The "Brain from Inside Out" discusses György Buzsáki's ideas about how to accelerate progress in neuroscience and applications of neuroscience in healthcare and artificial intelligence. Buzsáki's overarching thesis is that neuroscience research is stagnating because we are so tied down by what he calls the "Outside-In framework", which assumes that the goal of the brain is to truthfully represent features of the outside world. In this framework, in order to learn what the brain does, researchers typically go about manipulating features of stimuli in the outside world, and observing how the brain behaves in response to those manipulations. Instead, inside-out framework, which he advocates for, proposes treating the brain and its dynamics as the independent variable, and exploring how those mechanisms give rise to action or measurable performance in perception, emotion and cognition. Particularly, there should be an emphasis on how an animal's actions influence incoming signals in the brain because this comparison is how the brain generates meaning.
The first major thesis is that a major source of error in the outside-in approach comes from using pre-neuroscience terms inherited from psychology to direct our neuroscience experiments. This is easy to do with an outside-in approach because in that framework we are already looking for external things (like our own ideas about psychology, or the objects/features that we observe in the world) to correlate with internal mechanisms of the brain, so a dictionary of terms from psychology is a really convenient place to start. For instance, we look for "attention" inside the brain, even though we don't know if the brain is actually doing one specific thing that could be called attention. There are several chapters in the book that provide explicit examples of where our inherited terminology is likely inaccurate: From Chapter 5: we think of episodic memory and spatial navigation as separate "cognitive" tasks but they are actually represented by the same mechanism. The same seems to be true of structuring semantic knowledge and the entorhinal grid systems. From Chapter 9: we often look for how the brain constructs semantic memory…but the way that we think about declarative, semantic memory is really odd because of how many tools we make. Since we externalize so much of our thought, the process of episodic memory transforming into semantic memory isn't as clear-cut for humans as for other animals, and really difficult to study from the perspective of internalized function. From Chapter 10: many researchers study how we perceive space and time in the brain. However, in the chapter we learn that space and time are not represented by neurons specifically - instead neurons represent velocities in many ways by grounding to actions of the bodies and just link up sequences. From Chapter 11, we learn that even though many researchers go looking in the brain for something called "attention," attentional mechanisms are likely gain control mechanisms, but there are a huge variety of gain control mechanisms (implying that there may be multiple kinds of attentional mechanisms), and there are many gain control mechanisms that we don't consider to be "attention," such as moving between reference frames or abstraction. From Chapter 13 we learn that comparing ratios is something that the mechanisms of neural dynamics does intuitively, but that counting is not - even though we often talk about the two as if they are the same. For instance, these are approximately the same at very low numbers, and so we wrongfully identify animals that can "count" to low numbers rather than considering counting a sophisticated externalized technology. Cognitive researchers often equate looking for counting or math in the brain to looking for ratio comparison or pattern recognition. Buzsáki believes that mathematics and counting are technologies, while ratio comparison and pattern recognition are true features of the brain, and that we should be careful about separating them.
If we take an inside-out approach, we start by investigating attributes of the brain and then make hypotheses about how those attributes affect action or behavior. Therefore, the inside-out approach would let us not focus so much on believing in our intuitive ideas from psychology. Besides making progress in understanding the brain, the impact of moving away from inherited names from psychology to internal mechanisms would have a profound effect on healthcare, because we would better understand the range of brain mechanisms that could cause some subjective experience (say an auditory hallucination), and we would strive to prescribe a treatment according to the mechanism rather than for the symptom.
The second major thesis is that in order to understand the "code" of the brain, we need to understand how the code is read, and the best way to do that is by studying action. So basically, the logic is that something inside the brain "understands" what the brain is doing, even if we don't. Buzsáki calls whatever is decoding a message the "reader." Observing actions (which are defined as motor actions, autonomic activation, hormone release, or thoughts in the form of buffers to action plans) is a useful approach for two reasons. One is that the experimenter can understand actions even if we don't understand the non-motor action-readers right away. Action is inherently a "reader" that has processed all of the incoming information into some kind of output - so we can kind of follow it backwards to try to get some sense of how it is decoding whatever is going on in the brain. Another reason is that the brain initially needs to learn meaning in the world, and can only do that through observing the correspondence between its actions and how things change in the world. Because of this, he thinks that observing how actions influence incoming signals in the brain is a useful approach. He gives some very interesting examples connecting meaning to action. For instance, people with amygdala damage can't accurately predict if someone's face is showing fear in experiments, so a major narrative in neuroscience is that the amygdala processes fear stimuli. But people with amygdala damage do notice fearful people in the real world, when they aren't looking just at pictures in an experiment, they still report feeling fear themselves, and if explicitly directed to look at the eyes of the face in the pictures, those patients can accurately identify fear. Buzsáki thinks, instead, that circuits in the amygdala are probably involved in eye search patterns.
