Eric Jonas and Konrad Kording just released a provocative paper, “Could a neuroscientist understand a microprocessor?” In their paper, they ask whether current neuroscience techniques could discover the
operations of a simple microprocessor.
Their reasoning is as follows. The field of neuroscience is trying to understand the computational properties of the brain. If we think current neuroscience techniques are sufficient to understand something as complex as the brain, surely they will be able to handle a small microprocessor. If, on the other hand, current techniques are insufficient to understand even this simple CPU, it raises serious questions about the current approaches in the field. True, the brain is not a silicon processor but there are similarities (they list several in the paper). So let’s apply these techniques to this simpler computational system as a litmus test.
Their methodology includes applying an array of traditional techniques such as lesioning, examining statistics of bit patterns, analyzing tuning properties of transistors, dimensionality reduction, etc. They studied the microprocessor “in vivo” while playing a variety of video games (you can tell they had a lot of fun with this project!) They were able to discover that transistors exhibited very low pairwise correlations but were not actually independent (very similar to behavior of neurons). They showed strong spatiotemporal structure in the activity of various processor components. The resulting plots and charts look remarkably similar to those in neuroscience papers. Yet these techniques did not uncover the true computational nature of the microprocessor nor its functional structure.
Of course, in this process what they are really asking is “Could a neuroscientist understand the brain”? Their conclusion: an unequivocal “NO”.
We all know that “correlation does not imply causation”. Current statistical techniques report all sorts of correlations, but little regarding true underlying structure.
What does it mean to understand the brain?
So what does it really mean to understand the brain? Unfortunately the paper does not answer this question. They do make vague comments that the field should understand “how the output relates to the inputs”, and that it should reward “those who innovate methodologically.” These statements are unsatisfactory at best.
I propose a much stronger answer. As a computer scientist, I believe the only way to be certain you understand something is to build it. Write the program for it. We don’t need to create an exact replica, just a system that demonstrates the important properties. This methodology is harder but demands that you uncover underlying structure and function.
Let me give an analogy using a different paradigm. Suppose cars didn’t exist. Humans somehow get access to a luxury Mercedes sedan and the race is on to understand how it works. Let’s consider two alternative approaches.
The first approach involves calculating a number of statistical measures and building predictive models. These models might accurately predict the car’s gas mileage under different conditions, such as going uphill vs downhill. They might be able to plot precise acceleration profiles under different loads. They would know exactly how long it takes the air conditioning to cool the car in different climates. Scientists would publish thousands of peer-reviewed papers with all sorts of equations and charts proving the accuracy of these models. But would they really understand how the car works?
The second approach involves using the car to deduce fundamental mechanisms such as a power source, transmission, and steering. It would focus on the function of these subsystems, and less on details such as the strength of bolts or the efficiency of water pumps. To test our theories of function we would build a much simpler machine from scratch, perhaps something like a Ford Model T. This car would have a super simple engine and hand cranked starter. The controls might be awkward, the tires bad, and the seats uncomfortable. It would definitely have no air conditioning. It might not even be as fast as a horse! But, you could actually drive this car. Because we understand how our simple car works, over time we can improve it and eventually build vehicles even better than the Mercedes prototype.
What is our approach?
At Numenta we are using this second approach to understanding the brain. We are building the Model T equivalent of the neocortex. We use neuroscience discoveries and details to deduce the fundamental components of intelligence. For example, we know that the neocortex learns a predictive model of the world. It learns continuously without supervision. We know that behavior and sensory inference are not separate processes, but are intimately integrated such that learning cannot be achieved without behavior. Our theories are constrained by and consistent with a great many neurosciences details, but our software simulations only capture the functional properties of the brain and not all the details. We are often asked, “How do you decide what neuroscience details to include in your simulations and which to leave out?” The answer is we include neuroscience details when they are essential for function. When we hit stumbling blocks we return to experimental neuroscience to provide clues and hard constraints on how to solve problems. Our simulations also provide insights into the structure of cortex.
Compared to the brain, our software is at an early stage and primitive, like a Model T. But, you can actually take it for a spin, see how it performs, and then know what areas need improvement.
Can we understand the brain? We have made excellent progress with our approach. If we stay focused on large-scale functional theories and building systems based on those theories, the answer to this question is an unequivocal “YES”!
Footnotes and Citations
 Jonas, E., and Kording, K. (2016). Could a neuroscientist understand a microprocessor? Cold Spring Harbor Labs Journals bioRxiv doi:10.1101/055624. http://biorxiv.org/content/early/2016/05/26/055624.abstract
 Specifically Motorola 6507, similar to what was used in the Apple I and Atari video game consoles 40 years ago.
 Schneidman, E., Berry, M. J., Segev, R., and Bialek, W. (2006). Weak pairwise correlations imply strongly correlated network states in a neural population. Nature 440, 1007–12. doi:10.1038/nature04701.
 Note that the paper specifically targets computational neuroscience techniques, not experimental neuroscience. The amount of experimental data in neuroscience has been exploding exponentially, but without good theoretical guidance uncovering value is like finding a needle in a haystack.
 Hawkins, J., and Ahmad, S. (2016). Why Neurons Have Thousands of Synapses, a Theory of Sequence Memory in Neocortex. Front. Neural Circuits 10. doi:10.3389/fncir.2016.00023. http://journal.frontiersin.org/article/10.3389/fncir.2016.00023/full
 Ahmad, S., and Hawkins, J. (2016). How do neurons operate on sparse distributed representations? A mathematical theory of sparsity, neurons and active dendrites. arXiv:1601.00720 [q–bio.NC]. Available at: