Contact person: Christian Müller (cmueller@dfki.de)

Internal Partners:

1. German Research Centre for Artificial Intelligence (DFKI), Christian Müller
2. Volkswagen AG, Andrii Kleshchonok

Involving human knowledge into the learning model can be done in diverse ways, e.g., by learning from expert data in imitation learning or reward engineering in deep reinforcement learning. In many applications, however, the expert data usually covers part of the search space or “normal” behaviors/scenarios. Learning a policy in the autonomous driving application under the limited dataset can make the policy vulnerable to novel or out of distribution (OOD) inputs and, thus, produce overconfident and dangerous actions. In this microproject, we aim to learn a policy based on the expert training data, while allowing the policy to go beyond data by interacting with an environment dynamics model and accounting uncertainty in the state estimation with virtual sensors. To avoid a dramatic shift of distribution, we propose to use the uncertainty of environment dynamics to penalize the policy for states that are different from human behavior.

Results Summary

Our key idea in this project is to learn a representation with deep models in a way to incorporate rules (e.g., physics equations governing dynamics of the autonomous vehicle) or distributions that can be simply defined by humans in advance. The learned representations from the source domain (the domain whose samples are based on the defined equations/distributions) are then transferred to the target domain with different distributions/rules and the model adapts itself by including target-specific features that can best explain variations of target samples w.r.t. underlying source rules/distributions. In this way, human knowledge is considered implicitly in the feature space.

We aim to develop a robust and generalized model that can perform well on out-of-distribution or novel data/domains/environments. The key idea is to learn the fundamental feature distribution shared between both source (training) and the target domain (test) while learning and including target-specific features that account for different variations of the shared distribution on the target domain.

Contact person: Richard Niestroj, VW Data:Lab Munich, Yuanting Liu (liu@fortiss.org; yuanting.liu@fortiss.org) 

Internal Partners:

1. Volkswagen AG, Richard Niestroj
2. Consiglio Nazionale delle Ricerche (CNR), Mirco Nanni
3. fortiss GmbH, Yuanting Liu  

The goal is to build a simulation environment to test connected car data based applications. AI based car data applications save people‘s time by guiding drivers and vehicles intelligently. This leads to a reduction of the environmental footprint of the transportation sector by reducing local and global emissions. The development and usage of a simulation environment enables data privacy compliancy for the development of AI based applications.

Tangible Outcomes

1. Video presentation summarizing the project

Contact person: Haris Papageorgiou (Athena RC) (haris@athenarc.gr) 

Internal Partners:

1. ATHENA RC,ILSP , Haris Papageorgiou, haris@athenarc.gr
2. DFKI, Georg Rehm, georg.rehm@dfki.de

Knowledge discovery offers numerous challenges and opportunities. In the last decade, a significant number of applications have emerged relying on evidence from the scientific literature. ΑΙ methods offer innovative ways of applying knowledge discovery methods in the scientific literature facilitating automated reasoning, discovery and decision making on data. This micro-project focuses on the task of question answering (QA) for the biomedical domain. Our starting point is a neural QA engine developed by ILSP addressing experts’ natural language questions by jointly applying document retrieval and snippet extraction on a large collection of PUBMED articles, thus, facilitating medical experts in their work. DFKI will augment this system with a knowledge graph integrating the output of document analysis and segmentation modules. The knowledge graph will be incorporated in the QA system and used for exact answers and more efficient Human-AI interactions. We primarily focus upon scientific articles on Covid-19 and SARS-CoV-2.

Tangible Outcomes

1. Video presentation summarizing the project

Contact person: Mireille Hildebrandt (m.hildebrandt@cs.ru.nl) 

Internal Partners:

1. University of Sussex (UOS)
2. German Research Centre for Artificial Intelligence (DFKI)
3. Vrije Universiteit Brussel (VUB)  

HumanE-AI research needs data to advance. Often, researchers struggle to progress because of the lack of data. At the same time, collecting a rich and accurate dataset is no easy task. Therefore, we propose to share through the AI4EU platform the datasets already collected so far by different research groups. The datasets are curated to be ready-to-use for researchers. Possible extension and variation of such datasets are also generated using artificial techniques and published on the platform. A performance baseline is provided for each dataset, in the form of publication reference, developed model or written documentation. The relevant legal framework will be investigated with specific attention to privacy and data protection, as to highlight limitations and challenges for the use and extension of existing datasets as well as future data collection on the subject of multimodal data collection for perception modelling.

