The Human BrAIn: How Neuroscience Can Revolutionize Machine Learning

5 min readMar 12, 2025

AI has made incredible strides, but is it true intelligence? Jeff Hawkins’s A Thousand Brains: A New Theory of Intelligence offers a groundbreaking perspective on how intelligence emerges — not from a single, centralized model but from thousands of mini-models working in parallel. He challenges our understanding of cognition, presenting a radical approach to learning and perception. This biological approach to learning provides a fascinating comparison to how today’s artificial intelligence (AI) systems operate. In this blog, we’ll explore the core ideas of the book and how they challenge conventional AI learning paradigms.

*Cortical Columns depicting* depicting the complex connections in the motor cortex of a mammal. Visualization by *Nicolas Antille*

The Core Idea: A Thousand Models in Parallel

Traditional neuroscience often describes the brain as a single processing unit, but Hawkins introduces the “Thousand Brains Theory”. He argues that intelligence isn’t centralized but distributed. Each column in the neocortex builds its model of the world, running in parallel. These mini-models interact, compare notes, and refine each other, resulting in a cohesive understanding of reality. This distributed approach to intelligence challenges the way AI systems are designed today. Unlike the brain’s decentralized learning system, AI models rely on a vastly different approach to understanding data — one that comes with significant limitations.

How Does This Compare to AI?

Most AI models today are built on deep learning architectures, where a single neural network is trained on massive datasets. These systems learn through pattern recognition but lack the distributed, multi-model learning that Hawkins describes in biological intelligence. Key differences include:

AI relies on statistical correlation, while the brain builds predictive models.

Deep learning models recognize patterns in data but do not form an understanding of the underlying structure. On the other hand, our brain creates models that identify patterns and relationships and connect concepts dynamically through predictive modeling. Instead of relying on a single representation, the brain constructs thousands of independent models within cortical columns, each maintaining its reference frame. These reference frames allow the brain to relate memories, recognize connections, and predict future inputs based on past experiences. Sound familiar? (In large language models, context windows and token embeddings serve a similar role by structuring how models process and relate to sequential information, much like how the brain constructs models through reference frames). Unlike the brain, which continually refines its understanding, AI systems remain trapped within their training data. This lack of ongoing adaptation limits their ability to handle new or unexpected inputs.

AI models are static after training, while the brain is dynamic and constantly updating.

AI models learn once and then stop until retrained. Most models learn in a fixed training phase, where massive amounts of data are absorbed before deployment. Once trained, they don’t continuously update themselves in real time but instead rely on periodic retraining or fine-tuning to improve. The brain, on the other hand, never stops learning. It constantly refines its internal models, adjusting predictions as new sensory inputs come in. Every movement, every glance, and every experience fine-tunes its understanding, making it fluid and adaptable. While AI models recognize patterns, the brain lives in them, constantly updating its predictions to align with the ever-changing world around it.

3. The brain learns through movement, whereas AI models are passive.

Get Shiphalika Mahesh’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

The brain doesn’t solely learn through static observation — it learns by moving. Every step you take, every time you reach for an object, your brain updates its understanding of the world. Jeff Hawkins argues that intelligence isn’t just about recognizing patterns; it’s about physically interacting with the environment to refine those patterns. AI, on the other hand, lacks this kind of embodied learning. It doesn’t explore, touch, or move through space. Instead, it processes vast amounts of static data, recognizing correlations but never experiencing the world firsthand. Our brains map space using grid cells for physical navigation and, even more interestingly, conceptual thinking. While the brain builds knowledge by continuously engaging with its surroundings, AI models remain passive observers, disconnected from the reality they attempt to interpret.

How This Can Change AI

AI today is powerful, but it’s also rigid. Most systems rely on deep learning, which excels at recognizing patterns but struggles with adaptability. When faced with something new — data it hasn’t explicitly seen before — an AI model often falters, revealing its lack of true generalization. Unlike the brain, which continuously refines its understanding through movement and interaction, AI models remain static, fixed in the knowledge they were trained on. They don’t evolve in real time; they only improve when explicitly retrained with new data.

But the biggest limitation? AI doesn’t truly ‘understand’. A deep learning model doesn’t form concepts the way we do — it maps inputs to outputs based on statistical correlations without grasping the deeper structure of the world. It’s a sophisticated pattern-matcher but not a thinker. This becomes especially clear in large language models (LLMS), which are incredibly powerful tools but can be brittle. They generate impressively fluent text that reads naturally, yet they can fabricate sources or state contradictions. We call these hallucinations. The brain is beautifully designed to cross-check and refine its understanding by integrating across multiple “models” across cortical columns.

So what would it take for AI to break these constraints? If AI models could start understanding and integrating sensory inputs, they wouldn’t just process data; they would experience it. Imagine an AI-powered robot that learns the texture of objects by touching them or a self-learning system that refines its predictions by actively moving through its environment. Instead of passively analyzing pixels or text, these AI systems would develop dynamic, evolving models — just like the brain does. When AI begins to perceive, interact, and adapt in real time, it may finally cross the threshold from pattern recognition to genuine learning. If AI were to mimic the brain more closely, it would need to build multiple parallel models, each contributing to a broader, more adaptable intelligence.

Hawkins’s Thousand Brains Theory suggests a path forward. If AI could move beyond a single, rigid model of the world and instead build multiple, overlapping representations — constantly refining them through interaction — it could become far more flexible and resilient. Rather than memorizing data, AI would learn dynamically, adapting to new information the way humans do. With a shift toward parallel processing and real-world engagement, AI could finally take a step closer to true intelligence.

This shift is already underway in fields where researchers are working on Embodied AI, which integrates machine learning, computer vision, robot learning, and language technologies, all together creating a holistic “embodiment” of artificial intelligence. These agents refine their knowledge through trial and error, physical movement, and real-time feedback from the world. The embrace of sensorimotor learning would allow the space to reach beyond pattern matching and toward truly adaptive intelligence.

References