Search for: All records

Research in Responsible AI has developed a range of principles and practices to ensure that machine learning systems are used in a manner that is ethical and aligned with human values. However, a critical yet often neglected aspect of ethical ML is the ethical implications that appear when designing evaluations of ML systems. For instance, teams may have to balance a trade-off between highly informative tests to ensure downstream product safety, with potential fairness harms inherent to the implemented testing procedures. We conceptualize ethics-related concerns in standard ML evaluation techniques. Specifically, we present a utility framework, characterizing the key trade-off in ethical evaluation as balancing information gain against potential ethical harms. The framework is then a tool for characterizing challenges teams face, and systematically disentangling competing considerations that teams seek to balance. Differentiating between different types of issues encountered in evaluation allows us to highlight best practices from analogous domains, such as clinical trials and automotive crash testing, which navigate these issues in ways that can offer inspiration to improve evaluation processes in ML. Our analysis underscores the critical need for development teams to deliberately assess and manage ethical complexities that arise during the evaluation of ML systems, and for the industry to move towards designing institutional policies to support ethical evaluations. 

The Colorado River supplies >40 million people in the United States Southwest with their daily water supply and is unable to meet the current demands. New approaches are needed to enhance sustainability and resilience. A net zero urban water (NZUW) approach meets the needs of a given community with a locally available and sustainable water supply, without detriment to interconnected systems and the long-term water supply. Transitioning to a NZUW future will require considerable modifications to governance and policy across the Southwest and its cities specifically. We identify five areas of governance and policy challenges: diversified water sources and sinks; planning, design, and operation; monitoring and enforcement; coordination; and addressing equity and justice. Four case study cities are investigated: Albuquerque, Denver, Los Angeles, and Tucson. Across these cities, the policy priorities include supporting potable water reuse, coordinating policies across jurisdictions for alternative water sources, addressing equity and justice, developing and incentivizing water conservation plans, and making aquifer storage and recovery projects easier and more economical to pursue. We conclude that a NZUW transition in the Southwest faces considerable governance and policy challenges, but moving cities toward this goal is crucial. 

Abstract We have examined the roles of yeast mRNA decapping-activators Pat1 and Dhh1 in repressing the translation and abundance of specific mRNAs in nutrient-replete cells using ribosome profiling, RNA-Seq, CAGE analysis of capped mRNAs, RNA Polymerase II ChIP-Seq, and TMT-mass spectrometry of mutants lacking one or both factors. Although the Environmental Stress Response (ESR) is activated in dhh1Δ and pat1Δ mutants, hundreds of non-ESR transcripts are elevated in a manner indicating cumulative repression by Pat1 and Dhh1 in wild-type cells. These mRNAs show both reduced decapping and diminished transcription in the mutants, indicating that impaired mRNA turnover drives transcript derepression in cells lacking Dhh1 or Pat1. mRNA degradation stimulated by Dhh1/Pat1 is not dictated by poor translation nor enrichment for suboptimal codons. Pat1 and Dhh1 also collaborate to reduce translation and protein production from many mRNAs. Transcripts showing concerted translational repression by Pat1/Dhh1 include mRNAs involved in cell adhesion or utilization of the poor nitrogen source allantoin. Pat1/Dhh1 also repress numerous transcripts involved in respiration, catabolism of non-preferred carbon or nitrogen sources, or autophagy; and we obtained evidence for elevated respiration and autophagy in the mutants. Thus, Pat1 and Dhh1 function as post-transcriptional repressors of multiple pathways normally activated only during nutrient limitation. 

Degradation of most yeast mRNAs involves decapping by Dcp1/Dcp2. DEAD-box protein Dhh1 has been implicated as an activator of decapping, in coupling codon non-optimality to enhanced degradation, and as a translational repressor, but its functions in cells are incompletely understood. RNA-Seq analyses coupled with CAGE sequencing of all capped mRNAs revealed increased abundance of hundreds of mRNAs indcp2Δ cells that appears to result directly from impaired decapping rather than elevated transcription. Interestingly, only a subset of mRNAs requires Dhh1 for targeting by Dcp2, and also generally requires the other decapping activators Pat1, Edc3, or Scd6; whereas most of the remaining transcripts utilize nonsense-mediated mRNA decay factors for Dcp2-mediated turnover. Neither inefficient translation initiation nor stalled elongation appears to be a major driver of Dhh1-enhanced mRNA degradation. Surprisingly, ribosome profiling revealed thatdcp2Δ confers widespread changes in relative translational efficiencies (TEs) that generally favor well-translated mRNAs. Because ribosome biogenesis is reduced while capped mRNA abundance is increased bydcp2Δ,we propose that an increased ratio of mRNA to ribosomes increases competition among mRNAs for limiting ribosomes to favor efficiently translated mRNAs indcp2Δ cells. Interestingly, genes involved in respiration or utilization of alternative carbon or nitrogen sources are upregulated, and both mitochondrial function and cell filamentation are elevated indcp2Δ cells, suggesting that decapping sculpts gene expression post-transcriptionally to fine-tune metabolic pathways and morphological transitions according to nutrient availability. 

We consider networks, trained via stochastic gradient descent to minimize L2 loss, with the training labels perturbed by independent noise at each iteration. We characterize the behavior of the training dynamics near any parameter vector that achieves zero training error, in terms of an implicit regularization term corresponding to the sum over the datapoints, of the squared L2 of the gradient of the model with respect to the parameter vector, evaluated at each data point. This holds for networks of any connectivity, width,depth, and choice of activation function. We interpret this implicit regularization term for three simple settings: matrix sensing, two layer ReLU networks trained on one-dimensional data, and two layer networks with sigmoid activations trained on a single datapoint. For these settings, we show why this new and general implicit regularization effect drives the networks towards "simple" models. 

Abstract Motivated by the results of Scott and Patel about “untangling” closed geodesics in finite covers of hyperbolic surfaces, we introduce and study primitivity, simplicity and non-filling index functions for finitely generated free groups. We obtain lower bounds for these functions and relate these free group results back to the setting of hyperbolic surfaces. An appendix by Khalid Bou–Rabee connects the primitivity index function f prim ( n , F N ) to the residual finiteness growth function for F N . 

Abstract High nighttime urban air temperatures increase health risks and economic vulnerability of people globally. While recent studies have highlighted nighttime heat mitigation effects of urban vegetation, the magnitude and variability of vegetation-derived urban nighttime cooling differs greatly among cities. We hypothesize that urban vegetation-derived nighttime air cooling is driven by vegetation density whose effect is regulated by aridity through increasing transpiration. We test this hypothesis by deploying microclimate sensors across eight United States cities and investigating relationships of nighttime air temperature and urban vegetation throughout a summer season. Urban vegetation decreased nighttime air temperature in all cities. Vegetation cooling magnitudes increased as a function of aridity, resulting in the lowest cooling magnitude of 1.4 °C in the most humid city, Miami, FL, and 5.6 °C in the most arid city, Las Vegas, NV. Consistent with the differences among cities, the cooling effect increased during heat waves in all cities. For cities that experience a summer monsoon, Phoenix and Tucson, AZ, the cooling magnitude was larger during the more arid pre-monsoon season than during the more humid monsoon period. Our results place the large differences among previous measurements of vegetation nighttime urban cooling into a coherent physiological framework dependent on plant transpiration. This work informs urban heat risk planning by providing a framework for using urban vegetation as an environmental justice tool and can help identify where and when urban vegetation has the largest effect on mitigating nighttime temperatures.