Many real-life scenarios require humans to make difficult trade-offs: do we always follow all the traffic rules or do we violate the speed limit in an emergency? These scenarios force us to evaluate the trade-off between collective norms and our own personal objectives. To create effective AI-human teams, we must equip AI agents with a model of how humans make trade-offs in complex, constrained environments. These agents will be able to mirror human behavior or to draw human attention to situations where decision making could be improved. To this end, we propose a novel inverse reinforcement learning (IRL) method for learning implicit hard and soft constraints from demonstrations, enabling agents to quickly adapt to new settings. In addition, learning soft constraints over states, actions, and state features allows agents to transfer this knowledge to new domains that share similar aspects. We then use the constraint learning method to implement a novel system architecture that leverages a cognitive model of human decision making, multi-alternative decision field theory (MDFT), to orchestrate competing objectives. We evaluate the resulting agent on trajectory length, number of violated constraints, and total reward, demonstrating that our agent architecture is both general and achieves strong performance. Thus we are able to capture and replicate human-like trade-offs from demonstrations in environments when constraints are not explicit.

This document sums up our results forthe NLP lecture at ETH in the spring semester 2021. In this work, a BERT based neural network model (Devlin et al.,2018) is applied to the JIGSAW dataset (Jigsaw/Conversation AI, 2019) in order to create a model identifying hateful and toxic comments (strictly seperated from offensive language) in online social platforms (English language), inthis case Twitter. Three other neural network architectures and a GPT-2 (Radfordet al., 2019) model are also applied on the provided data set in order to compare these different models. The trained BERT model is then applied on two different data sets to evaluate its generalisation power, namely on another Twitter data set (Tom Davidson, 2017) (Davidsonet al., 2017) and the data set HASOC 2019 (Thomas Mandl, 2019) (Mandl et al.,2019) which includes Twitter and also Facebook comments; we focus on the English HASOC 2019 data. In addition, it can be shown that by fine-tuning the trained BERT model on these two datasets by applying different transfer learning scenarios via retraining partial or all layers the predictive scores improve compared to simply applying the model pre-trained on the JIGSAW data set. Withour results, we get precisions from 64% to around 90% while still achieving acceptable recall values of at least lower 60s%, proving that BERT is suitable for real usecases in social platforms.

Object detection is a fundamental task in computer vision. While approaches for axis-aligned bounding box detection have made substantial progress in recent years, they perform poorly on oriented objects which are common in several real-world scenarios such as aerial view imagery and security camera footage. In these cases, a large part of a predicted bounding box will, undesirably, cover non-object related areas. Therefore, oriented object detection has emerged with the aim of generalizing object detection to arbitrary orientations. This enables a tighter fit to oriented objects, leading to a better separation of bounding boxes especially in case of dense object distributions. The vast majority of the work in this area has focused on complex two-stage anchor-based approaches. Anchors act as priors on the bounding box shape and require attentive hyper-parameter fine-tuning on a per-dataset basis, increased model size, and come with computational overhead. In this work, we present DAFNe: A Dense one-stage Anchor-Free deep Network for oriented object detection. As a one-stage model, DAFNe performs predictions on a dense grid over the input image, being architecturally simpler and faster, as well as easier to optimize than its two-stage counterparts. Furthermore, as an anchor-free model, DAFNe reduces the prediction complexity by refraining from employing bounding box anchors. Moreover, we introduce an orientation-aware generalization of the center-ness function for arbitrarily oriented bounding boxes to down-weight low-quality predictions and a center-to-corner bounding box prediction strategy that improves object localization performance. DAFNe improves the prediction accuracy over the previous best one-stage anchor-free model results on DOTA 1.0 by 4.65% mAP, setting the new state-of-the-art results by achieving 76.95% mAP.

The recent advancements in artificial intelligence (AI) combined with the extensive amount of data generated by today's clinical systems, has led to the development of imaging AI solutions across the whole value chain of medical imaging, including image reconstruction, medical image segmentation, image-based diagnosis and treatment planning. Notwithstanding the successes and future potential of AI in medical imaging, many stakeholders are concerned of the potential risks and ethical implications of imaging AI solutions, which are perceived as complex, opaque, and difficult to comprehend, utilise, and trust in critical clinical applications. Despite these concerns and risks, there are currently no concrete guidelines and best practices for guiding future AI developments in medical imaging towards increased trust, safety and adoption. To bridge this gap, this paper introduces a careful selection of guiding principles drawn from the accumulated experiences, consensus, and best practices from five large European projects on AI in Health Imaging. These guiding principles are named FUTURE-AI and its building blocks consist of (i) Fairness, (ii) Universality, (iii) Traceability, (iv) Usability, (v) Robustness and (vi) Explainability. In a step-by-step approach, these guidelines are further translated into a framework of concrete recommendations for specifying, developing, evaluating, and deploying technically, clinically and ethically trustworthy AI solutions into clinical practice.

