This semester, I am taking a Neuroscience minor course in Groningen. Although at first glance, the topic may seem quite distant from the usual activities we engage in at W.I.S.V. ‘Christiaan Huygens’, the reality is that the fields of Computer Science and Neurobiology are becoming increasingly intertwined. For example, artificial intelligence (AI) draws significant inspiration from the structures found in mammalian brains, particularly neural networks. These biological systems have provided an example for creating algorithms that copy how neurons process information, leading to developments in AI.

An excellent example of this connection is Neuralink, a cutting-edge startup focused on developing brain-machine interfaces. Their goal is to restore sensory and motor function in patients with neurological disorders, pushing the boundaries of what is possible with technology and neurobiology.

To introduce you to the fascinating world of neuroscience, I will first provide an overview of basic brain anatomy. Following that, we will dive deeper into the similarities between neural networks in the mammalian brain and in machine learning systems. As a final treat, we will briefly explore the exciting advancements Neuralink has made in bridging the gap between machines and the human brain.

Neuroanatomy

To understand the functioning of our brain, we have to understand its smallest details. The brain and nervous system are composed of specialised cells known as neurons. While neurons share many similarities with other cells in the body, they also possess unique features that enable them to perform highly specialised functions. Let’s break down the key parts of a neuron:

Anatomy of a Neuron

• Soma: Also known as the cell body, the soma houses the nucleus and performs the basic functions of the cell, much like the cells we learned about in high school biology. It’s responsible for maintaining the cell’s health and overall function.
• Dendrites: These are tree-like extensions that branch out from the soma. Dendrites are specialised in receiving signals from other neurons by binding to chemicals called neurotransmitters. This is how neurons communicate and transmit information across the nervous system.
• Axon: The axon is a long, tail-like structure that extends from the neuron, designed for sending signals over long distances to other neurons or parts of the body. Some axons can be long, reaching from the spinal cord to your arms or legs.
• Terminal Buttons: At the end of the axon, terminal buttons release neurotransmitters into the synaptic cleft (the small gap between neurons). These neurotransmitters then bind to the dendrites of other neurons or different tissues, allowing the signal to be passed on.

Some neurons have incredibly long axons that allow signals to travel quickly over large distances. For example, certain neurons send signals from the brain to muscles in the arms and legs, allowing us to respond to stimuli with remarkable speed — up to 6 metres per second!

The many neurons (think billions) in our body form networks of interconnected ganglia (groups of neurons, also called nuclei) which are largely responsible for the functioning of our body.

This small introduction to neuroanatomy is already enough to see the similarities between the neural networks used in machine learning and the ones in our bodies.

Neural Networks

In a previous edition of the MaCHazine, the history of artificial intelligence was told, and we discussed one of the foundational concepts: the Perceptron, a simplified model of a neuron.

The image below presents a graphical comparison between a Perceptron and a biological neuron. While the Perceptron is a basic building block of machine learning models, its structure and function are strikingly similar to the anatomy of a biological neuron. Let’s dive deeper into the anatomy of the Perceptron and explore how it mirrors the workings of a biological neuron.

Anatomy of the perceptron

• Inputs (Dendrites): In biological neurons, dendrites receive signals from other neurons in the form of neurotransmitters. These inputs determine whether the neuron will activate or “fire”. In a Perceptron, the inputs are the equivalent of dendrites, where signals (represented by numerical values or features) are fed into the model. Each input is assigned a weight, similar to the varying importance of signals received by biological dendrites.
• Summation and Activation (Soma): In the biological neuron, the soma (cell body) processes the incoming signals and decides if the total input is strong enough to generate an action potential, which is the signal that will be sent down the axon. Similarly, in the Perceptron, all input signals are multiplied by their respective weights and then summed. This weighted sum is compared to a threshold (for example when using ReLU as the activation function), determining whether the Perceptron will “fire” (output a signal) or remain inactive. This mimics the decision-making role of the neuron’s soma.
• Output (Axon and Terminal Buttons): If a biological neuron “fires”, an electrical impulse travels down its axon to the terminal buttons, where neurotransmitters are released, passing the signal to the next neuron. In the Perceptron, if the weighted sum exceeds the threshold, the Perceptron outputs a value (typically 1), signifying that it has activated, just as the neuron sends a signal down the axon. If the threshold is not met, the Perceptron outputs a 0, akin to the neuron not firing.

Although the Perceptron is a highly simplified model, it is the foundation of more complex neural networks. Modern AI systems use layers of interconnected Perceptrons to simulate learning processes, inspired by how neurons in the brain work together to process information, recognize patterns, and make decisions. These systems continue to evolve, with newer architectures building even more sophisticated features of biological neurons.

Neuralink: Bridging Neuroscience and Technology

As we’ve explored the parallels between biological and artificial neural networks, let’s examine how these fields connect in groundbreaking ways. Neuralink sets an example by developing brain-machine interfaces that directly connect the human brain to computers.

Ultra-Fine Neural “Threads”

Inspired by biological neurons, Neuralink has developed extremely thin, flexible polymer probes called “threads”. Each thread contains 32 electrodes and is less than 6 μm in diameter — much thinner than a human hair. These threads are analogous to axons, serving as conduits for neural signals. Their small size and flexibility may improve biocompatibility compared to more rigid electrodes, mimicking the natural flexibility of neural tissue.

Mirroring the billions of interconnected neurons in our brains, Neuralink’s arrays can contain up to 3,072 electrodes distributed across 96 threads. This high-density approach allows for recording from many more individual neurons — a crucial step towards the complexity of biological neural networks.

Neurosurgical Robot

To achieve the precision required for inserting these delicate threads, Neuralink built a neurosurgical robot. It can insert up to 6 threads (192 electrodes) per minute with micron precision, while avoiding blood vessels on the brain’s surface. This robot’s ability to navigate the complex landscape of the brain showcases the intersection of neuroscience knowledge and advanced robotics.

Custom Electronics

Just as the soma of a neuron processes incoming signals, Neuralink developed a custom chip to amplify and digitise neural signals from thousands of electrodes simultaneously. This chip acts as an artificial “soma”, processing the vast amount of information collected by the electrode array.

Conclusion?

In early tests with rats, Neuralink’s system has demonstrated the potential to record from large numbers of neurons across multiple brain regions. This technology could enable breakthrough applications in treating neurological disorders and enhancing human-AI interaction. As we continue to unravel the mysteries of the brain and advance our understanding of artificial neural networks, the possibilities for innovation at this intersection are truly exciting. Neuralink’s work represents a significant step towards high-bandwidth brain-computer interfaces, bridging the gap between neuroscience and technology in ways that were once only imaginable in science fiction.

This article was published in “Machazine” 2024–2025 Q1 edition, the magazine of the Study Association W.I.S.V. Christiaan Huygens (TU Delft)