Showing papers by "Nello Cristianini published in 2018"
TL;DR: It is argued that the nature of the feedback commonly used by learning agents to update their models and subsequent decisions could steer the behaviour of human users away from what benefits them, and in a direction that can undermine autonomy and cause further disparity between actions and goals as exemplified by addictive and compulsive behaviour.
Abstract: Interactions between an intelligent software agent (ISA) and a human user are ubiquitous in everyday situations such as access to information, entertainment, and purchases In such interactions, the ISA mediates the user's access to the content, or controls some other aspect of the user experience, and is not designed to be neutral about outcomes of user choices Like human users, ISAs are driven by goals, make autonomous decisions, and can learn from experience Using ideas from bounded rationality (and deploying concepts from artificial intelligence, behavioural economics, control theory, and game theory), we frame these interactions as instances of an ISA whose reward depends on actions performed by the user Such agents benefit by steering the user's behaviour towards outcomes that maximise the ISA's utility, which may or may not be aligned with that of the user Video games, news recommendation aggregation engines, and fitness trackers can all be instances of this general case Our analysis facilitates distinguishing various subcases of interaction (ie deception, coercion, trading, and nudging), as well as second-order effects that might include the possibility for adaptive interfaces to induce behavioural addiction, and/or change in user belief We present these types of interaction within a conceptual framework, and review current examples of persuasive technologies and the issues that arise from their use We argue that the nature of the feedback commonly used by learning agents to update their models and subsequent decisions could steer the behaviour of human users away from what benefits them, and in a direction that can undermine autonomy and cause further disparity between actions and goals as exemplified by addictive and compulsive behaviour We discuss some of the ethical, social and legal implications of this technology and argue that it can sometimes exploit and reinforce weaknesses in human beings
83 citations
24 Oct 2018
TL;DR: This work applies a method developed in the context of domain adaptation to address the problem of "being right for the right reason", where a classifier is asked to make a decision in a way that is entirely 'agnostic' to a given protected concept.
Abstract: We consider the problem of a neural network being requested to classify images (or other inputs) without making implicit use of a “protected concept”, that is a concept that should not play any role in the decision of the network. Typically these concepts include information such as gender or race, or other contextual information such as image backgrounds that might be implicitly reflected in unknown correlations with other variables, making it insufficient to simply remove them from the input features. In other words, making accurate predictions is not good enough if those predictions rely on information that should not be used: predictive performance is not the only important metric for learning systems. We apply a method developed in the context of domain adaptation to address this problem of “being right for the right reason”, where we request a classifier to make a decision in a way that is entirely ‘agnostic’ to a given protected concept (e.g. gender, race, background etc.), even if this could be implicitly reflected in other attributes via unknown correlations. After defining the concept of an ‘agnostic model’, we demonstrate how the Domain-Adversarial Neural Network can remove unwanted information from a model using a gradient reversal layer.
29 citations
TL;DR: Overall, the authors see strong evidence that their language changes dramatically between night and day, reflecting changes in their concerns and underlying cognitive and emotional processes, which occur at times associated with major changes in neural activity and hormonal levels.
Abstract: The psychological state of a person is characterised by cognitive and emotional variables which can be inferred by psychometric methods. Using the word lists from the Linguistic Inquiry and Word Count, designed to infer a range of psychological states from the word usage of a person, we studied temporal changes in the average expression of psychological traits in the general population. We sampled the contents of Twitter in the United Kingdom at hourly intervals for a period of four years, revealing a strong diurnal rhythm in most of the psychometric variables, and finding that two independent factors can explain 85% of the variance across their 24-h profiles. The first has peak expression time starting at 5am/6am, it correlates with measures of analytical thinking, with the language of drive (e.g power, and achievement), and personal concerns. It is anticorrelated with the language of negative affect and social concerns. The second factor has peak expression time starting at 3am/4am, it correlates with the language of existential concerns, and anticorrelates with expression of positive emotions. Overall, we see strong evidence that our language changes dramatically between night and day, reflecting changes in our concerns and underlying cognitive and emotional processes. These shifts occur at times associated with major changes in neural activity and hormonal levels.
28 citations
24 Oct 2018
TL;DR: A rigorous way to measure some of these biases is presented, based on the use of word lists created for social psychology applications, and a simple projection can significantly reduce the effects of embedding bias.
Abstract: Many modern Artificial Intelligence (AI) systems make use of data embeddings, particularly in the domain of Natural Language Processing (NLP). These embeddings are learnt from data that has been gathered “from the wild” and have been found to contain unwanted biases. In this paper we make three contributions towards measuring, understanding and removing this problem. We present a rigorous way to measure some of these biases, based on the use of word lists created for social psychology applications; we observe how gender bias in occupations reflects actual gender bias in the same occupations in the real world; and finally we demonstrate how a simple projection can significantly reduce the effects of embedding bias. All this is part of an ongoing effort to understand how trust can be built into AI systems.
21 citations
TL;DR: In this article, a corpus of Italian newspapers published in 1873-1914 in Gorizia, the county town of an area in the North Adriatic at the crossroad of the Latin, Slavic and Germanic civilizations is presented.
