Publications

The MIT/IEEE/Amazon GraphChallenge.org encourages community approaches to developing new solutions for analyzing graphs and sparse data. Sparse AI analytics present unique scalability difficulties. The proposed Sparse Deep Neural Network (DNN) Challenge draws upon prior challenges from machine learning, high performance computing, and visual analytics to create a challenge that is reflective of emerging sparse AI systems. The Sparse DNN Challenge is based on a mathematically well-defined DNN inference computation and can be implemented in any programming environment. Sparse DNN inference is amenable to both vertex-centric implementations and array-based implementations (e.g., using the GraphBLAS.org standard). The computations are simple enough that performance predictions can be made based on simple computing hardware models. The input data sets are derived from the MNIST handwritten letters. The surrounding I/O and verification provide the context for each sparse DNN inference that allows rigorous definition of both the input and the output. Furthermore, since the proposed sparse DNN challenge is scalable in both problem size and hardware, it can be used to measure and quantitatively compare a wide range of present day and future systems. Reference implementations have been implemented and their serial and parallel performance have been measured. Specifications, data, and software are publicly available at GraphChallenge.org.

Recent works in deep learning have been driven broadly by the desire to attain high accuracy on certain challenge problems. The network architecture and other hyperparameters of many published models are typically chosen by trial-and-error experiments with little considerations paid to resource constraints at deployment time. We propose a fully automated model learning approach that (1) treats architecture selection as part of the learning process, (2) uses a blend of broad-based random sampling and adaptive iterative refinement to explore the solution space, (3) performs optimization subject to given memory and computational constraints imposed by target deployment scenarios, and (4) is scalable and can use only a practically small number of GPUs for training. We present results that show graceful model degradation under strict resource constraints for object classification problems using CIFAR-10 in our experiments. We also discuss future work in further extending the approach.