Professor Robert Tibshirani recently gave a talk about statistical learning. He reminded a fundamental formula about trying to extract insights from data: "try simple models first and move on to more complex methods, only if necessary".

Going deep directly is barely the best move unless you have a LOT of data. If you donâ€™t try simple things first, you will for sure miss important characteristics that could help to eventually use or design an appropriate model which goes deep. By directly going deep, you are giving up with interpretability which is often important in neuroscience though not always needed indeed. Most importantly, if you have wide data (number of features bigger than number of observations) or if you only have a moderate amount of observations (less than 500) then a simple LASSO will likely outperform a deep model. Deep neural networks (DNNs) usually need large quantity of data to perform well, which is often lacking in neuroimaging. In fact, He and colleagues (2018) have shown that DNNs did not outperform Kernel regression using almost 1000 subjects from the Human Connectome Project for predicting the fluid intelligence using functional connectivity (419 x 419 features). In neuroscience, many researchers used a DNN with less than 500 observations, still even more features and claimed a high accuracy. A model on such high dimensional data would face the curse of dimensionality (Hastie et al., 2009), thus not learn but overfit the observations and most likely fail to generalize. As Bzdok and Yeo (2017) stated: always keep in mind that the huge success of DNNs (in many different application domains) is partly due to colossal sample sizes with n > 1,000,000 (LeCun et al., 2015; Jordan et al., 2015). In neuroscience nowadays, the biggest datasets reach about 1000 participants (Human Connectome Project) to about 10,000 participants (UK Biobank Imaging). Therefore, despite the growing literature applying DNNs to neuroscience applications, if you have only a few data at hand, and you still reach a high accuracy through cross validation, remind yourself that it might be that you hacked the hell out of the hyperparameters and eventually overfitted your sample (including the test set).

Prof. Tibshirani Talk: https://www.youtube.com/watch?v=qTYDDhGUK10

Bzdok, Yeo (2017). Neuroimage, Inference in the age of big data: Future perspectives on neuroscience.

Hastie T, Tibshirani R, Friedman J (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer, New York City, USA

He, T. et al. (2018). Is deep learning better than kernel regression for functional connectivity prediction of fluid intelligence? In 2018 International Workshop on Pattern Recognition in Neuroimaging (PRNI), pp. 1â€“4, IEEE

M.I. Jordan, T.M. Mitchell (2015). Machine learning: trends, perspectives, and prospects. Science, 349, pp. 255-260

Y. LeCun, Y. Bengio, G. Hinton (2015). Deep learning. Nature, 521, pp. 436-444