Poster D1, Wednesday, August 21, 2019, 5:15 – 7:00 pm, Restaurant Hall
Post-hoc modification of linear models: take control of your machine learning algorithm
Marijn van Vliet1, Riitta Salmelin1;1Aalto University
Machine learning models have enabled the use of increasingly ambitious experimental designs for studying the neurobiology of language. For example, single trial analysis of neuroimaging data (e.g EEG, MEG, fMRI) is now possible. However, it can be daunting to figure out what a model is “learning” about the data and assert control over it. In this study, we propose a framework for understanding linear models (e.g. OLS, lSVM, logistic regression, LDA, etc.) that allows for a back-and-forth between the learning algorithm and the researcher. First, the model is fitted to the data as usual. Then, its weight matrix is decomposed into a covariance matrix, a pattern and a normalizer. These subcomponents are much easier to reason about than the original weights and it is, therefore, also straightforward to modify them based on domain information. Finally, the modified subcomponents are re-assembled into a weight matrix, yielding an updated linear model. We demonstrate the operation of this framework on EEG data recorded during a semantic priming experiment. The task for the machine learning model was to deduce the associative strength between two words based on the EEG response. The words were presented sequentially in written form. By manipulating the subcomponents of the model, we were able to improve its performance by leveraging domain knowledge about the characteristics of the EEG method, the time course of the N400 potential, and the recordings of other participants. The improved decoding performance serves as one example of the primary goal of this framework, which is to have more control over what a model is doing.
Themes: Computational Approaches, Methods
Method: Functional Imaging