This year's Turing lecture was special. It was a complex talk by one of the world's leading neuroscientists during which he attempted the high-wire balancing act of explaining aspects of the cognitive process while paying tribute directly to Alan Turing. In the lecture he aimed to shed new light on the ways Turing's towering intellectual and creative work is still contributing to one of the most important fields of research in science, if not the most important. It is hard for me to be objective since I hung on every word of the lecture. My notes here are liberally scattered with my own observations and thoughts as well as those of Professor Dolan. It was a lecture which dealt mainly with cognitive neuroscience, for me the most interesting subject possible and one that has consumed and inspired me since I began my own investigations into its history, discoveries, and heroes. It also dealt with a person who is a hero by so many definitions: wartime hero, human hero, scientific hero, intellectual hero and almost any other definition of hero you care to throw at him. The lecture was gracious, informative, and beautifully lucid. It had the satisfying sound of a tennis ball well hit.
a map of the world on the ceiling of the lecture hall
Modelling Professor Dolan
I wonder how it feels to be the most cited neuroscientist in the world. I try to imagine what it could be like to be here tonight lecturing the IET and the BCS on the most prestigious of occasions – the 100th anniversary of Alan Turing's birth and the 14th Turing lecture since they began in 1999. This may be the special year for Turing as far as the rest of the world is concerned, but the IET/BCS honours this unique genius with a lecture in his honour every year. Every year we invite someone on the leading edge of science to share their insights with us. In the lectures I have attended this has always been someone of whom, one suspects, Turing would have approved mightily. Professor Dolan certainly fits the bill.
As he stood there talking in his quiet authoritative voice it felt like part of my own mind was talking to me about the most interesting things in the world. He was giving the talk based on experience and his life's work. He job is the dream of philosophers and thinkers through the ages. His job is to pierce the vale that shrouds the inner sanctum of the mind and consciousness itself.
His lecture wound through the physiology of the brain, around the synaptic structure of cognitive perception, over information, belief and sensory input and arrived back at Turing and his magnificent mind.
The early appearance of a-priori
In my book “Managing Knowledge: Apply the discoveries of neuroscience and understand the human factor” there is a chapter called “Arm yourself with a-priori knowledge”. I had worried about this chapter title when writing the book. I was not advocating a-priori as a set of immovable beliefs - quite the contrary. Delusional belief is one of the world's major problems. The delusional a-priori belief that homosexuality is a perversion was the weapon used to drive Alan Turning to his death.
In writing that chapter I was pointing out that our model of the world is, by definition, a-priori. A-priori is what we think we know and the things we believe and assume. We filter information and experience through this model of the world. It colours our relationship with reality. If this is the case we need to be very careful of the a-priori knowledge we carry around with us and make sure it is open to redefinition. The more inflexible it is, the more distorted by generalisations or deletions it is, the more it distorts incoming information. The more exercised it is, the more it arms us to avoid delusion and inappropriate certainty. Too much certainty in our a-priori position halts progress and prevents the production of a-posteriori – or learning. The scientific process is one of constantly updating our a-priori hypothesis and theories based on incoming information - in other words managing it.
Bearing this in mind will help you understand why this speaker, his lecture and his subject gripped me from the outset and why I break out and wax lyrical as we proceed.
Alan Turing, enigma machines and more a-priori
Turing famously cracked the enigma code during world war II. An important fact to understand is that, in order to communicate, the receiving machine had to be in the same setting as the transmitting machine. The Nazi's were so certain of their a-priori belief in the strength of the enigma machine that they ignored signs that the allies had broken it. Dolan mentioned this almost in passing in his opening preamble and I wondered if he understood the significance of this statement. Of course he did! As he talked about uncertainty, inductive inference, and evidence accumulators, my neurology continued to light up like a small city in my head. All my a-priori jostled to the front and waited to become a-posteriori - either proven or redefined.
