277 — The Roles of Supervised Machine Learning in Systems Neuroscience

Glaser, Benjamin, Farhoodi & Kording (1805.08239)

Read on 24 May 2018
#neuroscience  #machine-learning  #neural-network  #random-forest  #brain  #architecture  #spiking-neural-network  #model  #systems-neuroscience  #supervised-learning 

Some huge number of the papers I’ve read so far for this project have involved — in some form or another — both neuroscience and machine learning in conjunction with one another. Sometimes that is in the form of computer vision for image segmentation; sometimes it’s in the form of using neuroscientific discoveries of the past in order to design more robust, more biofidelic, or more interesting artificial neural networks. And a few have even discussed how artificial neural networks suck compared to the biological counterparts.

This preprint looks specifically at how machine learning has influenced (and will continue to influence) neuroscience — effectively, the opposite direction of information flow compared to the conventional notion of designing biofidelic systems.

The authors classify these influences into four main categories:

Solutions to engineering challenges. Neuroscience research raises many complex questions regarding the relationships between input “knowns” and output predictions. ML assists — intuitively enough — in this sort of classic challenge of understanding input/output correlation.

Predictive variables. This is the evaluation of a large set of inputs, in pursuit of learning which correlate most closely wth a known output. The most clear-cut example of this is passing arbitrary neurological data to a ML algo in order to attempt to predict or diagnose disease.

Setting benchmarks for simple models. Machine learning techniques are very well-suited to the problem of distinguishing between how much of a model’s failure is attributable to shortcomings in the model and how much is attributable to noisy input data or low-fidelity signal.

Acting as a model of the brain. Computational models — artificial neural networks — are quite different from the endogenous, spiking neurons of the brain. But the two systems have similarities as well: Tuning curves appear to be a fundamental component of distributed-computing systems that follow the simple node-edge architecture. This will only shed more light on the brain as spiking neural network models grow in popularity and efficiency.