Results Summary

There were 2 main outputs:

— Datasets: The partners involved have created, curated and released datasets for Human Activity Recognition (HAR) tasks, in particular, the extended dataset OPPORTUNITY++ and Wearlab BeachVolleyball dataset. The participation in the microproject has offered the chance to get a closer look at the practices, doubts and difficulties emerging within the scientific community involved in the creation, curation and dissemination of training datasets. Considering that one of the goals of the HumanE-AI Net is to connect research with relevant use cases in European society and industry, the participation to the microproject has offered the occasion to situate dataset collection, curation, and release within the broader context of AI pipeline.

— A comprehensive report introducing the concept of “Legal Protection Debt”: the Report examines the potential issues that arise within current ML-practices and provides an analysis of the relevant normative frameworks that govern such practices. By bridging the gap between practices and legal norms, the Report provides researchers with the tools to assess the risks to fundamental rights and freedoms that may occur due to the implementation of AI research in real world situations and recommends a set of mitigating measures to reduce infringements and to prevent violations.

The Report acknowledges that datasets constitute the backbone infrastructure underpinning the development of Machine Learning. The datasets that are created, curated and disseminated by ML practitioners provide the data to train ML models and the benchmarks to test the improvement of such models in performing the tasks for which they are intended.

However, until recently, the practices, processes and interactions that take place further upstream the ML-pipeline, between the collection of data and the use of dataset for training ML-models, have tended to fade into the background.

The report argues that the practices of dataset creation, curation and dissemination play a crucial role in the setting of the level of legal protection that is afforded to all the legal subjects that are located downstream ML-pipelines. Where such practices lack appropriate legal safeguards, a “Legal Protection Debt” can mount up incrementally along the stages of MLpipelines.

In section 1.1., the Report provides a brief overview of how current data science

practices depend on and perpetuate an ecosystem characterised by a lack of structural safeguards for the risks posed by data processing. This can lead to the accumulation of “technical debt”. Such debt, in turn, can assume relevance in the perspective of compliance with legal requirements. Taking inspiration from the literature on technical and ethical debt, the Report introduces the concept of Legal Protection Debt. Because of this legal protection debt, data-driven systems implemented at the end of the ML pipeline may lack the safeguards necessary to avoid downstream harm to natural persons.

The Report argues that the coming about of Legal Protection Debt and its accumulation at the end of the ML pipeline can be contrasted through the adoption of a Legal protection by design approach. This implies the overcoming of a siloed understanding of legal liability that mirrors the modular character of ML pipelines. Addressing legal protection debt requires ML practitioners to adopt a forward looking perspective. Such perspective should situates the stage of development in practitioners are involved in the context of the further stages that take place both upstream and downstream the pipeline. The consideration of the downstream stages of the ML-pipeline shall, as it were, back propagate and inform the

choices as to the technical and organisational measure to be taken upstream: upstream design decisions must be based on the anticipation of the downstream uses afforded by datasets and the potential harms that the latter may cause. Translated into a legal perspective, this implies that the actors upstream the pipeline should take into consideration the legal requirements that apply to the last stages of the pipeline.

The Report illustrates how data protection law lays down a set of legal equirements that overcome modularity and encompass the ML pipeline in its entirety, connecting the actors upstream with those downstream. The GDPR makes controllers responsible for the effects of the processing that they carry out. In section 2, the Report shows how the GDPR provides the tools to mitigate the problem of many hands in ML-pipelines. The duties and obligations set by the GDPR require controllers to implement by design safeguards that conjugate the

need to address downstream harms with the necessity to comply with the standards that govern scientific research. In this perspective, the Report shows that the obligations established by data protection law either instantiate or harden most of the requirements set by the Open science and Open data framework and also the best practices emerging within the ML-community.

In section 2.1. the report illustrates the core structure of the regime of liability to which controllers are subject under the GDPR. Such a regime of liability hinges upon controllers’ duty to perform a context-dependent judgment. Such judgment must inform controllers’ decisions as to the measures to be adopted to ensure compliance with all the obligations established by the GDPR. Such judgment must be based on the consideration of the downstream harms posed by the processing.