The third major thesis of the book is that the brain has pre-formed connectivity and dynamic operations that constrain the potential activity in the brain. The brain matches pre-existing possible neural trajectories that exist within these constraints to new information about the world. This is actually very important to keep the brain stable, and he points out that ideas like Hebbian learning, where synaptic weights are increased by neurons that fire together, on its own would simply lead to run-away excitation and epilepsy. He talks about a lot of these internally organized brain patterns in chapters 6, 7, 8, and 12. He suggests that brain oscillations that are on a log-normal distribution, formed by cell assemblies synchronously spiking, are the "letters" of neuronal communication. A certain number of fast gamma oscillations can fit within slower oscillations, which are used to synchronize larger territories of brain. Interestingly, despite huge differences of sizes of mammal brains, the frequencies of oscillations are conserved across species, which suggests that the specific frequency ranges of oscillatory activity are very important for brain function. Buzsáki proposes that the hippocampal system and cortex trade-off being the "reader," especially during waking and sleeping, with different intrinsic rhythmic systems that send information. Waking is dominated by gamma oscillations embedded in theta rhythms that seem to mark sequence trajectories in navigation, which is interpreted an interaction between a sequence generator (the hippocampal system), which relates together the vast range of neural trajectories that could come from the cortex, where-as sleeping, sharp-wave ripples are a major synchronous event in the hippocampus that seem to consolidate memories in the cortex by fast replay of organized spike sequences forward and backward, that can then be used as sequence snippets in long-term memory. The oscillatory activity clearly serves functions and is understood by readers in the brain, but also the oscillatory dynamics are set up to make the brain energy efficient and stable. A constrained log-normal distribution of spiking frequency of neurons in the brain also makes the brain stable against run-away excitation, but also capable of balancing fast decision making capabilities with slower, more specific processing.
These two theses together, of the importance of action loops to ground meaning and constraints inspired by mammalian brain dynamics, are particularly applicable to the problem of creating better artificial intelligence or generalized artificial intelligence. He also gets really creative at one point, and suggests that certain mechanisms that define constraints on brain dynamics may be useful for human societies as well. Basically he observes that the log-normal distribution of neuronal firing rates, which are extremely important for brain function - because we have a balance of high firing rate, rigid, good enough, generalists, and very flexible, low firing rate, specific, specialists - would lead naturally to huge inequalities in the relative power of the 10-20% of high firing rate neurons vs. the 50% of lowest firing rate neurons. But the brain has mechanisms during sleep where the distribution is fixed to constrain the inequality within some limits. Instead, usually societies, which have power and wealth inequalities that also follow log-normal distributions, will have continuously growing inequality until there is a revolution.
Discussion Questions:
- Can you think about ways to do "inside-out" research that is non-invasive? It seemed like measuring and manipulating brain dynamics is actually pretty standard in well-funded behavioral and systems neuroscience labs. It seems difficult to treat the brain dynamics as an independent variable in other domains - like non-invasive brain imaging.
- Are you as pessimistic as Buzsáki about the outside-in approach? Terminology is constantly changing to match observations in conventional neuroscience with the outside-in approach. In fact, most of the examples, justifying the inside-out approach, came from outside-in studies.
- Throughout the book, I kept thinking that our hypotheses might be more affected by our technologies than our philosophy, as Buzsáki suggests. For instance, there is a strong narrative that the amygdala underlies fear processing, whereas Buzsáki proposes that the amygdala mainly directs eye movements for fearful stimuli. Perhaps we would think of brainstem areas as underlying fear processing if those areas were large enough to image in typical fMRI experiments.
- Buzsáki equates certain inhibitory activity to "shunting" in electronic circuits. I did not know what this meant, and if you want to learn more, here is the wikipedia page on electronic shunting.
