Jun 16, 2017 12 minutes read

I know it’s a weird way to start a blog with a negative, but there was a wave of discussion in the last few days that I think serves as a good hook for some topics on which I’ve been thinking recently. It all started with a post in the Simply Stats blog by Jeff Leek on the caveats of using deep learning in the small sample size regime. In sum, he argues that when the sample size is small (which happens a lot in the bio domain), linear models with few parameters perform better than deep nets even with a modicum of layers and hidden units. He goes on to show that a very simple linear predictor, with top ten most informative features, performs better than a simple deep net when trying to classify zeros and ones in the MNIST dataset using only 80 or so samples. This prompted Andrew beam to write a rebuttal in which a properly trained deep net was able to beat the simple linear model, even with very few training samples. This back-and-forth comes at a time where more and more researchers in biomedical informatics are adopting deep learning for various problems. Is the hype real or are linear models really all we need? The answer, as alwyas, is that it depends. In this post, I wanted to visit use cases in machine learning where deep learning would not really make sense to use as well as tackle preconceptions that I think prevent deep learning to be used effectively, especially for newcomers.

Breaking deep learning preconceptions

First, let’s tackle some preconceptions that folks outside the field have that turn out to be half-truths. There’s two broad ones and one a bit more technical that I’m going to elaborate on. This is somewhat of an extension to Andrew Beam’s excellent “Misconceptions” section in his post.

Deep learning can really work on small sample sizes

Deep learning’s claim to fame was in a context with lots of data (remember that the first Google brain project was feeding lots of youtube videos to a deep net), and ever since it has constantly been publicized as complex algorithms running in lots of data. Unfortunately, this big data/deep learning pair somehow translated into the converse as well: the myth that it cannot be used in the small sample regime. If you have just a few samples tapping into a neural net with a high parameter-per-sample ratio superficially may seem like asking to overfit. However, just considering sample size and dimensionality for a given problem, be it supervised or unsupervised, is sort of modeling the data in a vacuum, without any context. It is probably the case that you have data sources that are related to your problem, or that there’s a strong prior that a domain expert can provide, or that the data is structured in a very particular way (e.g. is encoded in a graph or image). In all of these cases, there’s a chance deep learning can make sense as a method of choice – for example, you can encode useful representations of bigger, related datasets and use those representations in your problem. A classic illustration of this is common in natural language processing, where you cam learn word embeddings on a large corpus like Wikipedia and then use those as embeddings in a smaller, narrower corpus for a supervised task. In the extreme, you can have a set of neural nets jointly learn a representation and an effective way to reuse the representation in small sets of samples. This is called one-shot learning and has been successfully applied in a number of fields with high-dimensional data including computer vision and drug discovery.

View the Original article