now used for large-scale memory devices, it first appeared on a single transistor and I had the opportunity to experiment with such transistors.

Following my three months in France and my success with automated Volmet, my remaining years at Marconi were all, by fortuitous circumstances, spent focused on voice technology.

I worked on voice technology for civil, military and telecommunication applications.

The military applications were principally low bit rate speech coding, where I built both channel and LPC military grade vocoders. I also worked on other waveform coding techniques for telecommunication applications.

I also started my association with speech recognition, as I managed a government funded research program on speech recognition in high noise environments.

As a spin-off of this research Marconi developed a standalone connected speech recogniser, the Marconi SR-128. This was the worlds first standalone connected speech recogniser and it was flown for evaluation purposes in a BAC 1-11 civil aircraft, a Wessex helicopter and a Buccaneer jet.

I count myself extremely fortunate that on one occasion I got to fly nap-of-the-earth in the co-pilot’s seat of a Wessex helicopter, as the pilot used an SR-128 to control his instruments by voice.

Clearly all of this took place in the early 1980s and so the technologies available were extremely limited – with limited processor speeds and memory capacities.

However, it was now clear that digital technology was going to greatly enhance the opportunities for new telecom services.

My division produced a number of digital network announcers for both BT and other third party service providers, and the key breakthrough came, in the UK, when BT offered Direct Dial In interconnects to Service Providers.

Now for the first time hundreds of different services could be offered on a telephone line, with the service selected by the number dialled – and for the first time, all such services were start at the beginning.

Until that moment we had had Broadcasting services, now suddenly Narrowcasting had arrived.

It seems almost impossible, even to myself – although I was involved in many of these technical changes, how different the world was then; no mobiles just pagers, no PCs, no Internet, no streaming – it was a very different world from today but the changes were exciting.

Perhaps fortunately from my personal perspective Marconi didn’t want to fully commit to this new future of digital services, and so in the summer of 1987 I left Marconi to found Telsis.

Telsis initially focused on Voice Services Equipment, later expanding into SMS text messaging technology and later still into ISDN video services equipment – and yes, we did it before YouTube!

I was chairman of Telsis for over 25 years and our sales averaged out at over £10m per year, so I’m very proud of our ability to have sold over £¼ billion worth of innovative products and services.

There was so much innovation in this period that although there is a trail of patents, it is now difficult to recreate an accurate historical record and so I will content myself with outlining a few of our innovative steps.

Our first product was Hi-Call, a disc-based Voice Services Equipment with Direct Dial In interfaces in many flavours. The approval number in Germany was 001, as Hi-Call was the first such equipment approved for the Deutsche Telecom network.

The system used dual discs for fault tolerance.

From today’s perspective perhaps the most amazing fact will be that we offered 10 hours of audio storage on a 170 MB disc – which was the largest disc of its type that one could buy in 1987. We also designed, built, gained approval for and supplied the first Hi-Call in under six months.

Over the following years we added a voice activity detector, DTMF tone detection and even rotary dial handset pulse detection capability. As there was, at that time, networks that still had rotary dial handsets – so very, very different from today’s mobile technology.

Hi-Call also offered a live feed capability, so that it was possible to dial in and listen to sports events. Later audioconferencing was added and we built a standalone network-based system that a US magazine rated as best for audio quality.

All of this innovation was to lead us to customers in over 40 countries’

There was one other development – which today may not seem to be especially unusual, but at the time was a sea change in telephony.

Historically, fixed telecoms networks had grown up with two strong belief systems. Firstly that telecoms was a natural monopoly – to the extent that even third-party service innovation was viewed as unnecessary and undesirable – a view that persisted in the USA until the 1972 Carter-phone decision, a copy of which I kept on my office wall for many years.

Secondly, network designs had developed on the principle that call arrivals were random, although there might be busy hour in a day, and so traffic capacity calculations were based on the assumption of random call arrivals.

But then television came into our lives, and one day somebody said on television, Call Now.

Suddenly, there was a high volume of synchronised calls – putting the telephone network under two types of pressure; firstly signalling capacity had not been provisioned for large volumes of synchronised call arrivals and secondly a network had no way of terminating such a large volume of calls.

Telsis worked with several networks on this problem – but I have to mention something that perfectly captures the rigid mindset of this period.

We had a meeting with a large European fixed network operator and explained the Call Now issue and the potential to both improve customer service and create new revenues.

Unfortunately, when the operator realised that it had the only network suitable for such a service in that country, it decided not to proceed – perhaps because this was a period when networks didn’t have customers just subscribers!

As the need for more complex network services developed, there was a focus on what were called Intelligent Network services, and therefore Telsis developed an Ocean family of Intelligent Network components.

There were three main components, a Service Control Point- or SCP, the brains of the system, a Service Switching Point – or SSP, for call routing, and an Intelligent Peripheral – or IP, for terminating calls.

Almost uniquely for that time our Ocean SSP was nonblocking, it also had extremely high level of fault tolerance with dual clocks and the ability to work unimpaired in the event of any single fault. It was even more impressive in that it could in itself terminate calls and provide services.

Services were written in a safe C programming variant called Nodal, a language created by Telsis that provided the benefits of C but which was safe. Some other suppliers used C – but then bad code could crash a system – this wasn’t possible with Nodal.

By this time Telsis had its own SS7 signalling stack and was able to offer extremely robust service solutions.

Mobile Telephony had now arrived and so there was a new communications channel, the Short Message Service i.e. SMS, which was to become a highly successful addition to the voice channel.

SMS was added in to the mobile standards almost as an afterthought, yet it went on to become a huge success and Telsis developed a range of products to both route and terminate SMS messages.

One of my key patents is from this period as I invented SMS Home Routing.

This network enhancement had two key benefits:- firstly, it greatly improved security, as the SMS channel was also used for service updates, and secondly, it provided the perfect way to provide new value-added text services.

As this was such our fundamental enhancement, it Is worth a moment to explain the issue.

In the classic GSM architecture used in Europe, most SMS messages sent to a mobile phone would not transit the customer’s home network. So the GSM default architecture prevented a mobile operator offering their customers value added text services. However SMS Home Routing provided the perfect solution for a wide range of value-added services.

Of all my innovations over this latter period, 4 of them standout.

Firstly, SMS Home Routing, as described.

Secondly, a simple to use audio service called Radio Choice.

This is a simple to use audio service that provides time shifted access to any radio program in the last seven days. The key benefit, still not replicated in today’s approaches, is that one can listen to any program from the beginning even if it had started just a few minutes ago. This is great if one drives a lot as now it no longer matters if you get into your car 10 minutes after your favourite program has started.

Thirdly, Mobile Assistant – this was I believe my last patent, filed in 2012, when I invented a network architecture that could provide all the traditional fixed line boss secretary features in a mobile environment. As an example it allowed the secretary, in a mobile environment, to answer calls to the boss and also call transfers could be done in either direction at any time.

Finally, the Telsis ISDN video services equipment was a real breakthrough in enabling ISDN video à la YouTube.

Unfortunately through a lack of network interest in ISDN, the limited video bandwidth and the crazy pricing of ISDN calls, video services only finally blossomed when broadband data arrived.

The lack of interest in the customer’s experience is still something that seems to pervade parts of high-tech industry, driven perhaps by strong belief systems.

In 2014, I started to work full time at Precise Intelligence Ltd – and in 2014 the company was called Solent Press Ltd, a name chosen for anonymity.

I started to research the problem of Natural Language Understanding (NLU), and as this is essentially a mathematical problem, I would refer to it as the Geometry of Language (GoL) – as it was clear that this was a problem in geometry.

I had learned during my research into speech recognition that it can be easier to solve a hard problem than a simple problem – for the simple reason that with a hard problem the solution space is more constrained.

The beauty of mathematics is to find simple solutions to seemingly complex problems. My first step, unlike that of many other researchers, was not to try to simplify the problem but to simplify the solution.

