1,320 research outputs found
Designing responsible artificial intelligence:hybrid approaches for aligning learning and reasoning
Full text linkArtificial Intelligence (AI) has become an integral part of our society: we have smart assistants with speech recognition on our phones, self-driving cars, and online algorithms that recommend what we should buy, watch or listen to. Most of these AI systems learn to make decisions based on data: large quantities of examples from the past. The exact internal reasoning of such AI systems that learn from data is difficult to determine, however. This can cause the AI system to behave irresponsibly. In this thesis, we introduce a method to evaluate the internal reasoning of AI systems that learn from data. We show that AI systems sometimes make the right decisions, but for the wrong reasons. For example, unbeknownst to us, an AI system can learn an undesirable, hidden bias from the data. The method that we describe in our thesis can not only evaluate the internal reasoning of an AI system, but can also adjust it and steer it in the right direction. Additionally, we also show how one can create an AI system with predefined reasoning, rather than making it learn its reasoning from data. This way, the system cannot accidentally learn to make the decisions for the wrong reasons. All of the methods we discuss in the thesis build upon the idea that we should use the domain knowledge of human experts when designing AI systems that learn from data. The thesis shows that this is essential for designing responsible artificial intelligence
Rationale Discovery and Explainable AI
Full text linkThe justification of an algorithm's outcomes is important in many domains, and in particular in the law. However, previous research has shown that machine learning systems can make the right decisions for the wrong reasons: despite high accuracies, not all of the conditions that define the domain of the training data are learned. In this study, we investigate what the system does learn, using state-of-the-art explainable AI techniques. With the use of SHAP and LIME, we are able to show which features impact the decision making process and how the impact changes with different distributions of the training data. However, our results also show that even high accuracy and good relevant feature detection are no guarantee for a sound rationale. Hence these state-of-the-art explainable AI techniques cannot be used to fully expose unsound rationales, further advocating the need for a separate method for rationale evaluation. </p
Discovering the Rationale of Decisions:Towards a Method for Aligning Learning and Reasoning
Full text linkIn AI and law, systems that are designed for decision support should be explainable when pursuing justice. In order for these systems to be fair and responsible, they should make correct decisions and make them using a sound and transparent rationale. In this paper, we introduce a knowledge-driven method for model-agnostic rationale evaluation using dedicated test cases, similar to unit-testing in professional software development. We apply this new quantitative human-in-the-loop method in a machine learning experiment aimed at extracting known knowledge structures from artificial datasets from a real-life legal setting. We show that our method allows us to analyze the rationale of black box machine learning systems by assessing which rationale elements are learned or not. Furthermore, we show that the rationale can be adjusted using tailor-made training data based on the results of the rationale evaluation
Hypertext: A Sacred (He)Art?: Cor ad Cor Loquitur from Augustine to Shelley Jackson
Full text linkSelf-discovery, self-exploration, the creation of the self or the Subject is a human preoccupation that goes beyond the postmodern era. The epigraphs that begin this paper show that the human concern with how language and representation play a crucial role in the formation of the subject flows back through time from our present to Augustine, the fourth-century master of the art of self-knowledge, and beyond. When Augustine started writing his Confessions, the self as something to write about, a theme or object (subject) of writing activity, was already well established. In his Confessions, Augustine uses cor ad cor loquitur, or to put it plainly, having a heart to heart with God. Such a conversation was meant to change his life by teaching him how to revise himself in Christ\u27s image. In other words, cor ad cor loquitur is a lesson in subjectivity.