Cryptanalysis and Bayesian inference
Dolan started by establishing his connection to Turing. He told us he was appointed consultant in 1986 to the National hospital for neurology and neurosurgery - home of neurology – and that its basement had hosted the Ratio club (which has been mentioned before on these blogs). Turing and his friends were interested in cryptanalysis. This lead Turing to the enigma work and breaking the unbreakable code. They were also interested in the brain and how they might build one. This led to the birth of modern computing and is leading us toward Artificial Intelligence of one sort or another. It was also an important parent of neuroscience as we shall hear.
While my mind reeled from the reflected genius of Turing the polymath, the ghosts of Bayes and Pierre Simon Laplace were being conjured up on stage. They were needed to provide the basis of inductive probability – combining uncertain data into a new belief. (Bayes has also been mentioned and commented upon in these blogs before). .
The following principles were discussed:
Inductive Inference - making inferences based on uncertain data provides a mathematical means of deciding if new data supports the computing hypothesis – given the nature of the enigma machine this was essential to Turing.
Conditional probability – you have some evidence that will effect the outcome.
In other words prior beliefs can distort how we see the world. Uncertainty describes how much faith you put in sensory input or a-priori. At this stage I was in heaven. My whole first book and a large part of my second rests on this theory.
Professor Dolan talked about how Turing needed to quantify hunches because much of the decoding depended on acting or not acting on hypothesis. He explained that the enigma code had to be cracked every day – it was not a one off event. For that process Turing had to have some way of judging when his hunches were solid enough to rely on. Turing and Jack Good came up with the idea of a ban or deciban which is the measure of the weight of evidence in favour of a hypothesis or hunch, The smallest change in weight of evidence accessible to human intuition. This started leading us back into neuroscience.
The brain decodes the state of the world by implementing statistical procedures very similar to those used by Turing and his colleagues. We entertain models of the world that are susceptible to Bayesian comparison in our brains. In this way it chooses the model that carries the highest degree of confidence. From a Darwinian perspective brains that use a Bayesian perspective on the world are more likely to be selected by evolution.
The following two ideas are so central to the talk that they were presented on slides:
“A Bayesian perspective provides a means of starting with a theory of how parameters produce data and inverts it into a theory of how that can be used to reveal the parameters that caused it”
“A central concern in systems neuroscience is to formulate precise models of the brain's model making processes.”
Then this paradigm changing thought from Alan Turing from 1950 – "the mind, rather than the brain itself, is the equivalent of the pattern of information processing supported by the brain ."
The brain makes sense of the world around it by decoding sensory input using Bayesian inference.
He illustrated his point with some visual illusions and by showing us a field of dots in which only 30% is moving in a constant direction and the other 70% is moving at random. He was explaining the process by which we decide whether the constant motion is to the left or right by looking for evidence to support one way or the other. The brain's evidence accumulator, in this case a structure at the back of the brain that detects motion either to the left or the right, examines the incoming information. As we find evidence to support one hunch we find it easier to find more evidence to support it until we reach a critical mass at which point we can make a decision.
If the field contains only 10% moving in a single direction and 90% moving at random there is more noise. The process takes longer but the critical mass is the same. (a certain number of Neurons firing).
The more noise there is in the data the more difficult it is to gather evidence and the longer it takes to build to critical mass. This is called the speed/accuracy trade off in psychology and should be familiar to anyone asked to make a difficult evidence based decision.
This was important to the code breakers because they had to know when to come to a decision in their analysis of the enigma code. They had to come to a decision quickly but not so quickly as to get it wrong.
The analogy is that sensory decoding and cryptanalysis both deal with:
weight of evidence for competing hypothesis
criterion for deciding between the the hypothesis
Uncertainty and perceptual experience
The next part of the talk resonated so strongly that I am not sure I did not start muttering to myself animatedly. It was all about pain, prior information and uncertainty.
