Jack Sklar

Adam Wunderlich, Jack Sklar

Random noise arising from physical processes is an inherent characteristic of measurements and a limiting factor for most signal processing and data analysis tasks. Given the recent interest in generative adversarial networks (GANs) for data-driven modeling, it is important to determine to what extent GANs can faithfully reproduce noise in target data sets. In this paper, we present an empirical investigation that aims to shed light on this issue for time series. Namely, we assess two general-purpose GANs for time series that are based on the popular deep convolutional GAN (DCGAN) architecture, a direct time-series model and an image-based model that uses a short-time Fourier transform (STFT) data representation. The GAN models are trained and quantitatively evaluated using distributions of simulated noise time series with known ground-truth parameters. Target time series distributions include a broad range of noise types commonly encountered in physical measurements, electronics, and communication systems: band-limited thermal noise, power law noise, shot noise, and impulsive noise. We find that GANs are capable of learning many noise types, although they predictably struggle when the GAN architecture is not well suited to some aspects of the noise, e.g., impulsive time-series with extreme outliers. Our findings provide insights into the capabilities and potential limitations of current approaches to time-series GANs and highlight areas for further research. In addition, our battery of tests provides a useful benchmark to aid the development of deep generative models for time series.

Erik Partridge, Jack Sklar, Omar El-lakany

Often, more time is spent on finding a model that works well, rather than tuning the model and working directly with the dataset. Our research began as an attempt to improve upon a simple Recurrent Neural Network for answering "simple" first-order questions (QA-RNN), developed by Ferhan Ture and Oliver Jojic, from Comcast Labs, using the SimpleQuestions dataset. Their baseline model, a bidirectional, 2-layer LSTM RNN and a GRU RNN, have accuracies of 0.94 and 0.90, for entity detection and relation prediction, respectively. We fine tuned these models by doing substantial hyper-parameter tuning, getting resulting accuracies of 0.70 and 0.80, for entity detection and relation prediction, respectively. An accuracy of 0.984 was obtained on entity detection using a 1-layer LSTM, where preprocessing was done by removing all words not part of a noun chunk from the question. 100% of the dataset was available for relation prediction, but only 20% of the dataset, was available for entity detection, which we believe to be much of the reason for our initial difficulties in replicating their result, despite the fact we were able to improve on their entity detection results.