Melanie Mitchell’s life changed on the New York City subway. During her post-college stint as a high-school math teacher in Manhattan, every subway ride was an opportunity to conquer a few more pages of Douglas Hofstadter’s Gödel, Escher, Bach. Reading it, she became fascinated with the way math, art, and music could help explain the emergent properties of intelligence. She realized she wanted to work with Hofstadter and become an AI researcher. 

Several decades later, Mitchell has achieved that goal and more. Her research has explored the limits of analogy-making machines, the particle physics of cellular automata, and the rules of genetic algorithms. 

Combined with a lifelong love of science and logic puzzles, Mitchell’s drive to “think about how we think” made Hofstadter’s research group a natural first step in her AI career. In the mid ‘80s, she worked to build Copycat, a computer program that looked for analogies between strings of letters. For example, if the string abc changes to the string abd, what is the analogous change to the string mrrjjj?  

“It's amazing what you can do in this very constrained domain,” Mitchell says. In Hofstadter’s group, the maxim was that constraints could breed creativity. 

While Copycat wasn’t particularly computationally expensive, one thing that set it apart was Mitchell’s brilliant visual interface that showed the computer’s actions in real time. For many computer scientists, Copycat provided immensely valuable insight into the thinking process of AI — and humans as well. 

At the University of Michigan, Mitchell also began to branch out into biologically-inspired AI with John Holland, one of SFI’s earliest profesors. Holland was the inventor of genetic algorithms, computer methods inspired by Darwinian evolution. He had developed a theory that predicted how these algorithms would solve problems. But when Mitchell and fellow Michigan graduate student Stephanie Forrest (now a member of SFI's external faculty) looked, they found something amiss. 

“We tried to probe his theory by coming up with the simplest example that, according to his theory, genetic algorithms would absolutely excel at — and we showed that it wasn't the case,” Mitchell says. Holland’s idea was that genetic algorithms would have a high fitness, or success, if they combined fit components together, like adding up multiple genes which all increase height. 

It turns out, just like biology, algorithms also suffer from a problem called hitchhiking: maladaptive genes located near fit genes on the chromosome can get carried along simply by association. 

The research would lead Mitchell to SFI in the nineties, where she spent much of the decade as the director of SFI’s Adaptive Computation program, where her goal was to make computational systems “more lifelike” by bringing in insights from natural adaptive systems such as insect colonies and immune systems, as well as biological evolution. At SFI, Mitchell worked with researchers across a breadth of fields, from population geneticists to physicists.

Since moving up to Oregon in the early aughts, Mitchell has remained an external faculty member at SFI, writing books about complexity and artificial intelligence. Her research over the past decade has moved back to AI, trying to understand image recognition from the perspective of analogy-making.  

What Mitchell wants to know is how our version of object recognition differs from a computer’s — how both top-down and bottom-up perspectives can be continually combined to enable visual analogies.

Though her research spans multiple disciplines, Mitchell finds unifying connections throughout. 

“In adaptive systems like evolution, there's this balance between what's called exploration and exploitation,” Mitchell says. Exploitation meaning repeating what works and exploration meaning trying new strategies. Individuals have to do some of both to succeed. Too much exploitation and you get stagnation; too much exploration and you lose stability. This isn’t just the case for carbon-based organisms — Copycat and other AI programs also need to balance exploitation and exploration to succeed.  

Lately, Mitchell has been working on an even more general principle — one at the heart of Gödel, Escher, Bach. “If you look at that book, it's really about intelligence. About how something like intelligence can emerge from a non-intelligent substrate of, say, neurons interacting with each other,” she says. 

This year, along with SFI External Professor Melanie Moses, Mitchell is investigating foundations of intelligence by hosting a series of interdisciplinary workshops to answer questions like where it can exist and whether human intelligence is the same as computer intelligence or the intelligence of a swarm. By bridging disciplinary divides in how intelligence is understood, they hope to discover the next frontier in AI research.

In other words, she’s still thinking about thinking.