How does the brain work?

How do we think?

What is intelligence?

We, humans, have been thinking about this forever. Since Aristotle and Plato, the origin of intelligence has been a riddle. Despite our efforts to create or improve intelligence, we are still far from reaching the level of the brain.

This is where Jeff Hawkins and the Thousand Brains of Intelligence theory enter. Based on the latest research, Hawkins proposes a new theory of intelligence to explain how the brain works. Ideas that can change the direction of artificial intelligence (AI), and even the future of humanity.

I found it beautiful. I hope you will too.

📱 A Few Words about the Author

Jeff Hawkins is one of those people who discovered their purpose early. To put it simply: his mission is to uncover intelligence. And he turned every stone to do so.

Already in his PhD, he proposed to research the mechanism of the neocortex, the brain region responsible for intelligence (we will get into this in a bit). This got rejected because the field was thought to be too complex with too much unknown. Too risky, no funding.

So what did he do instead? Hawkins followed his entrepreneurial drive and founded two companies: Palm Inc. and Theo. With these two companies, he built the foundation of handheld computing, creating the first pen-based “touchscreen”. 25 years before the launch of the iPhone.

But he did not abandon his dreams of neuroscience research.

Without public funding, he founded two institutes. Redwood Neuroscience Institute focused on the biological side of intelligence, whereas Numenta tackled AI. Hawkins and his team published breakthrough papers in neuroscience, rethinking fundamental elements of brain theory.

Three years later, Hawkins published these findings in a book. Here, he collects scientific insights mixed with metaphors and personal anecdotes.

A fantastic book that I would like to recommend to you today.

📖 3 reasons to read A Thousand Brains

To learn about the brain. Have you ever pondered the question of how we think or learn? If you are like me then you have.

The brain is the most critical organ in the body. Without the brain, there is no life, yet we know so little about it! If you are curious about how this fascinating organ works, you will love this book.

To learn about AI. This was an unexpected turn, Hawkins transitioned from biology into computer science. With incredible fluency, he shifts the conversation from biological intelligence to AI. He compares human and AI, what current AI lack compared to the neocortex, and proposes a novel strategy for a truly generic AI. I found it new and made me wonder about the future of AI.

Because it is science put simply. I admire people who can write about complex scientific concepts in an easy way. It demonstrates true knowledge. The way Dr Stephen Hawking did with “A Brief History of Time”, he was able to break down relativity and sparked people's interest in cosmology.

I believe Jeff Hawkins does the same. This time instead of spacetime, we focus on the neocortex. If you enjoy reading science, I found it to be one of the best in the genre.

🤯 3 things that blew my mind

There is a universal algorithm for learning and perception

It is easy to think that the mechanism of learning is wide-range. Learning language, maths, or music feels different.

However, the structure of the brain shows something different. The brain is homogenous. We do not find separate sections or compartments. This means that the same structure and mechanism drive all learnings.

I found it fascinating that there is a fundamental algorithm that drives all perception. It made me wonder how extremely versatile our brain actually is.

The brain makes tiny predictions. All the time.

The basic unit of the neocortex is a group of neurons called cortical columns. The role of the cortical columns is twofold. First, they map the world, creating reference models based on sensory input. Second, they are continuously making predictions.

We go into the details of how these happen later, but this was mind-blowing to me.

I used to think of the brain as an input-output system, reacting to its environment in real-time superfast. Instead, it's a dynamic, prediction-based system, creating multiple predictions simultaneously, and selecting and merging the most appropriate ones.

And how does this selection happen? That is the third mind-blowing idea.

The brain aggregates predictions based on voting

With more than 150 000 cortical columns in the brain, how do we have a single perception? How can we consolidate the different models and the different predictions with each other?

Historically, we thought the brain to be hierarchical, from general knowledge branching down to specific expertise. However, the structure of the brain shows a decentralised model.

In this decentralised model, Hawkins proposes that the columns vote. In a democratic fashion, each column of the neocortex (involved in perception) proposes its own predictive output. The outputs are consolidated. The ones with the greatest consensus (from the other columns) get strengthened, whereas the ones with fewer votes get suppressed.

The brain builds a predictive model based on the most relevant columns, leading to a smooth perception.

📈 3 ways how the book changed me

I can learn anything! I found this book empowering. It showed how the brain is a universal learning machine. Constantly forming updating the world model, creating new synapses (connections) and demolishing old ones. A dynamic system that we can shape with effort. And this is the empowering idea: we can virtually learn anything with effort.

