Twitter uses machine learning to crop images, where crops are centered around the part predicted to be the most salient. In fall 2020, Twitter users raised concerns that the automated image cropping system on Twitter favored light-skinned over dark-skinned individuals, as well as concerns that the system favored cropping woman's bodies instead of their heads. In order to address these concerns, we conduct an extensive analysis using formalized group fairness metrics. We find systematic disparities in cropping and identify contributing factors, including the fact that the cropping based on the single most salient point can amplify the disparities because of an effect we term argmax bias. However, we demonstrate that formalized fairness metrics and quantitative analysis on their own are insufficient for capturing the risk of representational harm in automatic cropping. We suggest the removal of saliency-based cropping in favor of a solution that better preserves user agency. For developing a new solution that sufficiently address concerns related to representational harm, our critique motivates a combination of quantitative and qualitative methods that include human-centered design.
相關內容
Fair ranking problems arise in many decision-making processes that often necessitate a trade-off between accuracy and fairness. Many existing studies have proposed correction methods such as adding fairness constraints to a ranking model's loss. However, the challenge of correcting the data bias for fair ranking remains, and the trade-off of the ranking models leaves room for improvement. In this paper, we propose a fair ranking framework that evaluates the order of training data in a pairwise manner as well as various fairness measurements in ranking. This study is the first proposal of a pre-processing method that solves fair ranking problems using the pairwise ordering method with our best knowledge. The fair pairwise ordering method is prominent in training the fair ranking models because it ensures that the resulting ranking likely becomes parity across groups. As far as the fairness measurements in ranking are represented as a linear constraint of the ranking models, we proved that the minimization of loss function subject to the constraints is reduced to the closed solution of the minimization problem augmented by weights to training data. This closed solution inspires us to present a practical and stable algorithm that iterates the optimization of weights and model parameters. The empirical results over real-world datasets demonstrated that our method outperforms the existing methods in the trade-off between accuracy and fairness over real-world datasets and various fairness measurements.
On the Importance and Shortcomings of Code Readability Metrics: A Case Study on Reactive Programming
Well structured and readable source code is a pre-requisite for maintainable software and successful collaboration among developers. Static analysis enables the automated extraction of code complexity and readability metrics which can be leveraged to highlight potential improvements in code to both attain software of high quality and reinforce good practices for developers as an educational tool. This assumes reliable readability metrics which are not trivial to obtain since code readability is somewhat subjective. Recent research has resulted in increasingly sophisticated models for predicting readability as perceived by humans primarily with a procedural and object oriented focus, while functional and declarative languages and language extensions advance as they often are said to lead to more concise and readable code. In this paper, we investigate whether the existing complexity and readability metrics reflect that wisdom or whether the notion of readability and its constituents requires overhaul in the light of programming language changes. We therefore compare traditional object oriented and reactive programming in terms of code complexity and readability in a case study. Reactive programming is claimed to increase code quality but few studies have substantiated these claims empirically. We refactored an object oriented open source project into a reactive candidate and compare readability with the original using cyclomatic complexity and two state-of-the-art readability metrics. More elaborate investigations are required, but our findings suggest that both cyclomatic complexity and readability decrease significantly at the same time in the reactive candidate, which seems counter-intuitive. We exemplify and substantiate why readability metrics may require adjustment to better suit popular programming styles other than imperative and object-oriented to better match human expectations.
Understanding the operation of biological and artificial networks remains a difficult and important challenge. To identify general principles, researchers are increasingly interested in surveying large collections of networks that are trained on, or biologically adapted to, similar tasks. A standardized set of analysis tools is now needed to identify how network-level covariates -- such as architecture, anatomical brain region, and model organism -- impact neural representations (hidden layer activations). Here, we provide a rigorous foundation for these analyses by defining a broad family of metric spaces that quantify representational dissimilarity. Using this framework we modify existing representational similarity measures based on canonical correlation analysis to satisfy the triangle inequality, formulate a novel metric that respects the inductive biases in convolutional layers, and identify approximate Euclidean embeddings that enable network representations to be incorporated into essentially any off-the-shelf machine learning method. We demonstrate these methods on large-scale datasets from biology (Allen Institute Brain Observatory) and deep learning (NAS-Bench-101). In doing so, we identify relationships between neural representations that are interpretable in terms of anatomical features and model performance.
(Im)balance in the representation of news has always been a topic of debate in political circles. The concept of balance has often been discussed and studied in the context of the social responsibility theory and the prestige press in the USA. While various qualitative, as well as quantitative measures of balance, have been suggested in the literature, a comprehensive analysis of all these measures across a large dataset of the post-truth era comprising different popular news media houses and over a sufficiently long temporal scale in a non-US democratic setting is lacking. We use this concept of balance to measure and understand the evolution of imbalance in Indian media on various journalistic metrics on a month-by-month basis. For this study, we amass a huge dataset of over four million political articles from India for 9+ years and analyze the extent and quality of coverage given to issues and political parties in the context of contemporary influential events for three leading newspapers. We use several state-of-the-art NLP tools to effectively understand political polarization (if any) manifesting in these articles over time. We find that two out of the three news outlets are more strongly clustered in their imbalance metrics. We also observe that only a few locations are extensively covered across all the news outlets and the situation is only slightly getting better for one of the three news outlets. Cloze tests show that the changing landscape of events get reflected in all the news outlets with border and terrorism issues dominating in around 2010 while economic aspects like unemployment, GST, demonetization, etc. became more dominant in the period 2014 -- 2018. Further, cloze tests clearly portray the changing popularity profile of the political parties over time.