In essence, the duty to anticipate and address potential downstream harms requires controllers to adopt a forward-looking approach. In order to ensure compliance with the GDPR, controllers must engage in a dynamic, recursive practice that addresses the requirements of present processing in the light of the future potential developments. At the same time, the planning effort required by the GDPR is strictly connected with the compliance with obligations set by other normative frameworks. In this sense, compliance with the GDPR and compliance with obligations such as those imposed by the Open science and Open data framework go hand in hand. Compliance with the GDPR is a pre-requisite

for complying with Open science and Open data framework. Simultaneously, the perspective of open access and re-usability of datasets affects the content of the obligations set by the GDPR.

As a result, the consideration of “what happens downstream” – i.e., the potential uses of datasets, potential harms that the latter may cause, further requirements imposed by other normative frameworks – back propagates, determining the requirements that apply upstream.

In section 2.2. we show how the compliance with the documentation obligations set by the GDPR can contrast the accumulation of a documentation debt and ensure controllers’ compliance with the obligations established by other normative frameworks, such as Open Data and Open Science. The overlapping between the documentation requirements established by such different frameworks shows firstly that a serious approach to the compliance with the GDPR can provide the safeguards necessary to contrast the accumulation of a documentation debt. In this way, compliance with the documentation obligations set by the GDPR can prevent the accumulation of other forms of technical debt and, eventually, of legal protection debt. At the same time, the convergence between the requirements set by the GDPR and those established by the FAIR principle and the Horizon DMP template shows how the performance of the documentation obligations established by the GDPR can also facilitate compliance with requirements specific to data processing conducted in the context of scientific research.

A correct framing of the practices of dataset creation, curation and release in the context of research requires to make an effort towards the integrity of the legal framework as a whole, taking into consideration the relations between Open data, Open science and data protection law. First, it is first important to stress that compliance with data protection law represents a pre-requisite for the achievement of the goals of Open Data and Open Science framework.

In section 2.3. the report analyses the requirements that govern the release and downstream (re)use of datasets. Compliance with the requirements set by the GDPR is essential to avoid that dataset dissemination gives rise to the accumulation of legal protection debt along ML pipelines. Based on the assessment of adequacy and effectiveness required for all forms of processing, controllers can consider the adoption a range of measures to ensure that data

transfer are compliant with the GDPR. Among such measures, the Report examines the use of licenses, the providing of adequate documentation for the released dataset, data access management and traceability measures, including the use of unique identifiers.

The Report contains an Annex illustrating the provisions of the GDPR that establish a special regime for the processing carried out for scientific research purposes. We highlight how most of the provisions contained in the GDPR are not subject to any derogation or exemption in view of the scientific research purpose of the processing. All in all, the research regime provided by the GDPR covers the application of a limited number of provisions (or part of provisions). A process that is unlawful in that it does not comply with the general provisions set by the GDPR cannot enjoy the effects of the derogations provided by the research regime. The derogations allowed under the special research regime concern

almost exclusively the GDPR provisions on the rights of data subjects, while no derogation is possible for the general obligations that delineate the responsibility of the controller. The derogations provided under the special research regime allow controllers to modulate their obligations towards data subjects where the processing of personal data is not likely to affect significantly the natural persons that are identified or identifiable through such data. As it were, the decrease of the level of potential harm makes possible the lessening of the safeguards required to ensure the protection of data subjects. Even in such cases, however, no derogation is allowed with respect to the requirements different than those concerning the rights of the data subject. This circumstance makes manifest that the system established by the GDPR aims at providing a form of protection that goes beyond the natural persons whose personal data are processed at that time by controllers.

Contact person: Joao Gama ( jgama@fep.up.pt) 

Internal Partners:

1. INESC TEC, Joao Gama
2. Università di Pisa (UNIPI), Dino Pedreschi
3. Consiglio Nazionale delle Ricerche (CNR), Fosca Giannotti  

Nowadays ML models are used in decision-making processes in real-world problems by learning a function that maps the observed features with the decision outcomes. However, these models usually do not convey causal information about the association in observational data, thus not being easily understandable for the average user, therefore not being possible to retrace the models’ steps, nor rely on its reasoning. Hence, it is natural to investigate more explainable methodologies, such as causal discovery approaches, since they apply processes that mimic human reasoning. For this reason, we propose the usage of such methodologies to create more explicable models that replicate human thinking, and that are easier for the average user to understand. More specifically, we suggest its application in methods such as decision trees and random forest, since by themselves are highly explainable correlation-based methods.

