It's Sunday morning in New York, and I really need a break from grant writing... So here a few cranky thoughts from a cranky guy.
Is modern neuroscience on the right path – or showing signs of early demise?
Two of the most exciting areas of contemporary science are to understand the structure of the universe and the structure and function of the human brain. The problem, incidentally, turns out to be similar in scope. Our neighborhood galaxy, the Milky Way, is estimated to contain about 100 billion stars. The human brain is estimated to contain about 86 billion neurons. Although different in kind, the scale of these scientific questions suggests that we need experimentation and quantitative analysis over enormous numbers and astonishing complexity.
The galaxy is big - and distant. The brain is small and between our ears. But the brain may be the bigger challenge. Because of the almost perverse connectivity of cells with each other and the extremely rapid processing timescales (milliseconds), understanding how the brain is organized and in which way it forms the basis for everything we do - from regulating are heartbeat to making complex decisions to enjoying music - is clearly one of the most profound challenges for modern science. That is presumably why we like it and do it, no?
As neuroscientists, we celebrate a veritable orgy of successes, suggesting that from molecule to mind, our techniques are steadily illuminating each aspect of brain function. We capitalize on cutting-edge developments, and practically every one of our graduate students is now a literate in big data and deep learning.
This is all wonderful, and keeps us employed. Scratching on the surface however, reveals that there are deep underlying problems – and the acute danger that human brain science may be merely ‘descriptive,’ epistemologically flawed, and perhaps going down a completely dangerous and even sterile path. Indeed, it appears that in the neurosciences, "science" is being replaced by engineering... Provocative papers have begun to point out that this ever growing field is departing from approaches that have been successful, for better or for worse (e.g. Krakauer er al, 2017, Neuron).
The history of science tells us that concepts such as ‘explanation,’ ‘mechanism,’ ‘theory’ lie at the foundation of modern science since the enlightenment.
Remarkably, contemporary neuroscience has increasingly divorced itself from the standard approaches of the sciences. This may be a good thing and legitimate and ‘epistemologically innovative’ – or it may be catastrophic and lead to the sterility of the neurosciences ... If explanation and mechanism are replaced by concepts such as prediction (“As long as my model predicts with a good fit, I am happy” prominent senior neuroscientist), if understanding is in fact no more than model-based prediction, then this exciting and growing branch of the sciences will have departed from the aims that we typically pursue in the sciences. We will continue to embrace big data - and abandon big theory.
To be sure, nobody argues against having access to large amounts of data. Whether collecting genomes, transcriptomes, connectomes, any-other-kind-of-omes, physiological recordings from hundreds of cells, or brain imaging from tens of thousands of locations, everyone realizes the value of the ever-growing body of data we can measure and analyze. There is very little risk in having more information. Perhaps the marriage of big data and relentless reductionism really is the answer to our questions.
That being said, there is a huge risk in having too few ideas, too few guiding principles, too few (warning: old school concept ahead…) “hypotheses” that provide the intellectual substrate that motivates why we do the work to begin with! And let’s be clear: just saying ”I have a computational model” that underlies my data analysis a theory not makes. In short, we are at a critical juncture in the neurosciences. We will, that goes without saying, continue to adopt the methods associated with big data and deep learning, because they are important expansions of the toolboxes we have to examine these most complex set of phenomena. But if we replace the fundamental, mechanistic, explanatory understanding of biology, brain, cognition, with data-driven correlation, regression, model fitting – when our data have become our theory – will the price be too high?
Examining these ideas -- that is to say engaging in a systematic, sustained, and perhaps countercultural and unpopular reflection on where this most important field is going -- will illuminate whether the next 30 to 50 years in the neurosciences will constitute the success of correlation over explanation, or whether there is a path forwards that recaptures the value of ideas that form the basis of fundamental understanding. With computation, broadly construed, sitting centrally at the issues facing this field, it will be critical to assess (and recommend) whether we should continue to replace scientists with engineers and whether we should abandon the Baconian ideals of how the sciences have worked for the last several hundred years.