Abstract: We have digitised a corpus of Italian newspapers published in 1873–1914 in Gorizia, the county town of an area in the North Adriatic at the crossroad of the Latin, Slavic and Germanic civilizations...
11 citations
Posted Content•
TL;DR: This paper used word lists created for social psychology applications to measure gender bias in data embeddings and demonstrate how a simple projection can significantly reduce the effects of embedding bias, which is part of an ongoing effort to understand how trust can be built into AI systems.
Abstract: Many modern Artificial Intelligence (AI) systems make use of data embeddings, particularly in the domain of Natural Language Processing (NLP). These embeddings are learnt from data that has been gathered "from the wild" and have been found to contain unwanted biases. In this paper we make three contributions towards measuring, understanding and removing this problem. We present a rigorous way to measure some of these biases, based on the use of word lists created for social psychology applications; we observe how gender bias in occupations reflects actual gender bias in the same occupations in the real world; and finally we demonstrate how a simple projection can significantly reduce the effects of embedding bias. All this is part of an ongoing effort to understand how trust can be built into AI systems.
7 citations
TL;DR: In this paper, a Domain-Adversarial Neural Network (DANN) is proposed to remove unwanted information from a model using a gradient reversal layer, where the classifier makes a decision in a way that is entirely 'agnostic' to a given protected concept (e.g. gender, race, background etc.).
Abstract: We consider the problem of a neural network being requested to classify images (or other inputs) without making implicit use of a "protected concept", that is a concept that should not play any role in the decision of the network. Typically these concepts include information such as gender or race, or other contextual information such as image backgrounds that might be implicitly reflected in unknown correlations with other variables, making it insufficient to simply remove them from the input features. In other words, making accurate predictions is not good enough if those predictions rely on information that should not be used: predictive performance is not the only important metric for learning systems. We apply a method developed in the context of domain adaptation to address this problem of "being right for the right reason", where we request a classifier to make a decision in a way that is entirely 'agnostic' to a given protected concept (e.g. gender, race, background etc.), even if this could be implicitly reflected in other attributes via unknown correlations. After defining the concept of an 'agnostic model', we demonstrate how the Domain-Adversarial Neural Network can remove unwanted information from a model using a gradient reversal layer.
4 citations
Posted Content•
TL;DR: The History Playground as mentioned in this paper is an interactive web-based tool for discovering trends in massive textual corpora, which makes use of scalable algorithms to first extract trends from textual corpus, before making them available for real-time search and discovery, presenting users with an interface to explore the data.
Abstract: Recent studies have shown that macroscopic patterns of continuity and change over the course of centuries can be detected through the analysis of time series extracted from massive textual corpora. Similar data-driven approaches have already revolutionised the natural sciences, and are widely believed to hold similar potential for the humanities and social sciences, driven by the mass-digitisation projects that are currently under way, and coupled with the ever-increasing number of documents which are "born digital". As such, new interactive tools are required to discover and extract macroscopic patterns from these vast quantities of textual data. Here we present History Playground, an interactive web-based tool for discovering trends in massive textual corpora. The tool makes use of scalable algorithms to first extract trends from textual corpora, before making them available for real-time search and discovery, presenting users with an interface to explore the data. Included in the tool are algorithms for standardization, regression, change-point detection in the relative frequencies of ngrams, multi-term indices and comparison of trends across different corpora.
2 citations
24 Oct 2018
TL;DR: A method to support fact-checking of statements found in natural text such as online news, encyclopedias or academic repositories, by detecting if they violate knowledge that is implicitly present in a reference corpus is demonstrated.
Abstract: We demonstrate a method to support fact-checking of statements found in natural text such as online news, encyclopedias or academic repositories, by detecting if they violate knowledge that is implicitly present in a reference corpus. The method combines the use of information extraction techniques with probabilistic reasoning, allowing for inferences to be performed starting from natural text. We present two case studies, one in the domain of verifying claims about family relations, the other about political relations. This allows us to contrast the case where ground truth is available about the relations and the rules that can be applied to them (families) with the case where neither relations nor rules are clear cut (politics).
2 citations
24 Oct 2018
TL;DR: A comparative study of sentiment in news content across several languages is presented, assembling a new multilingual corpus and demonstrating that it is possible to detect variations in sentiment through machine translation.
Abstract: Rapid changes in public opinion have been observed in recent years about a number of issues, and some have attributed them to the emergence of a global online media sphere [1, 2]. Being able to monitor the global media sphere, for any sign of change, is an important task in politics, marketing and media analysis. Particularly interesting are sudden changes in the amount of attention and sentiment about an issue, and their temporal and geographic variations. In order to automatically monitor media content, to discover possible changes, we need to be able to access sentiment across various languages, and specifically for given entities or issues. We present a comparative study of sentiment in news content across several languages, assembling a new multilingual corpus and demonstrating that it is possible to detect variations in sentiment through machine translation. Then we apply the method on a number of real case studies, comparing changes in media coverage about Weinstein, Trump and Russia in the US, UK and some other EU countries.
1 citations