Deep Learning Algorithms: Theory and Practice

We begin this talk by discussing Machine Learning Networks (that typically have one hidden layer) and their shortcomings by providing a use-case called “Voice of Patients”. We then define and discuss Deep Learning Networks, which usually contain several layers of nonlinear processing units (rather than one) and where each layer uses the output from the previous layer as the input. Just like classical single-layer machine learning algorithms, these deep learning algorithms can also be either supervised or unsupervised in nature. Next, we focus on a specific kind of deep learning networks called recurrent neural networks and why such networks tend to yield good empirical results for use-cases that depend on time-series or are longitudinal in nature. We end this talk by discussing the promise that deep learning networks provide empirically versus the potential lack of any “theoretical guarantees”. Finally, rather than going into the math behind such networks and/or algorithms, this talk will mainly focus on advantages and disadvantages of using them for real-live examples.