Many researchers have tried to simplify the problem of Natural Language Understanding by ignoring anaphors, common sense, context, explainability and even statements, which are critical to human-quality conversations.

The problem with this approach is that one is then not solving the actual problem but a subset thereof – such that any apparent solution will not be relevant to the real world problem.

The first phase on my research was a review of the state-of-the-art, both from a theoretical and practical perspective. This was a dispiriting exercise as Chomsky’s analysis of language seemed to be overly complex, to the extent that at some point in these early years I wrote a one page article on the subject of ‘simple’. 

However, not only did language models lack a sound theoretical basis, but also much research seemed to be of the ‘mad scientist’ variety, with lots of code but no mathematical theory and little attention to the critical issues.

The phrase that seemed to capture the seeming insanity of this period was ‘bag of words’, and if you don’t know what this refers to, my advice is don’t bother to look it up.

Fortunately, by disregarding previous work, I was finally able to identify the invariant features of language so that I could classify sentences, and I use the term Abstract to refer to this classification.

As to how many Abstracts exist, in the near infinity of possible sentences, is as yet unknown, but I suspect that there are less Abstracts than there are distinct Chinese characters.

One Abstract I uncovered seemed to be especially complex, and it was to take me six months to fully understand it – I then realised it was simple!

However, that is the nature of research in mathematics. One must attack problems by searching for insight and then developing terse and powerful forms of representation so that the way forward becomes clearer.

Now that the discovery of invariant Abstracts allowed the solution to Natural Language Understanding to be simplified, the company’s research variant of an AI Assistant, HodgePodge Data (HPD), was used to investigate issues of context in question answering.

To ensure I work with human-quality language, I use short conversational tests, which I call Smoking Gun tests, as these provide a focus on context. One of the issues with traditional linguistics has been an unhealthy focus on single standalone sentences.

There were two key issues with conventional AI research at that time, and unfortunately these issues seem to have been carried forward to this day. Firstly, confusion between form and meaning, and secondly, the seeming invisibility of statements.

Statements are a key component of human conversations and hence understanding statements, which require the creation of unstructured cognitive memories, Is critical to Natural Language Understanding.

Fortunately, I discovered The Fundamental Theorem Of Cognition (TFTOC) in 2016 – and this powerful yet simple concept can, like E=mc², be expressed in just five symbols, but it was to take me five years to fully understand all the implications.

A partial implementation of TFTOC was then added to HPD, and by late 2019 it was possible to carry out a large number of Smoking Gun tests – which verified that TFTOC is essential for Natural Language Understanding.

One question, that may then naturally come to mind, is why hasn’t this new mathematics been discovered previously.

I think the answer is that prior to my discovery of TFTOC, mathematics had no concept of mind and no concept of cognitive memory.

Essentially a mathematical problem was always seen from a single third-party perspective.

Yet when one looks at a conversation between two people who stop and talk to each other in the street, the meaning of a sentence spoken by a third person will clearly be interpreted differently by each of the two individuals – as its meaning will depend on each individual’s cognitive memories.

Following on from the first generation of TFTOC code in HPD I have continued to work on two key research areas:

1. Representation – given the novelty of Abstracts and TFTOC, it is  important to have terse and simple representations – so this is an ongoing activity.
2. Language theory – given the novelty of the new mathematics it remains important to continue to simplify it.

As an example of my ongoing research into language, I found, in 2021, a  further key mathematical insight which explains many of the features of sentence structures that have baffled linguists.

I believe my explanation of the importance of Cognitive Memory, statements, common sense and explainability, plus the demonstrations of capability from the many Smoking Gun tests prove that my research provides the right way forward.

From my own perspective, from inside the citadel so to speak, I can confirm that TFTOC is essential and therefore necessary to anyone who wants Natural Language Understanding.

My career has spanned a fascinating period and I like to think my intelligence found the way forward.

However as I look around at the literally piles of paper at home and in my office, I am reminded of Edison’s comment about trial and error.

The combination of mathematics and language has been and remains a most amazing journey