Discover more from Overcoming Bias
Separate Top-Down, Bottom-Up Brain Credit
Recently I decided to learn more about brain structure and organization, especially in humans. As modularity is a key concept in complex systems, a key question is: what organizing principles explain which parts are connected how strongly to which other parts? (Which in brains is closely related to which parts are physically close to which other parts.) Here are some things I’ve learned, most of which are well known, but one of which might be new.
One obvious modularity principle is functional relation: stuff related to achieving similar functions tends to be connected more to each other. For example, stuff dealing with vision tends to be near other stuff dealing with vision. But as large areas of the brain light up when we do most anything, this clearly isn’t the only organizing principle.
A second organizing principle seems clear: collect things at similar levels of abstraction. The rear parts of our brains tend to focus more on small near concrete details while the front parts of our brain tend to focus on big far abstractions. In between, the degree of abstraction tends to change gradually. This organizing principle is also important in recent deep learning methods, and it predicts the effects seen in construal level theory: when we think about one thing at a certain level of abstraction and distance, we tend to think of related things at similar levels of abstraction and distance. After all, it is easier for activity in one brain region to trigger activity in nearby regions. The trend to larger brains, culminating in humans, has been accompanied by a trend toward larger brain regions that focus on abstractions; we humans think more abstractly than do other animals.
A key fact about human brain organization is that the brain is split into two similar but weakly connected hemispheres. This is strange, as usually we’d think that, all else equal, for coordination purposes each brain module wants to be as close as possible to every other module. What organizing principle can explain this split?
There seems to be a lot of disagreement on how best to summarize how the hemispheres differ. Here are two summaries:
The left hemisphere deals with hard facts: abstractions, structure, discipline and rules, time sequences, mathematics, categorizing, logic and rationality and deductive reasoning, knowledge, details, definitions, planning and goals, words (written and spoken and heard), productivity and efficiency, science and technology, stability, extraversion, physical activity, and the right side of the body. … The right hemisphere specializes in … intuition, feelings and sensitivity, emotions, daydreaming and visualizing, creativity (including art and music), color, spatial awareness, first impressions, rhythm, spontaneity and impulsiveness, the physical senses, risk-taking, flexibility and variety, learning by experience, relationships, mysticism, play and sports, introversion, humor, motor skills, the left side of the body, and a holistic way of perception that recognizes patterns and similarities and then synthesizes those elements into new forms. (more)
The [left] is centered around action and is often the driving force behind risky behaviors. This hemisphere heavily relies upon emotional input leading it to make brash and uncalculated decisions. … The [right] … relies primarily on critical thinking and calculations to reach its decisions. As such the conclusions reached by the [right] often result in avoidance of risk taking behaviors and overall inaction. … . In environments of scarcity, … taking risks is the foundational approach to survival. … However, in environments of abundance, as humans have observed, it is far more likely to die to damaging stimuli. … In areas of prosperity, … [right] domination is prevalent. … In areas of scarcity where cold and limited food are concerns [left] domination is prevalent. (more)
After reading a bit, I tentatively summarize the difference as: the right hemisphere tends to work bottom-up, while the left tends to work top-down. (In a certain sense of these terms.) Inference tends to be bottom-up, in that we aggregate many complex details into inferring fewer bigger things. For example, in a visual scene we start from a movie of pixels over time, and search for sets of possible objects and their motions that can make sense of this movie. In contrast, design tends to be top-down, in that to design a path to get us from here to there, we start with an abstract description of our goal, such as the start and end of our path, and then search for concrete details that can achieve that goal.
The right hemisphere tends to watch, mostly looking out to infer danger, while the left tends to initiate action, and thus must design actions. The right has a wide span of attention, watching the world looking out for surprises, most of which are bad, while the left has a narrow focus of attention, which supports taking purposive action, from which it expects good results. So the right hemisphere tends to do bottom-up processing, while the left does top-down processing.
In bottom-up processing, to explain one set of details one must consider many possible sets of abstractions, while in top-down processing, one set of goals gives rise to many possible specific details to achieve those goals. As a result, we should expect bottom-up work to need more resources at high abstraction levels, while top-down work needs more resources at detailed levels. And it fact, this is what we see in brain structure: the right hemisphere has a larger front abstract end, while the left hemisphere has a larger back concrete end. Our brains are “twisted” in this predicted way.
Why would it make sense to separate bottom-up from top-down thinking? A key problem in the design of intelligent systems is that of how to distribute reward or credit. And a common solution to this problem is to create a standard of good in one part of the system, today often called a “cost function” in AI circles, and then reward or credit other parts of the system for getting closer to achieving that standard. In inference, the standard is typically some form of statistical fit: how well a model of the world predicts the data that one sees. In design, the standard is more naturally centered on goals: how well does a plan achieve its goals?
Top-down and bottom-up styles of processing seem to me to use incompatible systems of credit assignment. That is, it seems hard to design a system that simultaneously credits abstract world scenarios for predicting details seen, while also rewarding details chosen for achieving abstract goals. Credit assignment systems work better when they have a single common direction in which credit flows. One can allow multiple design goals at a similar high level of abstraction, as then the design process can give credit for synergy, and search for details that satisfy all the goals. And one can allow multiple sources of detail, like sight and sound, and combine their statistical credit to infer which objects are moving how. But it seems hard to combine the two systems of credit.
And so that is my proposal for a third organizing principle of brains: separate bottom-up from top-down systems of credit assignment. I haven’t heard anyone else say this, though I wouldn’t be surprised if someone has said it before.
Added 1Sep: The main risk of mixing credit directions is creating self-supporting credit cycles not well connected to real needs. This may be why the connections between the two hemispheres are mostly inhibitory, reducing activity.