Modern machine learning as benchmark for encoding models¶
Foreword
This chapter provides a machine learning tool to verify the assumptions of simpler, interpretable encoding models. By acting as performance benchmarks, machine learning methods can identify if nonlinearity remains to be explained. This work is also an instructive demonstration of which machine learning methods are most successful at neural prediction. This paper is an example of Role 3: machine learning as a benchmark for simpler models.
In a collaboration with Prof. Lee Miller at Northwestern University, we applied this approach to datasets collected from motor and somatosensory cortices. Many of these machine learning methods had not previously been used to predict neural activity, and we found they empirically worked quite well and outperformed GLMs at predicting spikes in three separate brain areas. This result is a warning that GLMs may generally fail to capture neural nonlinearity and can mischaracterize stimulus/response relationships for naïvely chosen sets of features.
This chapter is reproduced from a paper now published in Frontiers in Computational Neuroscience , for which it was selected as Editor’s Pick 2021. It was also presented at two conferences, in the form of a poster at COSYNE 2017 and a talk at SAND8 in 2018.