My take:
I really loved this book. If someone asked me for a general book about what we know right now about how the brain works, I might suggest this book. It's not an introduction to the brain, but includes so much of the important modern findings that frame what we know. We obviously have an enormous amount to learn about how the brain works still, but it always rubs me the wrong way when people say that we "don't have any idea how the brain works" or similar sentiments. This book does a wonderful job drawing the outline of what is known and what is not known, uses experiments well to justify the knowledge, and I never feel like it oversteps. Buzsáki is conservative and thoughtful, and says things like, "we only completely mapped this in two mammalian species, so it seems like it might be conserved, but we need to do more research." That might seem obvious to some readers, but is refreshingly cautious to me, when I see so many scientists discuss general truths about neuroscience or biology that are based only on data from one strain of mouse or, worse, yeast, often without even bothering to mention the source of the knowledge.
While the book is ambitious and compelling, I think that it could have been made a bit more coherent in a few ways. First, the main thesis of the book is clear, but I think that there are maybe three sub-theses that pop-up in multiple chapters and their connection is hard to fully synthesize. The book meanders and feels almost conversational, like we're having a long lunch date together. It's extremely pleasant and interesting to read. But there are a lot of ideas that take time to unpack, and I think I would have preferred a little more scaffolding. I'm also a neuroscientist myself, and so I felt like if I was a layperson, I probably would have felt a bit lost in understanding the point at times. I think the book could have been structured in a better way - like maybe an introduction with the overall idea, and then Part I - about philosophy, history, and our historical lexicon of psychology words and why that's problematic. Part II - about the importance of action grounding. Part III - about how the brain uses action ground to attach meaning to already existing brain dynamics. And then a synthesis showing how they're all connected and lumping a bunch of examples of how to do inside-out experiments and how useful they are.
Second, there are numerous examples throughout the book justifying the inside-out framework, and other examples of research within the inside-out framework that Buzsáki is advocating for. I didn't count, but it felt like most of the information was more on the "justification" side - what might just be called the "inside" framework. For example, Buzsáki describes the log-normal distributions of many features of the brain. Great; this is really interesting stuff! Then he describes psychometric experiments showing log-normal distributions in sensory thresholds. Also, very cool! We can call psychometrics the "outside" framework, since it doesn't measure the brain at all. And then he hypothesizes that the two are connected - that the "inside" observation causes the "outside" observation. I'M ON THE EDGE OF MY SEAT! I'm thinking of experiments that could connect them, but they're all outside-in because I'm a foolish, brainwashed human neuroimaging person. I would try doing silly things like measuring the intensity of brain activity, while manipulating sensory stimuli features. Then if I saw that the intensity of brain activity increased as a function of log-intensity of the stimulus, I would feel like I was on the right track. So I'm waiting to hear how to do the experiment the inside-out way. But it never comes in the book. I'm not saying that it doesn't exist. I just feel like Buzsáki holds my hand very nicely when describing the outside-in approach, which I know pretty well and actually don't need too much help to understand, but then he assumes that I'm smart enough to come up with the inside-out approach on my own (which I'm not).
Basically, I wish that my attention was explicitly drawn to how to do inside-out research. I understand that part of the purpose of the book is to inspire neuroscientists to think critically about inventing new paradigms or approaches to do inside-out research, but it would be nice to have clear directions. And there are examples of the inside-out framework in action throughout the book, but Buzsáki doesn't explicitly say, "Now pay attention here, buddy-boy, because this is how you should be doing it."
Finally, I think that György Buzsáki frames neuroscience as a whole as if it is shackled by the Outside-In Framework, while only he has freed himself from this imprisonment. I think the reality is more complex and differs based on what subfield of neuroscience you are looking at. Computational neuroscience, reinforcement learning people, motor people, and some memory people will probably not find Buzsáki's thesis too contentious or novel, but they would recognize Buzsáki's major role in the shifting perspective and themes of investigation. A lot of behavioral neuroscience and human neuroimaging research is characterized by the outside-in framework, but even this is shifting. For example, in functional human imaging, there is a huge push towards understanding spontaneous internal dynamics, and in EEG, there is an exploding interest in the 1/f characteristics of oscillatory power. There are widespread conversations about how to refine the definitions of the words that think map onto real cognitive processes in the brain. The inside-out framework, where you treat the brain as the independent variable, is hard to do in human neuroimaging research because our best tools for manipulating the brain are invasive. I don't think that Buzsáki would argue with this point either. I just wish that he would have contextualized his thesis as a slowly changing perspective in neuroscience research that he has been a major part of, and that the point of the book was to explicitly outline this experimental change and to provide examples of how to set up your experiments within this framework.