From Social Data to Neural Networks: Robust Decision-Making in Machine Learning

Abstract

In this thesis we develop algorithms to control and understand the behavior of a system despite erroneous, noisy or otherwise problematic input. In our first line of work, we focus on the decision-making for social data pipeline that usually consists of three steps: (1) collect and label relevant data, (2) learn a surrogate model representing the real world as accurately as possible, and (3) perform optimization/decision-making on the surrogate model. We begin by tackling the first problem, showing how to aggregate labels and reduce uncertainty in crowdsourcing platforms. The data collected from crowdsourcing is often noisy and hence may yield models that are biased or erroneous. We present a graphical model to aggregate the different labelers' decisions and estimate their uncertainty. This allows us to utilize reviewers more efficiently and improve the quality of our data. Subsequently, we focus on the next two steps of the pipeline and use information dissemination in social networks as a case study, ultimately aiming to design robust algorithms for that problem. We propose a new hyperparametric model that utilizes information about the nodes of the social graph and allows us to approximately maximize the number of nodes that are will be influenced under the worst possible choice of the hyperparameter (on expectation). This accounts for errors introduced during learning the model from data of past cascades, i.e. observations of how does diffusion propagate among the nodes of the network. In the second part of the thesis, we focus on neural networks that have been in the frontline of the machine learning revolution we have been observing over the last few years. We investigate the connections between robustness and interpretability for genomic datasets and finally, try to shed some light upon the generalization behavior of neural networks by studying the implicit bias of Stochastic Gradient Descent (SGD), i.e. the underlying reason why we are able to recover generalizable solutions in modern overparametrized networks, by claiming that neural networks trained with SGD learn functions of increasing complexity as training progresses. That is, they learn meaningful correlations with the data and retain them throughout the training process, even after overfitting to the noise in the dataset.