The First Idea - Stanley Greenspan and Stuart Shanker

Update (8 April 2022): I just had the chance to chat with Nima Fazeli about this book. The discussion really heightened my awareness of how I have been socialized to talk about autism; I find it challenging to be articulate and not oversimplify things. Looking back on this blog post, I want to be more sensitive and compassionate before leaping to conclusions about what people need or want going forward. Of course, everything is a learning process, so I will leave the post as it was when I first wrote it, as is my usual approach.

Original text from here on:

This is a marvelous work! Every artificial intelligence researcher and roboticist should read this book. I sincerely believe it has insights that are critical to building a better understanding of machine/deep learning.

One small caveat: I have only finished about 3/4 of the book, which is the portion that conveys the central idea. The following chapters apply this idea to explain how to help people with autism, explore the development of society, and examine possible futures of our civilization of conscious beings. So, it's not too far off from my usual sci-fi reading!

By the way, the full title is actually The First Idea: How Symbols, Language, and Intelligence Evolved from our Primate Ancestors to Modern Humans.

The central idea of this book is that emotions are the basis for symbolic processing, logical reasoning, and consciousness. I don't think I can put it better than the mic drop that is page 292 of the book:

"We have shown how emotional signaling may explain the vital components of the relationship between emotion and cognition in terms of the brain and cognitive theory... As emotions are used to signal, they can be modulated through emotional interactions with others. The perception of the emotion, such a fear or anger or joy, can be experienced as a freestanding perception or image. This image can then acquire meaning through further emotional interactions... As it acquires emotional meaning, the image and its associated emotional patterns become a symbol. As a growing symbol, it can be united with other symbols into concepts and ultimately become part of an integrated system of reflective thinking. What becomes conscious, therefore, is the patter of emotional signaling into which the primary emotions have been transformed."

The authors thoroughly justify this bold claim in a variety of ways, of which I have chosen three to highlight. Each of these ways is an avenue of study that the authors pursued over several decades of their careers in child psychology and development, and each is supported by a plethora of citations.

First, the authors clearly explain how and why emotions were left out of the work of famous and critical thinkers from Descartes to Chomsky. The traditional view is that emotions and reason are separate; but in The First Idea, the authors instead show that emotions shore up many of the assumptions and awkward leaps of logic that these thinkers made.

Second, the authors demonstrate the growth of emotion using the growth of humans and primates. They demonstrate that the increase in symbolic reasoning that a person experiences in childhood is also present in the archaeological record; through human history, we gradually became more and more capable of symbol formation in parallel to developing larger and more complex social groups. To really drive the point home, the authors show that primates go through similar development. The more closely related a primate is to humans, the farther along in this symbolic development it can achieve. In particular, one of the authors conducted research with bonobos and chimpanzees who were able to learn sign language and use picture boards to engage with humans in complex linguistic interactions.

Third, the authors use their research in child development to show how an individual matures emotionally and symbolically in parallel. A person proceeds from catastrophic emotion (reactivity such as fight or flight) to calm reason (the ability to introspect and control emotions) by building ever more symbolic and emotional complexity. In particular, we "double-code" experiences as both physical and emotional, which allows us to reason over experiences by imbuing them with emotion. Importantly, the authors show that this progression through emotions takes a typical path in most people who are raised by supportive caregivers. The caregiver provides emotional interaction and teaches a growing child how to regulate their own emotions. The authors emphasize this importance of the emotional interaction by highlighting the challenges in symbol formation and usage faced by children who grow up without heavily-involved caregivers (e.g., in an orphanage) or with biological challenges (e.g., neurological features that lead to autism).

These three points above explain why emotions are so critical to building an understanding of consciousness and human experience. But, the authors take it further; they then explain how this research let them develop thorough explanations for autism, and methods to help people with autism. That is to say, this book is not just philosophical musing, but rather an avenue for practical applications in psychology.

So, how does this book relate to machine learning? For me, it poses an interesting research direction along with a key warning.

The research direction is how to use analogs of emotions to improve a computer's ability to understand its surroundings. Currently, we develop cost/loss functions that one may think of as a coarse approximation of emotional experience. However, the way that machine agents are trained, we essentially require them to immediately react (analogous to fight or flight) to sensory input. Perhaps deeper or more complex models build some internal representation that allows for the type of symbolic development that we see humans go through. Unfortunately, the way that machine agents are trained and deployed, we ultimately ask them to directly transform sensory input into action; so, it is hard to say conclusively that we are in any way replicating the self-regulating utility of emotional experience.

However, this leads immediately to the question: should we replicate emotions? In The First Idea, the authors show that the gradual development of more complex emotions leads naturally to a sense of self, and to consciousness. Assuming that we have the hardware to support computational emotion, is it humane even to attempt to create conscious beings? And (wearing my engineer hat), do we need computers to experience a sense of self to effectively solve the types of problems that we care about? To be clear, I am specifically talking about problems that are so large and challenging that the only way we know how to solve them is by training massive neural networks. For example, the problem of identifying pedestrians for an autonomous car. Currently, neural networks do an amazing job, but also experience severe and unpredictable failures. So, do machine agents need emotional experience to be reliable for safety-critical tasks?

My current answer is no, based on the notion of competence without comprehension. To better understand this idea, check out From Bacteria to Bach and Back by Dan Dennett (which I'll blog about eventually). To see this idea cast in a sci-fi setting, check out Blindsight by Peter Watts.

As a final note, I have massively enjoyed The First Idea, and strongly encourage everyone to read it!