What do we understand about the brain? That’s a surprisingly tricky question. But first, here’s another one: what does it mean to “understand” a biological system? There’s currently a lot of disagreement about the first question, which I believe lies primarily in unspoken differences in how we answer the second. So let’s break it down a bit.
The many faces of understanding
In colloquial usage, “understanding” has several meanings, which cluster into two camps: (1) comprehension and (2) explanation. Comprehension allows you to reduce a complex system down to simpler principles. These principles are simple enough that you could explain them to another person in relatively plain English - a good example is the Wired YouTube series in which an expert explains a single concept to different people (ranging from 7-year-olds to graduate students) at different levels. Having to verbalize this information is a pretty high bar, because it requires that you’ve also encoded the meaning behind the system’s components - you see how they fit together and form an integrated whole. Full disclosure: I’m on Team Comprehension.
In contrast, explanation is a more strictly statistical term. It’s often used to refer to how well a model - such as a regression - can account for variance in the system’s behavior (you can capture this in several ways, such as R-squared or AIC. Cross-validation techniques that assess a model’s predictive abilities can measure how well a model accounts for the data more robustly, but can still support explanation without comprehension). For some cognitive neuroscientists, finding a model that explains the data well is tantamount to understanding the brain.
If you grant that, a natural conclusion is that you should include as many parameters in your model as possible, because by chance some proportion of those will have explanatory power. This is true even if you’re using a predictive framework with cross-validation. Cross-validation can mitigate the risks of over-fitting (generally cross-validated performance will fall if you include too many predictors and throw off your data’s rank), but even if you regularize your model to shrink the influence of unimportant predictors, with too many to choose from you could end up with a well-performing model by chance - so, if you simply want to explain the data, adding in lots of parameters is a great strategy.
Maybe this is biologically reasonable - the brain is highly complex, so perhaps you need millions of moving parts to build a model brain. Many prominent scientists consider this to be the central project of cognitive neuroscience. Why build an artificial brain? You could use it to build a robot with human-level social processing or try to interpret locked-in patients’ brainwaves – both totally legitimate. You could learn how it works by taking it apart, like a car engine. That’s very tempting, but peering into the engine of a million-parameter model is daunting. Maybe not impossible - researchers are making headway in developing new methods for interpreting high-dimensional models like deep neural networks - but still unlikely.
Case study: neural networks
In recent years, various species of deep learning networks are the high-dimensional model of choice for the brain. They’re capable of human-level performance on some visual recognition tasks and their hierarchical architectures are a reasonable match for the ventral pathway of the visual system, both in theory and in fMRI data, at least in some domains. Can they help us to understand the brain, in either sense?
I think that you can, if you’re careful about the kind of understanding you seek. Concluding that an area of the brain is doing the same thing as a neural net because it fits activation patterns in that region well is shaky. In principle, any model could fit the brain well with enough parameters (even with regularization, as explained above). More to the point, just because a model behaves like a human doesn’t mean that they’re executing the same processes to produce that behavior, and we shouldn’t assume otherwise without extensive validation tests.
However, I think that we can gain both explanation and comprehension from neural nets in some cases. They are a good case for proof-of-concept arguments - e.g., if a neural net can classify images given their percept alone, then a human brain, in principle, could do the same (there probably is still some semantic structure introduced into a model by providing labeled categories for the output layers, but that’s another story). Similarly, if a neural net that processes motion and form features more successfully replicates MEG patterns from watching action videos than one that only processes form features, that suggests that motion features are also important for the human perceptual system. But in all of these cases (as in most of science), the comprehension is partial and needs to be coupled with complementary data.
Moving toward understanding
No doubt this debate will rage on, and I’m curious to see where it leads the field. Obviously, it’s optimal to have both explanation and comprehension - explanation lends statistical rigor to any understanding, and it’s better to have a model that explains real data than a lonely box-and-arrow theory with tractable concepts. But when pushed, I would rather sacrifice some explanatory power to gain better comprehension. In my view, that is the kind of understanding that really drives science, and mere explanation is simply unsatisfying.
