Article illustration 1

A decade ago, intelligence seemed an impenetrable enigma. Today, large language models and transformer-based AI systems generate art, compose symphonies, and debate philosophy with eerie proficiency. As these artificial neural networks—originally inspired by the brain—approach human-like cognition, a provocative question emerges: Have we demystified human intelligence? Not quite, argues computational neuroscientists like the author of this DendWrite perspective. While AI offers a powerful proof of concept for connectionism, the brain’s biological intricacies reveal profound unresolved puzzles that could redefine both neuroscience and AI development.

The AI Revolution and Its Illusions of Mastery

Artificial neural networks now rival humans in tasks once deemed uniquely ours, powered by simple ingredients: nonlinear activation functions, massive datasets, and scalable architectures. This success fuels the connectionist view—that brains operate similarly. Yet, this parallel is superficial. As the author notes, "Artificial neural networks give us a vocabulary to understand brain computation," but they don’t capture the full story. For instance, AI’s reliance on backpropagation—a method that propagates errors backward through layers—has no clear biological counterpart. Brains use local rules like Hebbian plasticity or neuromodulation, raising questions about how learning truly occurs in vivo.

"I have sometimes been asked to explain ‘on one foot’ how the brain works. I often answer that the brain is basically a deep artificial neural network, and the rest is commentary. Today, though, this is not sufficient."

Where Brains Defy Silicon: The Critical Gaps

AI’s advancements obscure lingering biological advantages. Humans learn efficiently from minimal data, a feat often attributed to evolutionary "pre-training" rather than architectural superiority. However, neuroscientists point to mechanisms like hippocampal-cortical interactions for memory consolidation, which mitigate catastrophic forgetting—a chronic AI flaw where new learning overwrites old knowledge. Physical problem-solving in dynamic environments also highlights AI’s limitations, though this may yield to more data and compute.

The real divergence lies in the brain’s architecture:
- Recurrency Over Feedforward Design: Unlike AI’s linear transformers, the brain thrives on pervasive recurrent connections, creating a dynamic system that processes information across multiple timescales. This recurrency may enable real-time adaptation absent in static models.
- Cortical Conundrums: Why six layers in the neocortex? What computational roles do excitatory vs. inhibitory neurons play? Dendritic structures suggest sophisticated information integration, but their purpose remains elusive.
- Modular Mysteries: Brains are a patchwork of specialized regions (e.g., fly vision circuits) versus general-learning modules. Deciphering how evolution balances custom engineering with flexibility is key to understanding human cognition’s uniqueness.

Neuroscience’s New Mandate: Beyond Mimicry

As AI evolves, neuroscience must pivot to validating biological hypotheses. Techniques like electron microscopy allow mapping synaptic connectivity to explore "nature vs. nurture" in information storage. Theoretical work on backprop-free learning rules needs experimental backing in complex tasks. Understanding cortical layers or cell types isn’t academic trivia—it could unveil fundamental computational principles, much like transformers did for AI.
The quest isn’t just about closing gaps; it’s about redefining intelligence. AI forces us to confront what makes us human, and neuroscience holds the lens. As the author concludes, the minds of man and machine haven’t converged—yet. Unlocking the brain’s secrets could inspire next-gen AI, but more profoundly, it safeguards the irreplaceable value of biological cognition in an automated world.

Source: Adapted from What remains of the mysteries of the brain? by DendWrite.