What is our brain for? Making predictions. Why? Because that’s one way for a living organism to survive and possibly thrive in an environment that’s constantly changing. In the words of Andy Clark, author of Surfing Uncertainty, the brain is “an action-oriented engagement machine, adept at finding efficient embodied solutions that make the most of body and world.” I’m glad Clark provided that pithy summary at the end of his book. Because I’m not a neuroscientist, it took me a while to work my way through his argument. But I’m glad I did because it made me think a lot about how humans learn and about my origin-of-life research; both are key topics I think about a lot in my professional life.
I haven’t fully digested his argument which is essentially using a model he calls Predictive Processing (PP) to explain what the brain does and why. Many open questions remain, and Clark early on acknowledges that the specific details of his model may turn out to be wrong, but that the overarching idea of top-down predictive processes coupled with bottom-up error-signaling processes work together in concert to home in on a best guess of any encountered situation. But this isn’t an isolated brain in a jar. Embodied action is a critical part of honing the process. I will quote parts that really struck me and muse about them briefly in a meandering way. Like a surfer perhaps. This may be fitting given the title of his book.
More than a decade ago, when I first encountered the notion of System 1 and System 2 thinking (made famous by Daniel Kahneman’s Thinking Fast and Slow), I was enamored by the idea. But over time I’ve found the separation a little too clean. Clark argues they are one multi-faceted system. We might “use some quick-and-dirty heuristic strategy to identify a context in which to use a richer one, or use intensive model-exploring strategies to identify a context in which a simpler one will do. The most efficient strategy is simply the (active) inference that minimizes overall complexity costs… system 1 and system 2… are now just convenient labels for different admixtures of resource and influence, each of which is recruited as circumstances dictate.” I have a feeling Clark is correct and that his emphasis on multi-timescale processes is a key part of how organisms do what they do. I don’t quite understand how the longer timescale ‘higher-level’ brain processes couple to shorter timescale sensory signals, but I suspect the dynamic coupling of such processes is the beating heart of life.
Thermodynamic terms show up in Clark’s treatise. There’s free energy minimization when the brain tries to be efficient and make a prediction at the lowest cost. It’s why we continue to make mistakes (and learn from them) as we encounter new situations or variations of what we thought were things we knew. Entropy is defined in terms of surprisal; when prediction goes awry and we have an oops moment, this allows us to recalibrate. As a chemist, I define these terms differently, but I see a kinship between how I think about thermodynamics and what Clark is trying to do with these terms. However, having seen thermodynamic principles invoked in multiple areas, in my opinion I see more and more muddied thinking that may introduce more confusion than clarity.
I very much appreciated Clark’s emphasis on perception and action being inseparable. He writes that they are “two sides of a single computational coin. Rooted in multilevel prediction-error minimizing routines, perception and action are locked in a complex circular causal flow… Percepts and action-recipes here co-emerge, combining motor prescriptions with continuous effort at understanding our world.” While I mostly thought of sensory signals as exteroception, I appreciated Clark’s reminder that proprioception and interoception are just as important, and our brain needs to make sense of all three incoming channels. This made me ponder how to include all three in origin-of-life modeling, and also how to structure the seeming digital-analog divide. Information is efficiently stored digitally, but the action of life is analog. I’m sure that different timescales are important here, but I haven’t figured out how these could or should be modeled.
In Chapter 6, “Beyond Fantasy”, Clark delves into the idea that “perception is controlled hallucination”. He thinks we should be circumspect about the notion that our brains and thoughts are akin to virtual reality. Action on our part is important to continuously update the “probabilistic prediction-driven learning… able to see past superficial noise and ambiguity in the sensory signal, revealing the shape of the distal realm itself.” But our brain has also evolved to be an efficient computing machine, and this means pruning out or ignoring a lot of the sensory stimuli to focus on what is salient. I’m reminded about the mystery of learning, especially when it comes to the nonintuitive subject of chemistry. When the aha moment occurs, it’s a gestalt experience. After that I can’t unsee what I now know. It also blinds me as a teacher through the curse of knowledge. It reminds me that I constantly have to work hard at teaching because things obvious to me are not obvious to students encountering it for the first time. I can provide helpful scaffolding but how one actually learns is still mysterious. And my learning needs to be continuously updated. I’m sure I have erroneous notions I’m still passing along to students, but they’re in my blind spot – and I won’t know until I’m surprised by them.
Uncertainty surfaces when you least expect it. Perhaps that’s the moral of the story.
No comments:
Post a Comment