I should not always trust my brain. Hawkins also discussed the relationship between the “old” and the “new brain”. The old brain is driven by biological desires and influences emotions. The new brain is responsible for critical thinking, but it requires effort, and it can easily be hijacked. Even if I don’t want to. Understanding that there are gaps and biases in our brain, helped me build my self-awareness and to develop the habit of questioning my perception and thoughts.

I became more optimistic about AI. Coming from a life-sciences background, it helped how Hawkins first laid the biological foundations and then shifted towards computer science. He provided a holistic image of AI, and how the universal learning machine mechanism from neuroscience can be implemented. He also covered the risk concerning AI. Understanding a general more in-depth, made me more aware and excited about its future potential.

📕 Learnings, Notes, and Insights

⚡ Let’s start with the neuron

The brain consists of neurons or neural cells.

A neuron is connected to many others upstream and downstream.

When a neuron receives an electric signal strong enough, action potential occurs. This is the firing of a neuron. The signal travels down the axon and activates other neurons downstream.

This downstream activation occurs via synapses, these are the junctions between neurons.

The more we use neural pathways, the more synapses the brain creates, strengthening these links. Old synapses fade away or are removed entirely.

This is the basis of learning and forgetting.

🧠 Next stop: the Neocortex

The neocortex is the organ of intelligence.

Whenever we are using the cognitive function, the neurons of the neocortex are firing.

• you reading this → neocortex
• me writing this → neocortex
• you thinking about what you’re reading → neocortex
• us perceiving the world → neocortex
• us remembering an event from the past → neocortex
• vision, language, science, music → neocortex

We used to think that the brain is compartmentalised. Different regions of the brain are responsible for specific kinds of learning. They feel different.

However, the different regions of the brain show no major biological differences.

This indicates that all types of intelligence use the same mechanism of learning.

What is this mechanism?

🏛️ The basic unit of the neocortex: cortical columns

A cortical column is the basic organisational unit of the neocortex encapsulating a group of neurons organised into columns and layers.

It is a repeating structure of the neocortex.

Understanding the mechanism of a cortical column would reveal how the neocortex works.

Vernon Mountcastle first proposed the idea of a general-purpose method of learning. He argued that there is a universal principle, a function based on which the brain can learn virtually anything.

Hawkins proposed that the cortical columns are working based on this universal principle.

If the cortical columns are a basic input-output system, how can we create a complete perception of the world?

🚄 To create a complete model of the world we need continuous predictions

The input of the cortical columns is clear: visual, audio, or other input. We constantly see, hear, and use our other senses.

What is the output? How can the brain create a smooth model of the world?

The key: the brain makes predictions.

Let’s use the train example: if you’re sitting on the train, your eyes jump every 3 seconds, yet your perception of the world is blended. How?

• Input: the visual of the world
• Output: mental model constantly updates a real-time predictive model of the world

This continuously updated predictive model creates a single perception.

💥 The mechanism of predictions: dendrite spikes

Classic spike or action potential, as discussed, travels down the neuron to active other downstream neural cells.

It is a binary switch: no action potential, no downstream activation.

Hawkins and his team revealed, a new type of spikes: dendrite spikes.

An electric signal comes and it travels down the dendrite, but it does not activate the classical spike. It does not reach action potential.

It just raises the voltage of the neuron to almost make the neuron spike.

This way when the next signal comes, the classical spike would trigger faster.

Think of it as a balloon.

• a classic spike would blow the balloon all the way until it pops. the loud pop would trigger the next neurons.
• a dendrite spike would blow the balloon to a point where it almost pops. the smallest touch would lead to a pop leading to the activation of downstream neurons.

This phenomenon is called priming, putting neurons in a predictive state. “Primed” neurons will spike faster than other neurons (lacking dendrite spike).

Because neurons with dendrite spikes fire faster, they are ideal for predictions.

So when a certain pattern of activity is detected, dendrite spikes will prime neurons in the pathway, ready to fire earlier than other neurons.

What are these patterns of activity?

🗺️ Reference Frames and Movement

Hawkins proposed two key elements to effectively predict perception: reference frames and movement tracking within these reference frames.

Coffee cup example: moving your finger on the coffee cup.

