loader


xAUC metric
Honesty is a serious concern where machine-learned forecast risk scores inform high-level results such as sentencing to bail and criminal justice. Recent work has categorized the different impact that such risk scores can have when used for binary classification work and has provided tools to audit and correct the resulting classifiers. However, this does not take into account the very diverse downstream applications of risk scores and their non-binary nature. To better account for this, in this study, we examine the rationale for forecast risk scores from the perspective of a two-party ranking mission, where one seeks to evaluate positive examples rather than negative examples. We introduce xAUC inequality as a metric for estimating the differential impact of risk scores, and define it as the difference in probabilities for ranking a random positive example from a protected group. We present the decomposition of bilateral ranking loss as a component of the inconsistency and components that embody the pure predictive potential within each group. We further provide an explanation of the xAUC conflict in terms of resource allocation reasoning and make links to existing reasonable measurements and changes. We empirically evaluate xAUC in datasets on recidivism forecasting, income forecasting, and cardiovascular forecasting, where it describes inequalities that are not clear from simply comparing group forecasting performance. …

Multi-Objective Neural Architecture Search (Monas) Google


Recent studies on neural structure search have shown that automated neural networks work just as well as man-made structures. Most of the existing works on neurological structural search are aimed at finding structures that are optimal for prognostic accuracy. These methods can create complex structures that consume too much energy, which is not suitable for the computer environment with limited power budgets. We propose Monas, a multi-objective neural architecture quest with novel reward functions that takes into account both predictive accuracy and power consumption when examining neurological structures. Monas design effectively explores space and searches for structures that meet a given needs. Experimental results demonstrate that the structures discovered by Monas achieve comparable or better accuracy than sophisticated models, while having better energy performance. …

Wasserstein Percentre Google
The Wasserstein Percenter is a single distribution that summarizes the set of input operations while respecting their geometry. …

Wasserstein’s different assumption Google
This article introduces Wasserstein’s different hypothesis, which is a new form of approximate Bayesian hypothesis based on the optimal transport theory. Wasserstein’s differential assumption uses a new family variation, which includes both f-divergence and Wasserstein distance as special events. The gradient of Wasserstein contrast loss is obtained by retraction through synchronous repetitions. This technique results in a more stable feasibility training method that can be used with implicit distributions and probability schemes. Using Wasserstein’s different hypothetical framework, we introduce many new forms of autoencoding and test their strength and effectiveness against existing different autoencoding techniques. …



Source link

Leave a Reply