Glitches as Features: What John Hopfield Says Artificial Neural Networks Are Missing

Asked by Lex Fridman about the most profound difference between biological and artificial neural networks, John Hopfield points to the many components and properties of real neurons. In evolutionary biology a small quirk in how a molecule works can be sharpened by evolution into a useful feature rather than remaining a glitch. He expects neurobiology to have captured all kinds of ways to make neurons do useful work, and says that aspect has been completely suppressed in artificial networks.

His example is synchronization. Oscillators with similar rhythms can undergo a phase transition in which everyone suddenly falls into step. He recalls London's Millennium Bridge around 2001, where pedestrians who normally do not walk in lockstep began to synchronize with the bridge's sway because the engineers had not accounted for the side-to-side motion of walking. The bridge was closed for about two years for stiffening.

Nerve cells loosely coupled and firing action potentials at similar rates can likewise lock together under some circumstances, and other cells will notice when they fire together. Hopfield says an evolving brain can make a computational feature out of this. Most artificial neural networks, by contrast, do not even have action potentials, let alone a mechanism for synchronizing them.

On how evolution leverages biological messiness, Hopfield explains that a duplicated stretch of DNA lets one copy keep encoding a protein while the other drifts and may gain a new function, under evolutionary pressure to improve it. Fridman likens this to companies, which are born and go bankrupt on a somewhat shorter timescale.

Hopfield adds that companies struggle to launch new products that compete with their old ones, citing how IBM built its first PC in an isolated internal unit and, for the first time, did not insist on IBM components. In his telling, IBM entered a very different market only by changing its culture, whereas biology finds new niches in a more adaptive way and is better at that kind of integration.

Hopfield says adaptation is everything, but distinguishes learning over evolutionary time from learning during an individual's life, and biology has both. What makes neurobiology hard, he says, is that the mathematical systems computer scientists build are based on a different, non-evolutionary kind of system. Computer chips are essentially two-dimensional and struggle with three-dimensional wiring, whereas the neocortex is sheet-like and sits on white matter with about ten times the volume of the gray matter, holding the wiring.

He argues that the effect of a computer's structure on what is easy and hard is immense, and biology makes some things easy that are hard to do computationally even though it cannot do simple floating-point arithmetic. Fridman asks which kind of learning he finds more fascinating. Hopfield is in awe of evolution but calls it a black hole for research, whereas development and early life can be teased apart, including the stage of greatest cell death in the brain during infancy.

Asked which field will bring the biggest breakthroughs in understanding the mind, Hopfield says he sees the world through physics, the field he grew up in with two physicist parents. The key thing he took from it was a feeling that the world is an understandable place, where enough experiments and relevant mathematics let you work out how things function.

He says he has a nebulous notion of understanding that he has never known how to express well. When he taught freshman physics he wanted students to grasp the laws rather than memorize examples, so that they could anticipate a ball's trajectory. He says simple feed-forward networks do amazing things yet contain nothing of the essence of what he would call understanding, which is more than an enormous lookup table.

On whether feed-forward networks will ever understand, Hopfield says real systems depend on feedback. He knows that a mathematical system with feedback could in principle be unrolled into a feed-forward one, though he cannot prove it, but doing so would take an exponential expansion in the amount of machinery. So he regards feedback as essential.

Fridman asks whether there is a fundamental difference between the two kinds of networks or just surface detail. Hopfield notes that a neurosurgeon can judge from brain rhythms whether a brain is recoverable, and these rhythms are absent from anything running at Google. Fridman then asks whether they might ever show up in AI, and Hopfield compares today's systems to a child first learning the piano, while a great musician plays the same instrument.

Hopfield expects a repeating pattern. The AI and computer science community takes the current model of neurobiology, pretends it is good enough, and does everything it can with it. After a while it grinds to a halt, someone realizes something else is needed from neurobiology, and a new idea lets the field go much further. He thinks this could take several generations, with each pushing machines further toward passing the Turing test in broader ways, and his sense is that a couple more might be enough.

AI researchers might dismiss brain rhythms as unimportant, he says, using a story about his first car, a 1936 Dodge coupe, and a quirk that told him when he was past 45 miles an hour. A designer would never build a car that way, but in biology, if such a signal were useful, evolution would simply capture it. He adds that laptops have around 10 to the 10 transistors and physical systems with so many similar parts show collective properties, such as earthquakes, but he says artificial networks use none, so it will take generations to learn whether biology uses them.

Hopfield says one of his biggest surprises was how useful manifestly non-biological learning systems have turned out to be.