• there are two key questions:
• what is the object you’re touching? this is the reference frame.
• where is your finger moving on the coffee cup? this shows we’re changing our location on the reference frame.
• with these two pieces of information: the current location of the finger on the cup and the direction of the movement, the neocortex will be able to predict the new location relative to the cup

I admit this concept was harder for me to grasp. Let me try to help you by providing my example:

• Let’s say you have a map of Budapest. This is the reference frame.
• Let’s say you know the current location. Ors vezer square.
• If I give you the movement (travel down 3 stops with Metro 2), will you be able to “predict” where will end up?
• Most likely yes, because we’re moving within a well-established reference frame (at Keleti train station)

The requirements for effective predictions and single perception are, therefore: reference frames and the movement within those frames.

If this is still unclear, can you please reach out and provide me with your feedback? I'm still looking for ways to make this easier.

🔭 Multiple reference frames

This is the role of cortical columns: to create reference frames and track movement within.

Each cortical column will be exposed to only part of an object. Each will correspond to a specific sensory organ.

With each cortical column creating its own model, there will be thousands of reference frames and locations at the same time, covering all aspects of perception.

This is the key to a single perception.

And this is the pattern of activity (movement within the reference frame) recognised and initiated by the dendrite spikes.

💡 The mechanism of learning

To account for universal learning, Hawkins expands this theory beyond the coffee mug and physical perception.

Reference frames can be flexible. Once we learn what a car or house is like, we can manipulate this reference frame to fit any new cars or houses.

Reference frames can also be non-physical.

• Solving math questions: basic arithmetics (addition, multiplication) can be viewed as movements in the reference frame of numbers
• Language: using words to move in the reference frames of grammatical rules and vocabulary

fMRI studies have shown that recognition of an image of a bird triggered a similar pattern as moving across the house. But instead of moving across rooms, the brain moved from the legs, to the torso, and to the head of the bird, creating a single perception.

🗳️ A Thousand Brains Theory of Intelligence

Historically, the brain was thought to be hierarchical. However, the neural links in the brain seem to be decentralized and distributed.

How does the brain consolidate all thousands of predictive models created at the same time?

Columns vote.

Based on their own reference frames, the columns will lead to different predictions. They all send signals (dendrite spikes + classical spikes).

These predictions are aggregated. The most common ones will suppress the least common ones.

This consensus will govern perception and cognition.

The brain needs consensus, which explains our challenge with optical illusions

• we can only see the vase or the face
• we can only see this dress in blue/black or white/gold

We cannot perceive both ways because our cortical columns are pushing for a single consensus.

So to summarise:

A thousand models are created in a thousand columns.

These columns aggregate their predictive output, vote, and reach consensus.

This is our perception.

How awesome is this?

❓ Remaining Questions

This is only a theory.

It explains many key questions around the brain but cannot answer everything (yet). It still has its limitations.

For example:

• how are reference frames created and stored?
• what are the parameters stored?
• how is the reference frames different based on sensory output (e.g. visual or audio reference frames)?
• what is the capacity of a column?
• what is the exact mechanism of voting across the columns?

🤖 Transferring brain theory into computer science

This new brain theory can be applied to AI as well.

Today’s AI is very specific, it can do one role outstandingly (e.g. play chess). However, it cannot learn universally.

Brain theory can support this new direction of building generic AI, the universal learning algorithm of the neocortex can support the development of Artificial General Intelligence (AGI).

🎓 The new measure of AI

We can shift from measuring how well AI can perform a single task.

The measure can be adaptability: how well a machine learns, and stores new information about the world.

Following the mechanism of the neocortex, the expansion of AI can be tremendous:

• new sensors can lead to new types of input beyond what we already have (e.g. visual or audio input)
• the columns to store and learn new information can be linked virtually, allowing the connection of millions of columns creating a more robust predictive-learning model

🦺 Incorporating safety into AI

AI is already being used to track people, spread propaganda, and influence elections.

The new AIs can even have a larger impact. We need to think about safety and address these key risks.

This was my first time encountering Isaac Asimov’s three laws for safety:

1. A robot may not injure a human being or, through inactivation, allow a human being to come to harm
2. A robot must obey orders given by human beings except where such orders would conflict with the First Law
3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law

🌎 The Existential Risk

1. Self-replication: any system capable of replicating itself can be dangerous e.g. a biological or a computer virus
2. Motivation: biological motivations are a consequence of evolution, it is a drive for living things to replicate better than the competition
3. Intelligence: Hawkins argues that intelligence in itself does not necessarily pose an existential threat. Unless coupled with the previous two, intelligence in itself does not seek to harm and destroy.

Thank you for reading the article!

I'm grateful for your time.

If you have any feedback, question, or other ideas, please feel free to reach out. Always happy to hear others' ideas.