2019 Rice Data Science Conference

​WATCH VIDEO​​​

Two common techniques for image segmentation - level set methods and convolutional neural networks (CNN) - rely on alternating convolutions with nonlinearities to describe image features: neural networks with mean-zero convolution kernels can be treated as upwind finite difference discretizations of differential equations. Such a comparison provides a well-established framework for proving properties of CNNs, such as stability and approximation accuracy. We test this relationship by constructing a level set network, a CNN where forward-propagation is equivalent to solving the level set equation. The level set network achieves comparable segmentation accuracy to solving the level set equation, while not obtaining the accuracy of a CNN. We therefore analyze which convolution filters are present in our CNN, to see whether finite difference stencils are learned during training. We observe certain patterns form in the decoding layers of the network, where kernels cannot be accounted for by finite difference stencils alone.

​WATCH VIDEO​​​

We propose a simple yet powerful framework for modeling integer-valued data, such as counts, scores, and rounded data. The integer-valued data are modeled by Simultaneously Transforming And Rounding (STAR) a continuous-valued process, where the transformation may be known or learned from the data. STAR produces a flexible class of integer-valued processes, which can account for zero-inflation, bounded or censored data, and over- or underdispersion. Scalable computation is available via an efficient MCMC algorithm, which provides a mechanism for direct adaptation of successful Bayesian methods for continuous data to the integer-valued data setting. Using the STAR framework, we develop new additive models and Bayesian Additive Regression Trees (BART) for integer-valued data. The predictive and inferential capabilities of STAR are illustrated using a medical utilization dataset and an animal abundance dataset, with exceptional predictive and computational performance.

​WATCH VIDEO​​​

Convolutional neural networks (CNNs) have been increasingly deployed to edge devices. Hence, many efforts have been made towards efficient CNN inference on resource-constrained platforms. This paper attempts to explore an orthogonal direction: how to conduct more energy-efficient training of CNNs? We strive to reduce the energy cost during training from three complementary levels: stochastic mini-batch dropping on the data level; selective layer update on the model level; and sign prediction for low-cost, low-precision back-propagation, on the algorithm level. Extensive simulations and ablation studies, with real energy measurements from an FPGA board, confirm the superiority of our proposed strategies and demonstrate remarkable energy savings for training. For example, when training ResNet-110 on CIFAR-100, an over 84% training energy saving comes at the small accuracy costs of 2% (top-1) and 0.1% (top-5).

​WATCH VIDEO​​​

Deep learning (particularly deep convolutional neural networks) have revolutionized the field of data-driven modeling techniques. Through an increase in data coverage, computational horsepower, and model complexity we have regularly seen records set and broken in computer vision.

One fundamental limitation of deep convolutional networks is the data they apply to. A “convolution” is in essence a dense matrix-matrix multiply. It operates on dense, regularly structured data (e.g. a digital image with RGB values for every pixel in a regular grid). 

This presentation offers an overview of graph convolutional networks (GCNs) – an extension to classical convolutional networks for irregular data. It includes an overview of GCNs and a walkthrough of two example applications of GCNs (one unsupervised and one semi-supervised) on a real world streaming dataset from the GDELT Project.

​WATCH VIDEO​​​

Recently, neural networks have been generalized to process data on graphs, with cutting-edge results in traditional tasks such as node classification and link prediction. These methods have all been formulated in a way suited only to data on the nodes of a graph, based on spectral graph theory. Using tools from algebraic topology, it is possible to reason about oriented data on higher-order structures by relying on the so-called Hodge Laplacian. Our goal is to develop techniques for applying the Hodge Laplacian to process data on higher-order graph structures using graph neural networks. To illustrate the practical value of this framework, we tackle the problem of flow interpolation: Given observations of flow over a subset of the edges of a graph, how can flow over the unobserved edges be inferred? We propose an architecture based on recurrent neural networks for performing flow interpolation, and demonstrate it on urban traffic data.

​WATCH VIDEO​​​

Mimicking the behavior of an expert player in a Reinforcement Learning (RL) Environment to enhance the training of a novice agent from scratch is called Imitation Learning (IL). In most RL environments, the state sequences an agent encounters is a Markov Decision Process. This makes mimicking difficult as it is unlikely that a new agent may encounter similar state sequences as an expert. Prior research in IL proposes to learn a mapping between expert's states and actions, needing considerable number of state-action pairs to achieve good results. We propose an alternative to IL by appending the novice's action space with frequent action sequences of the expert. This modification improves the exploration and significantly outperforms alternatives like Dataset-Aggregation. We experiment with popular Atari games and show significant and consistent growth in the score that the new agents achieve using just a few expert action sequences.