With the availability of granular geographical data, social scientists are increasingly interested in examining how residential neighborhoods are formed and how they influence attitudes and behavior. To facilitate such studies, we develop an easy-to-use online survey instrument that allows respondents to draw their neighborhoods on a map. We then propose a statistical model to analyze how the characteristics of respondents, relevant local areas, and their interactions shape subjective neighborhoods. The model also generates out-of-sample predictions of one's neighborhood given these observed characteristics. We illustrate the proposed methodology by conducting a survey among registered voters in Miami, New York City, and Phoenix. We find that across these cities voters are more likely to include same-race and co-partisan census blocks in their neighborhoods. Net of other factors, White respondents are 6.1 to 16.9 percentage points more likely to include in their neighborhoods a census block composed entirely of White residents compared to one with no White residents. Similarly, Democratic and Republican respondents are 8.6 to 19.2 percentage points more likely to include an entirely co-partisan census block compared to one consisting entirely of out-partisans. Co-partisanship exhibits a similar, independent, influence. We also show that our model provides more accurate out-of-sample predictions than the standard distance-based measures of neighborhoods. Open-source software is available for implementing the proposed methodology.
The problem of Approximate Nearest Neighbor (ANN) search is fundamental in computer science and has benefited from significant progress in the past couple of decades. However, most work has been devoted to pointsets whereas complex shapes have not been sufficiently treated. Here, we focus on distance functions between discretized curves in Euclidean space: they appear in a wide range of applications, from road segments to time-series in general dimension. For $\ell_p$-products of Euclidean metrics, for any $p$, we design simple and efficient data structures for ANN, based on randomized projections, which are of independent interest. They serve to solve proximity problems under a notion of distance between discretized curves, which generalizes both discrete Fr\'echet and Dynamic Time Warping distances. These are the most popular and practical approaches to comparing such curves. We offer the first data structures and query algorithms for ANN with arbitrarily good approximation factor, at the expense of increasing space usage and preprocessing time over existing methods. Query time complexity is comparable or significantly improved by our algorithms, our algorithm is especially efficient when the length of the curves is bounded.
Transformer architectures show significant promise for natural language processing. Given that a single pretrained model can be fine-tuned to perform well on many different tasks, these networks appear to extract generally useful linguistic features. A natural question is how such networks represent this information internally. This paper describes qualitative and quantitative investigations of one particularly effective model, BERT. At a high level, linguistic features seem to be represented in separate semantic and syntactic subspaces. We find evidence of a fine-grained geometric representation of word senses. We also present empirical descriptions of syntactic representations in both attention matrices and individual word embeddings, as well as a mathematical argument to explain the geometry of these representations.
We present an open-source tool for visualizing multi-head self-attention in Transformer-based language representation models. The tool extends earlier work by visualizing attention at three levels of granularity: the attention-head level, the model level, and the neuron level. We describe how each of these views can help to interpret the model, and we demonstrate the tool on the BERT model and the OpenAI GPT-2 model. We also present three use cases for analyzing GPT-2: detecting model bias, identifying recurring patterns, and linking neurons to model behavior.
Recently, the state-of-the-art models for image captioning have overtaken human performance based on the most popular metrics, such as BLEU, METEOR, ROUGE, and CIDEr. Does this mean we have solved the task of image captioning? The above metrics only measure the similarity of the generated caption to the human annotations, which reflects its accuracy. However, an image contains many concepts and multiple levels of detail, and thus there is a variety of captions that express different concepts and details that might be interesting for different humans. Therefore only evaluating accuracy is not sufficient for measuring the performance of captioning models --- the diversity of the generated captions should also be considered. In this paper, we proposed a new metric for measuring the diversity of image captions, which is derived from latent semantic analysis and kernelized to use CIDEr similarity. We conduct extensive experiments to re-evaluate recent captioning models in the context of both diversity and accuracy. We find that there is still a large gap between the model and human performance in terms of both accuracy and diversity and the models that have optimized accuracy (CIDEr) have low diversity. We also show that balancing the cross-entropy loss and CIDEr reward in reinforcement learning during training can effectively control the tradeoff between diversity and accuracy of the generated captions.
This paper discusses and demonstrates the outcomes from our experimentation on Image Captioning. Image captioning is a much more involved task than image recognition or classification, because of the additional challenge of recognizing the interdependence between the objects/concepts in the image and the creation of a succinct sentential narration. Experiments on several labeled datasets show the accuracy of the model and the fluency of the language it learns solely from image descriptions. As a toy application, we apply image captioning to create video captions, and we advance a few hypotheses on the challenges we encountered.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191