Today, as someone who is a medievalist, theologian, and techno-geek, I find myself pondering how this ancient and never-ending conversation echoes still, even in the realm of hypertext. And yes! I did say hypertext. As theologian and medievalist, I wander on my pilgrim way in many different worlds, antique and contemporary. For me, the hypertext world of Cyberia (that computerized technological world in to which we are presently evolving) continues the ancient trail of a conversation, of heart speaking to heart, in which subjectivity evolves. The mechanism of self-reflection, central to cor ad cor loquitur, resides in the rhetorical structure of hypertext. Contemporary pilgrims negotiating their way as author and audience through the lexias[i] and byways of Cyberia\u27s hypertext find themselves following in the footsteps of their medieval ancestors who pondered on author and audience in the book of the heart known as cor ad cor loquitur. I invite you to accompany me as I use the medievalist’s lens to investigate how hypertext is the latest evolution in cor ad cor loquitur
Improving Rationales with Small, Inconsistent and Incomplete Data
Full text linkData-driven AI systems can make the right decisions for the wrong reasons, which can lead to irresponsible behavior. The rationale of such machine learning models can be evaluated and improved using a previously introduced hybrid method. This method, however, was tested using synthetic data under ideal circumstances, whereas labelled datasets in the legal domain are usually relatively small and often contain missing facts or inconsistencies. In this paper, we therefore investigate rationales under such imperfect conditions. We apply the hybrid method to machine learning models that are trained on court cases, generated from a structured representation of Article 6 of the ECHR, as designed by legal experts. We first evaluate the rationale of our models, and then improve it by creating tailored training datasets. We show that applying the rationale evaluation and improvement method can yield relevant improvements in terms of both performance and soundness of rationale, even under imperfect conditions
Discovering the Rationale of Decisions:Experiments on Aligning Learning and Reasoning
Full text linkIn AI and law, systems that are designed for decision support should be explainable when pursuing justice. In order for these systems to be fair and responsible, they should make correct decisions and make them using a sound and transparent rationale. In this paper, we introduce a knowledge-driven method for model-agnostic rationale evaluation using dedicated test cases, similar to unit-testing in professional software development. We apply this new method in a set of machine learning experiments aimed at extracting known knowledge structures from artificial datasets from fictional and non-fictional legal settings. We show that our method allows us to analyze the rationale of black-box machine learning systems by assessing which rationale elements are learned or not. Furthermore, we show that the rationale can be adjusted using tailor-made training data based on the results of the rationale evaluation
The XAI Paradox:Systems that perform well for the wrong reasons
Full text linkMany of the successful modern machine learning approaches can be described as ``black box'' systems; these systems perform well, but are unable to explain the reasoning behind their decisions. The emerging sub-field of Explainable Artificial Intelligence (XAI) aims to create systems that are able to explain to their users why they made a particular decision. Using artificial datasets whose internal structure is known beforehand, this study shows that the reasoning of systems that perform well is not necessarily sound. Furthermore, when multiple combined conditions define a dataset, systems can preform well on the combined problem and not learn each of the individual conditions. Instead, it often learns a confounding structure within the data that allows it to make the correct decisions. With regards to the goal of creating explainable systems, however, unsound rationales could create irrational explanations which would be problematic for the XAI movement.<br/
USP44 Is an Integral Component of N-CoR that Contributes to Gene Repression by Deubiquitinating Histone H2B
No full textSummaryDecreased expression of the USP44 deubiquitinase has been associated with global increases in H2Bub1 levels during mouse embryonic stem cell (mESC) differentiation. However, whether USP44 directly deubiquitinates histone H2B or how its activity is targeted to chromatin is not known. We identified USP44 as an integral subunit of the nuclear receptor co-repressor (N-CoR) complex. USP44 within N-CoR deubiquitinates H2B in vitro and in vivo, and ablation of USP44 impairs the repressive activity of the N-CoR complex. Chromatin immunoprecipitation (ChIP) experiments confirmed that USP44 recruitment reduces H2Bub1 levels at N-CoR target loci. Furthermore, high expression of USP44 correlates with reduced levels of H2Bub1 in the breast cancer cell line MDA-MB-231. Depletion of either USP44 or TBL1XR1 impairs the invasiveness of MDA-MB-231 cells in vitro and causes an increase of global H2Bub1 levels. Our findings indicate that USP44 contributes to N-CoR functions in regulating gene expression and is required for efficient invasiveness of triple-negative breast cancer cells
Taking the Law More Seriously by Investigating Design Choices in Machine Learning Prediction Research
Full text linkApproaches to court case prediction using machine learning differ widely with varying levels of success and legal reasonableness. In part this is due to some aspects of law, such as justification, being inherently difficult for machine learning approaches. Another aspect is the effect of design choices and the extent to which these are legally reasonable, which has not yet been extensively studied. We create four machine learning models tasked with predicting cases from the European Court of Human Rights and we perform experiments in order to measure the role of the following four design choices and effects: the choice of performance metric; the effect of including different parts of the legal case; the effect of a more or less specialized legal focus; and the temporal effects of the available past legal decisions. Through this research, we aim to study design decisions and their limitations and how they affect the performance of machine learning models.</p
- …
