171 | Bayesian Brain Surprises
Colombo, Sri Lanka
The audio for this post is available. It’s rough, but listen if you would rather. As I say below, I’ll be experimenting with this medium 🎤🙌
Friday 26 January. I know nothing about Sri Lanka and can’t yet comment about its people, history or culture. I arrived at 5am, it’s now 10am and I’ve eaten and walked and drank some coffee and worried about how smashed up my bicycle is in it’s cardboard box. We taped it up good, but one of the axels sticks out errantly: if it’s broken it’s a problem. I don’t yet know the vibrations of Colombo, perhaps I’ll go out tonight — it is Friday.
Before we kick off with this week’s missive about our individual hallucinations, I’d like to share Emerson’s position on seeking. It reminded me of some of you.
Let me remind the reader that I am only an experimenter. Do not set the least value on what I do, or the least discredit on what I do not, as if I pretended to settle any thing as true or false. I unsettle all things. No facts are to me sacred; none are profane; I simply experiment, an endless seeker… Yet this incessant movement and progression which all things partake could never become sensible to us but by contrast to some principle of fixture or stability in the soul.
You’ll note that I’ve recorded the audio of this week’s note. I thought it would be cool to do some audio recordings of the trip — and add these to my weekly updates. I might experiment with conversations at some point, too.
Although it’s so long ago that the details come only faintly, when I was a foetus, doctors tested me for some rare disease. The test was 95% accurate and came out positive, at face value, suggesting I would be born with the disease. The pregnancy was early, and the doctors said it would be better for me in the long run if I was terminated.
This was devastating for my parents to hear.
Before the test was done, the initial probability of any newborn having the disease was ~1 in 700 or 0.14%. Because the test is not 100% accurate, the original conjecture that I would have a ‘95% chance of being born with a disease that affects 0.14% of foetuses’ is misleading, although technically true. Because the test was inaccurate 5% of the time, Bayes showed that the likelihood of me having the disease amounts to only 2.65%. Furthermore, if the test were 99.9% accurate, the chance of the foetus having the disease would be 58.8%, high but still not as high as I would have expected.
With a positive result printed out and in their hands, my parents could use the new information to adjust their prior expectations. My parents decided to ignore the test, the chance still being quite small, so here I am.
Bayes’ Theorem takes an original guess (‘Prior Probability’) and updates the guess depending on new information, making the prediction more accurate. With the test result, my parents could see it was more likely that I would have the disease, but because the test was not 100% accurate, the probability of me having the disease was still low.
Whether my parents knew about Bayes’s Theorem or not, the theorem has many surprising implications. Bayes’s Theorem explains how we interact with the world: welcome to the Bayesian Brain hypothesis.
It’s hypothesised that the brain creates the world by forming probabilistic models. The brain recursively updates the model as new data is revealed.
We don’t interpret all sensory data (sight, smell, touch, etc.) when it arrives; we don’t build a picture of the world from the ground up, light beam by light beam, sniff by sniff.
Instead, we create and maintain a hallucination in our minds, which is updated when confronted with data that disagrees with the Prior Probability of the internal model, our hallucination. It’s like we’re dreaming, and the dream is being tested against our sense perceptions and updated in real-time. This is why optical illusions are invisible, as they don’t appear in our model, then once seen, they can’t be unseen: the model updates with new information, and the probability of seeing a blue whale in random dots is increased.
The Batinah goat is a long-haired breed I’ve been chasing around Oman’s Jabal Al Ahdhar mountains in the last few weeks. Having grown up around Yorkshire terriers, however, and not Batinah goats, the model in my mind (Prior Probability) immediately shows a terrier trotting across the road. It is, after all, more likely — given my upbringing — that a terrier is on the road.
Only when I get closer does my internal model — or hallucination — get updated with the new evidence: the horns and hoofs and aptitude to climb trees. Suddenly, the majestic Batinah goat forms in my mind’s eye. First, I see a Yorkshire terrier; moments later, a goat.
We hallucinate first, test second with our sense perceptions, update probabilities and alter our hallucination. Critically, it’s all a hallucination, before and after, and we continually update it based on the sense data. It’s ‘top-down’ — literally a pre-conception.
This is why we’re so often victims of our misconceptions. We do not see — do not hallucinate — reliably.
You’d be right if you think this sounds like starting a business.
While cycling, I am often puzzled about how we at Yokeru walked in the wilderness for two years in honest denial about the reality of the ‘unsuccess’ of our product.
I truly, honestly, completely saw what my optimistic Prior Probabilities showed me and didn’t update my probabilities with the new evidence. I refused to; I didn’t poke holes in my hallucination. More concretely, making very few sales should have triggered alarms, but we didn’t want to see them. Instead, our hallucination was determined to see the positive sales conversations and customer feedback (in sweet words, not in signed contracts, and so on).
If we remember we are in a hallucination, we can go to efforts to test it. To reveal inconsistencies in a hallucination, we have to be willing for it to fall apart. We can’t hold on too tightly; failure is a good thing — it reveals the truth — but this is antithetical to many of our ideals.
We often so badly want something to be true that we get overwhelmed and won over by our preconceptions. Our mind’s eye is a liar, but we are an accomplice by letting it continue to lie. It’s a pleasure, I have found, to live deluded. We must somehow internalise that the hallucination is all we have; it’s powerful and can be wrong.
We must build ways of testing our hallucinations — friends can tell us, therapists can point out inconsistencies, data can damn us. In startups, money dries up. Ultimately, we just need to ask more honest questions and be willing to hear the answers.
What evidence do you have that you are ignoring, and if you did update your probabilities, following Bayes’ Theorem, what would you change? — let me know!
My week in books…
Reading: Stalin: Paradoxes of Power by Stephen Kotkin. A Distant Mirror by Barbara Tuchman.
Live well,
Hector