We present a simple case study, demonstrating that Variational Information Bottleneck (VIB) can improve a network's classification calibration as well as its ability to detect out-of-distribution data. Without explicitly being designed to do so, VIB gives two natural metrics for handling and quantifying uncertainty.

Predicting potential credit default accounts in advance is challenging. Traditional statistical techniques typically cannot handle large amounts of data and the dynamic nature of fraud and humans. To tackle this problem, recent research has focused on artificial and computational intelligence based approaches. In this work, we present and validate a heuristic approach to mine potential default accounts in advance where a risk probability is precomputed from all previous data and the risk probability for recent transactions are computed as soon they happen. Beside our heuristic approach, we also apply a recently proposed machine learning approach that has not been applied previously on our targeted dataset [15]. As a result, we find that these applied approaches outperform existing state-of-the-art approaches.

Deep neural networks (DNN) are increasingly being accelerated on application-specific hardware such as the Google TPU designed especially for deep learning. Timing speculation is a promising approach to further increase the energy efficiency of DNN accelerators. Architectural exploration for timing speculation requires detailed gate-level timing simulations that can be time-consuming for large DNNs that execute millions of multiply-and-accumulate (MAC) operations. In this paper we propose FATE, a new methodology for fast and accurate timing simulations of DNN accelerators like the Google TPU. FATE proposes two novel ideas: (i) DelayNet, a DNN based timing model for MAC units; and (ii) a statistical sampling methodology that reduces the number of MAC operations for which timing simulations are performed. We show that FATE results in between 8 times-58 times speed-up in timing simulations, while introducing less than 2% error in classification accuracy estimates. We demonstrate the use of FATE by comparing to conventional DNN accelerator that uses 2's complement (2C) arithmetic with an alternative implementation that uses signed magnitude representations (SMR). We show that that the SMR implementation provides 18% more energy savings for the same classification accuracy than 2C, a result that might be of independent interest.

The skill or prediction error of a model must be estimated, and as an estimate, it will contain error. This is made clear by distinguishing between the true error of a model and the estimated or sample error. One is the error rate of the hypothesis over the sample of data that is available. The other is the error rate of the hypothesis over the entire unknown distribution D of examples.

This blog offers a deeper dive into the machine learning training process for performing automatic target detection. Samples of automatic target detection were recently presented at the Machine Learning: Automate Remote Sensing Analytics to Gain a Competitive Advantage webinar. Machine learning (ML) applications, from object recognition and caption generation, to automatic language translation and driverless cars, have increased dramatically over the last few years, powered mainly by the increase of computing power (using GPUs), reduced cost of storage, wider availability of training data, and development of new training techniques for the machine learning models. In the last five years, Harris Corporation has made a multi-million dollar investment into applying machine learning to solve customer challenges using remote sensing data. In response to the increased interest from our customers in evaluating how machine learning can solve their problems using geospatial data, I set out to train some of my coworkers on how to build a ML model to perform automatic feature detection on 2D overhead imagery.

ML 2.0: In this paper, we propose a paradigm shift from the current practice of creating machine learning models - which requires months-long discovery, exploration and "feasibility report" generation, followed by re-engineering for deployment - in favor of a rapid, 8-week process of development, understanding, validation and deployment that can executed by developers or subject matter experts (non-ML experts) using reusable APIs. This accomplishes what we call a "minimum viable data-driven model," delivering a ready-to-use machine learning model for problems that haven't been solved before using machine learning. We provide provisions for the refinement and adaptation of the "model," with strict enforcement and adherence to both the scaffolding/abstractions and the process. We imagine that this will bring forth the second phase in machine learning, in which discovery is subsumed by more targeted goals of delivery and impact.

Recidivism prediction scores are used across the USA to determine sentencing and supervision for hundreds of thousands of inmates. One such generator of recidivism prediction scores is Northpointe's Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) score, used in states like California and Florida, which past research has shown to be biased against black inmates according to certain measures of fairness. To counteract this racial bias, we present an adversarially-trained neural network that predicts recidivism and is trained to remove racial bias. When comparing the results of our model to COMPAS, we gain predictive accuracy and get closer to achieving two out of three measures of fairness: parity and equality of odds. Our model can be generalized to any prediction and demographic. This piece of research contributes an example of scientific replication and simplification in a high-stakes real-world application like recidivism prediction.

In this work, we characterize the doctor-patient relationship using a machine learning-derived trust score. We show that this score has statistically significant racial associations, and that by modeling trust directly we find stronger disparities in care than by stratifying on race. We further demonstrate that mistrust is indicative of worse outcomes, but is only weakly associated with physiologically-created severity scores. Finally, we describe sentiment analysis experiments indicating patients with higher levels of mistrust have worse experiences and interactions with their caregivers. This work is a step towards measuring fairer machine learning in the healthcare domain.

Classifiers can be trained with data-dependent constraints to satisfy fairness goals, reduce churn, achieve a targeted false positive rate, or other policy goals. We study the generalization performance for such constrained optimization problems, in terms of how well the constraints are satisfied at evaluation time, given that they are satisfied at training time. To improve generalization performance, we frame the problem as a two-player game where one player optimizes the model parameters on a training dataset, and the other player enforces the constraints on an independent validation dataset. We build on recent work in two-player constrained optimization to show that if one uses this two-dataset approach, then constraint generalization can be significantly improved. As we illustrate experimentally, this approach works not only in theory, but also in practice.

One of the most pressing concerns that keeps retail professionals up at night is how to combat fraud. Retailers could lose upwards of $71 billion from fraudulent online transactions over the next few years, yet some executives feel that publicly acknowledging a fraud issue would harm their brand. One of the most significant fraud concerns merchants face today are false positives -- i.e., transactions attempted by legitimate customers that are tagged as suspicious by fraud prevention systems, ultimately leaving money on the table. Because their effect is so difficult to accurately measure, false positives are often ignored, and their cost greatly underestimated. However, a majority of retailers say that fraudulent transactions that aren't detected cost more than a legitimate transaction that's inaccurately declined, despite some evidence that the opposite is true. What's more, relatively few companies track false positives.