Malignant melanoma is a common skin cancer that is mostly curable before metastasis -when growths spawn in organs away from the original site. Melanoma is the most dangerous type of skin cancer if left untreated due to the high risk of metastasis. This paper presents Melatect, a machine learning (ML) model embedded in an iOS app that identifies potential malignant melanoma. Melatect accurately classifies lesions as malignant or benign over 96.6% of the time with no apparent bias or overfitting. Using the Melatect app, users have the ability to take pictures of skin lesions (moles) and subsequently receive a mole classification. The Melatect app provides a convenient way to get free advice on lesions and track these lesions over time. A recursive computer image analysis algorithm and modified MLOps pipeline was developed to create a model that performs at a higher accuracy than existing models. Our training dataset included 18,400 images of benign and malignant lesions, including 18,000 from the International Skin Imaging Collaboration (ISIC) archive, as well as 400 images gathered from local dermatologists; these images were augmented using DeepAugment, an AutoML tool, to 54,054 images.

In this paper, we develop a new classification method based on nearest centroid, and it is called the nearest disjoint centroid classifier. Our method differs from the nearest centroid classifier in the following two aspects: (1) the centroids are defined based on disjoint subsets of features instead of all the features, and (2) the distance is induced by the dimensionality-normalized norm instead of the Euclidean norm. We provide a few theoretical results regarding our method. In addition, we propose a simple algorithm based on adapted k-means clustering that can find the disjoint subsets of features used in our method, and extend the algorithm to perform feature selection. We evaluate and compare the performance of our method to other closely related classifiers on both simulated data and real-world gene expression datasets. The results demonstrate that our method is able to outperform other competing classifiers by having smaller misclassification rates and/or using fewer features in various settings and situations.

As artificial intelligence enters the legal space, it is essential to recognize biases in legal data and ensure that they are not replicated and reinforced with legal technology [7, 13, 18]. Furthermore, understanding biases in legal data and developing discrimination-free technology could help the legal space to become fairer and more widely accessible. We typically find two types of biases in legal data: First, representation biases, i.e., certain social groups are over-or underrepresented in a data set. Second, sentencing disparities, i.e., the outcome of legal proceedings for similar cases varies across social groups. Representation biases may reflect disparities in policing (arrest rates) or in offense rates.

Modern streaming services are increasingly labeling videos based on their visual or audio content. This typically augments the use of technologies such as AI and ML by allowing to use natural speech for searching by keywords and video descriptions. Prior research has successfully provided a number of solutions for speech to text, in the case of a human speech, but this article aims to investigate possible solutions to retrieve sound events based on a natural language query, and estimate how effective and accurate they are. In this study, we specifically focus on the YamNet, AlexNet, and ResNet-50 pre-trained models to automatically classify audio samples using their respective melspectrograms into a number of predefined classes. The predefined classes can represent sounds associated with actions within a video fragment. Two tests are conducted to evaluate the performance of the models on two separate problems: audio classification and intervals retrieval based on a natural language query. Results show that the benchmarked models are comparable in terms of performance, with YamNet slightly outperforming the other two models. YamNet was able to classify single fixed-size audio samples with 92.7% accuracy and 68.75% precision while its average accuracy on intervals retrieval was 71.62% and precision was 41.95%. The investigated method may be embedded into an automated event marking architecture for streaming services.

In recent years, machine learning (ML) algorithms have been deployed in safety-critical and high-stake decision-making, where the fairness of algorithms is of paramount importance. Fairness in ML centers on detecting bias towards certain demographic populations induced by an ML classifier and proposes algorithmic solutions to mitigate the bias with respect to different fairness definitions. To this end, several fairness verifiers have been proposed that compute the bias in the prediction of an ML classifier -- essentially beyond a finite dataset -- given the probability distribution of input features. In the context of verifying linear classifiers, existing fairness verifiers are limited by accuracy due to imprecise modelling of correlations among features and scalability due to restrictive formulations of the classifiers as SSAT or SMT formulas or by sampling. In this paper, we propose an efficient fairness verifier, called FVGM, that encodes the correlations among features as a Bayesian network. In contrast to existing verifiers, FVGM proposes a stochastic subset-sum based approach for verifying linear classifiers. Experimentally, we show that FVGM leads to an accurate and scalable assessment for more diverse families of fairness-enhancing algorithms, fairness attacks, and group/causal fairness metrics than the state-of-the-art. We also demonstrate that FVGM facilitates the computation of fairness influence functions as a stepping stone to detect the source of bias induced by subsets of features.

We investigate novel parameter estimation and goodness-of-fit (GOF) assessment methods for large-scale confirmatory item factor analysis (IFA) with many respondents, items, and latent factors. For parameter estimation, we extend Urban and Bauer's (2021) deep learning algorithm for exploratory IFA to the confirmatory setting by showing how to handle user-defined constraints on loadings and factor correlations. For GOF assessment, we explore new simulation-based tests and indices. In particular, we consider extensions of the classifier two-sample test (C2ST), a method that tests whether a machine learning classifier can distinguish between observed data and synthetic data sampled from a fitted IFA model. The C2ST provides a flexible framework that integrates overall model fit, piece-wise fit, and person fit. Proposed extensions include a C2ST-based test of approximate fit in which the user specifies what percentage of observed data can be distinguished from synthetic data as well as a C2ST-based relative fit index that is similar in spirit to the relative fit indices used in structural equation modeling. Via simulation studies, we first show that the confirmatory extension of Urban and Bauer's (2021) algorithm produces more accurate parameter estimates as the sample size increases and obtains comparable estimates to a state-of-the-art confirmatory IFA estimation procedure in less time. We next show that the C2ST-based test of approximate fit controls the empirical type I error rate and detects when the number of latent factors is misspecified. Finally, we empirically investigate how the sampling distribution of the C2ST-based relative fit index depends on the sample size.