The role of deep learning in neuroscience

"Hidden Treasure" by Joan Belmar
The brain is truly one of the final frontiers of human exploration. Understanding how the brain works has vast consequences for human health and for computation. Imagine how computers might change if we actually understood thinking and even consciousness. Neuroscience has focused on the detailed implementation of computation, studying neural codes, dynamics and circuits. In machine learning (ML), however, artificial neural networks tend to eschew precisely designed codes, dynamics or circuits in favor of brute force optimization of a cost function, often using simple and relatively uniform initial architectures. 

Two recent developments have emerged within machine learning that create an opportunity to connect these seemingly divergent perspectives. First, structured architectures are used, including dedicated systems for attention, recursion and various forms of short- and long-term memory storage. Second, cost functions and training procedures have become more complex and are varied across layers and over time. Here we think about the brain in terms of these ideas. In this post, I'll overview 3 hypotheses about the brain's storage of information and what they would entail. The first is that the brain optimizes cost functions, the second that the cost functions are diverse and differ across brain locations and over development, and, finally, optimization operates within a pre-structured architecture matched to the computational problems posed by behavior. 

In support of these hypotheses, we argue that a range of implementations of credit assignment through multiple layers of neurons are compatible with our current knowledge of neural circuitry, and that the brain's specialized systems can be interpreted as enabling efficient optimization for specific problem classes. Such a heterogeneously optimized system, enabled by a series of interacting cost functions, serves to make learning data-efficient and precisely targeted to the needs of the organism. We suggest directions by which neuroscience could seek to refine and test these hypotheses.

From my experience, which is albeit limited in both neuroscience and computer science, machine learning communities are absolutely picking up steam right now with the latest developments of the algorithms and the approaches. I almost want to say entrenched, but you know, the biological sciences have been around for a long time, and then the publication methods are the way they are for a reason. But there's a cultural change that the three of us have only been out of our PhD's now for, you know, less than a decade. We are definitely sort of seeing this enormous change and it's fun to be at a place that really sort of embraces that cultural change.

The artificial neural networks now prominent in machine learning were, of course, originally inspired by neuroscience (McCulloch and Pitts, 1943). While neuroscience has continued to play a role (Cox and Dean, 2014), many of the major developments were guided by insights into the mathematics of efficient optimization, rather than neuroscientific findings (Sutskever and Martens, 2013). The field has advanced from simple linear systems (Minsky and Papert, 1972), to nonlinear networks (Haykin, 1994), to deep and recurrent networks (LeCun et al., 2015; Schmidhuber, 2015). Backpropagation of error (Werbos, 1974, 1982; Rumelhart et al., 1986) enabled neural networks to be trained efficiently, by providing an efficient means to compute the gradient with respect to the weights of a multi-layer network. Methods of training have improved to include momentum terms, better weight initializations, conjugate gradients and so forth, evolving to the current breed of networks optimized using batch-wise stochastic gradient descent. These developments have little obvious connection to neuroscience.

In my research (currently on zebrafish neuroscience), I have a lot of data that has the potential to be used for ML projects. With the open-sourced nature of other sorts of data, this leads to many goals an ideas for projects at any given moment in AI (artificial intelligence). There's the chance for me to create goals in image recognition, signal processing, or simply redefining existing bioinformatics and computational neuroscience pipelines in efficiency and effectiveness. We often try to observe cells in movies and images we create of neuroscientific phenomena, and then extracting the activity of those cells. And then at the end of a pipeline of steps to analyze and process the data, of that kind of processing ML pipeline, a bunch of this data then comes back to me, where now I have signals, and I have the record of the images represented on the screen, I've got other data about when the zebrafish moves, when it didn't I'm trying to take this data and try to make sense of it. If I'm just looking at the activity, can I decode what was on the screen from that activity? If I'm looking at the activity, can I predict what the organism's choices were at any given time, whether it chose to lick or whether it chose not to lick, in the context of his performance on the game.

As I continue my research in computational neuroscience, it's just great to see that more and more people are seeing how important this is, especially from a government point of view. It's great that a president has realized that this is one of the biggest frontiers that we really have left to solve. We know so little, and so it's really great to see that the community as a whole is investing in this sort of research. From my perspective I feel like that was one of the most important things was the recognition in any setting (government, academic, or other places) that brain science is maturing into the type of thing that needs large data centers, it needs large sharing and collaboration tools, it needs large investigations to really start to make a difference in people's lives. The science and the community have matured to the point where we can really start taking that standard forward and making a big impact. I really use very standardized packages just because I want to stay away from people having to install and use roughly written code and the things I use the most are NumPy, SciPy, Stats Package. These are the tools in front of me I need to use to make the most out of my experience as a scientist. Whatever it takes for me to get the job done.

Cox, D. D., and Dean, T. (2014). Neural networks and neuroscience-inspired computer vision. Curr. Biol. 24, R921–R929. doi: 10.1016/j.cub.2014.08.026

Haykin, S. S. (1994). Neural Networks: A Comprehensive Foundation. New York, NY: Macmillan.

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521, 436–444. doi: 10.1038/nature14539

Minsky, M. L., and Papert, S. (1972). Perceptrons: An Introduction to Computational Geometry. Cambridge, MA: MIT Press.

McCulloch, W. S., and Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. Bull. Math. Biophys. 5, 115–133. doi: 10.1007/BF02478259

Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986). Learning representations by back-propagating errors. Nature 323, 533–536. doi: 10.1038/323533a0

Schmidhuber, J. (2015). Deep learning in neural networks: an overview. Neural Netw. 61, 85–117. doi: 10.1016/j.neunet.2014.09.003

Sutskever, I., and Martens, J. (2013). “On the importance of initialization and momentum in deep learning,” in Proceedings of the 30 th International Conference on Machine Learning (Atlanta: JMLR:W&CP).

Werbos, P. (1974). Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Sciences. Doctoral Dissertation, Harvard University, Harvard.