In an experiment subjects were given prior information about how painful or not a stimulus was going to be. They were given the same stimulus and asked to rate the pain they experienced against the pain they had expected.
People who had been primed to expect a lot of pain felt more pain than they had expected. People who had been primed to expect less pain experienced less pain than they had expected. In fact those who were anticipating pain felt pain even before the stimulus was applied.
Uncertainty was also manipulated to extend the experiment – that is the uncertainty given to the information with which they were primed. It was manipulated - it was not clear to me how this was done but I assume doubt over the prior information was introduced in a controlled fashion. As uncertainty increased the reaction to the pain stimulus increased. Uncertainty undid the effect of the prior information. i.e. the more uncertain they were about the a-priori the less effect it had. There is a Bayesian avalanche going on here somewhere – the implication is that we are Bayesian inference machines.
The neural responses were also observed with fMRI and I was not surprised at the results (you will find out why in a minute – let's just say my Bayesian evidence accumulator was well primed).
So what does all this tell us? Let me try to explain what it told me.
The opening assertion was that we use prior beliefs to filter sensory information.
if you believe something is going to hurt, you pay more attention to the evidence that proves this. It seems that you may even invent evidence so that you experience more pain to prove yourself right. You filter out other information as noise while you strive to prove your a-priori belief that you will be hurt.
The converse is also true. If you believe you are not going to be hurt you are biased to the information or sensory input that proves that model of a-priori belief.
Professor Dolan used the example of a doctor who tells you an injection will hurt – it hurts like billy-o. If the same doctor first rubs on some cream that he tells you is anaesthetic, then the injection does not hurt, even if the cream is an inert placebo. Here we have the placebo response. He said the placebo RESPONSE not the placebo effect. I was ecstatic – it is not a random effect it is a response that can be elicited. It is all about how you weight the a-priori against the evidence.
(My primed a-priori is years of watching doctors telling my daughter that this will sting or that will hurt. I am blue in the face pleading with doctors and nurses to stop priming her for pain. I have even had nurses arguing with her that something must have hurt when she tells them it does not. Would that I could carry a portable Professor Dolan with me to show them the brain scans, graphs and charts of his experimental proof )
A statistical belief is a probability distribution over some future state. An optimal belief is one which is open to experience and updated on the basis of evidence/experience. A delusion is the pathology of a belief not susceptible to change.
But what is a belief in the brain? What is belief to a neuroscientist?
Daniel Bernoulli, an 18th century mathematician, was called to the witness stand, and, not for the first time, these hallowed halls were introduced to gambling as a way to explaining uncertainty and belief. Bernoulli decided that the best gambling strategy is to look at the magnitude of the outcome and multiply it by the probability to determine expected value. This means that you have to lay your bets according to the size of the reward/prize and your likelihood of getting it. It appears our brain operates this strategy and it uses Bayesian inference to work it out.
I have avoided getting too deeply involved in Professor Dolan's elegant explanations of the physical neurology of the brain but I have to talk here about Dopamine and its part in encoding belief. Belief is built up by using Dopamine to report the degree of surprise in the environment (how much the sensory input is diverging from your a-priori).
Imagine you are gambling and you simply have to bet on left or right. You know that one direction will pay out 80% of the time and the other 20% of the time but not which. Say you start with a weak belief that choosing left is the 80% route because you got a win there first time but it could have been a fluke. Now each time you choose left you win. The more you win the more your belief strengthens based on feedback. Your brain will get dopamine feedback relative to the accuracy of your predictions of the outcome. In this way dopamine is encoding a belief in a quantitative fashion. Essentially we update beliefs based on the degree of surprise in our environment and the dopamine level reports how well your expectations match the outcomes.
Again there was fMRI evidence and the dopamine release was proportional to the degree of expectations matching outcomes – in this way dopamine weights our predictions of the future based on past experience- this is belief encoding. We physically encode our experiences into beliefs. This is a constant feedback mechanism. Working healthily it should base our beliefs on our actual experience.
What this confirms to me and gives me a great big dopamine hit, is that we become what we do. If we surround ourselves with only evidence that confirms our model of the world our beliefs become deeper and harder to shift and we stop recognising facts that might challenge this view. What this tells me is that it is necessary to challenge everything and avoid that rut at all costs.
Models of the self and others
The next witness was Phineas Gage – he was a railway worker in the 1800's who survived having a spike driven through his head destroying much of his left frontal lobe. While much of his mental abilities remained largely unaffected he no longer knew the value of anything.
Professor Dolan used Phineas to talk about the phenomenon of models. We have models of ourselves and models of other people in our brain.
How do we create value models of other people? What do we base them on and how do we use them?
It seems that we infer them based on our observations. This lines up with everything we have heard so far.
When we are making decisions involving other people we invoke the models we have of their values. fMRI slides showed that when we are making decisions based on our own values we also light up those parts of the brain that store other people's values whether they are relevant or not. We do the same when making decisions for other people. Our values are lit up whether they are relevant or not. Essentially we have the ability to predict and model other people's values even if they are different to our own and access them as mental models.
This was really about empathy – in Dolan's opinion sociopaths do not have the ability to model another's pain. This ability to model is very sophisticated and among our higher cognitive functions.
Inferring what goes on in the brain (a mathematical microscope)
The final part of the talk was about using this knowledge as a mathematical microscope on the brain. Can we decode what is happening in the brain using the mathematical tools we have been discussing.
Can we look inwards and decode what is going on in the brain – what are the causes of the critical functions of the brain. This was a discussion around synaptic gaps, neurotransmitters and our old friend, Bayesian inference.
A useful piece of information was that memory performance increases when you are given L-Dopa.
However, the critical point is that Bayesian inference can be used to understand what is happening in the brain at a level that we cannot observe. It does this by looking at the production of certain neurotransmitters.
Right through this talk professor Dolan linked the work of Turing and the codebreakers to leading edge neuroscience. The brain is an inferential machine that can be used to look into itself. He finished by speculating that if Turing had survived that he would have gone into neuroscience since that appeared to be his real interest. He called for Turing's contributions to neuroscience to be recognised.
When professor Dolan stopped speaking I felt as though sensory evidence was flaming along the pathways of my brain.
I looked down and found I had about 20 pages of notes, thoughts and observations. I was bubbling with excitement and the audience was being invited to ask questions. I could not figure out where to start and no question could encompass all that I was thinking. I was quite aware that if I opened my mouth, too many thoughts would try to spill out.
The enigma machine had to have the transmitter and receiver with the same settings inorder to communicate – just like brains modelling each other in empathy.
The Nazis were so certain of their a-priori that they missed signs that the allies had cracked enigma – a certainty, a-priori immune to change, is a delusion. Bayesian inference interrupted.
I was reeling from the proof of suspicions, theories and heresies that are now main stream science. I felt that this remarkable scientist and his colleagues had been working to prove the assertion that neuroscience is unlocking the deep structure of consciousness and what it is to be human. It is teaching us how to manage our beliefs and how to make better decisions.
I made my way around to my countryman when the ceremony of the occasion had played out and I thanked him for his wonderful talk. When I saw him at the post-talk networking he was deep in conversation with someone and I was deep in conversation myself with a charming retired computer scientist. Time beckoned and I strolled out onto the embankment toward the tube station to catch my train back to the midlands.
The embankment, someone told me recently, is actually built up around one of the major sewers in London and one is of the major feats of civil engineering of the 19th century. Tonight it was full of people and lights. How many of those lights do we owe to Turing I wondered. How much do we owe to his wartime work and his contributions to the logic that enables the modern world to exist.
On the train I was looking through my notes and wondering how I would write all this up. I was lost in my thoughts. The only announcement I heard was when we passed through Bletchley.