Results Summary

In recent years, the study of causal relationships has become a crucial part of the Artificial Intelligence community, as causality can be a key tool for overcoming some limitations of correlation-based Machine Learning systems. Causality research can generally be divided into two main branches, that is, causal discovery and causal inference. The former focuses on obtaining causal knowledge directly from observational data. The latter aims to estimate the impact deriving from a change of a certain variable over an outcome of interest. The result of this project is a survey aiming at covering several methodologies that have been developed for both tasks. This survey does not only focus on theoretical aspects. But also provides a practical toolkit for interested researchers and practitioners, including software, datasets, and running examples. The published paper containts sections covering the following items. In Section 2, some basic definitions and notations are introduced. In Section 3, causal discovery techniques, tools, datasets, metrics, and examples are presented, organized by data type (cross-sectional, time-series, longitudinal). Section 4 covers causal inference techniques for several causal effects, tools, datasets, and a running example. Some remarks regarding the intersection between ML and causality are presented in Section 5, where some of the current open issues are also highlighted. Finally, conclusions are drawn.

Tangible Outcomes

1. Nogueira, Ana Rita, Andrea Pugnana, Salvatore Ruggieri, Dino Pedreschi, and João Gama. “Methods and tools for causal discovery and causal inference.” Wiley interdisciplinary reviews: data mining and knowledge discovery 12, no. 2 (2022): e1449. https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1449
2. Github repository of datasets, and papers related to causal discovery and causal inference research https://github.com/AnaRitaNogueira/Methods-and-Tools-for-Causal-Discovery-and-Causal-Inference
3. Github repository of software related to causal discovery and causal inference research https://github.com/AnaRitaNogueira/Methods-and-Tools-for-Causal-Discovery-and-Causal-Inference

Contact person: Hamraz javaheri (hamraz.javaheri@dfki.de) 

Internal Partners:

1. DFKI  

External Partners:

1. Hospital Saarbrücken “Der Winterberg”  

In this project, we successfully implemented and clinically evaluated an AR assistance system for pancreatic surgery, enhancing surgical navigation and achieving more precise perioperative outcomes. However, the system’s reliance on preoperative data posed challenges, particularly due to anatomical deformations occurring in the later stages of surgery. In future research, we aim to address this by integrating real-time data sources to further improve the system’s accuracy and adaptability during surgery.

Results Summary

Throughout our project, we developed and clinically evaluated ARAS, an augmented reality (AR) assistance system designed for pancreatic surgery. The system was clinically evaluated by field surgeons during pancreatic tumor resections involving 20 patients. In a matched-pair analysis with 60 patients who underwent surgery without ARAS, the ARAS group demonstrated a significantly shorter operation time compared to the control group. Although not statistically significant, the ARAS group also exhibited clinically noticeable lower rates of excessive intraoperative bleeding and reduced need for intraoperative red blood cell (RBC) transfusions. Furthermore, ARAS enabled more precise tumor resections with tumor-free margins, and patients in this group had better postoperative outcomes, including significantly shorter hospital stays. In this project, we published 2 journal papers (1 is accepted and will be published soon), 1 conference paper, 1 demo paper (Best Demo Paper Award), and 2 more conference papers are currently under submission. The success of our project got also several international and local news and media attention including Deutsche Welle news channel (Example links provided)

Tangible Outcomes

1. Beyond the visible: preliminary evaluation of the first wearable augmented reality assistance system for pancreatic surgery, Journal of International Journal of Computer Assisted Radiology and Surgery (https://doi.org/10.1007/s11548-024-03131-0 )
2. Enhancing Perioperative Outcomes of Pancreatic Surgery with Wearable Augmented Reality Assistance System: A Matched-Pair Analysis, Journal of Annals of Surgery Open ( https://doi.org/10.1097/AS9.0000000000000516)
3. Design and Clinical Evaluation of ARAS: An Augmented Reality Assistance System for Pancreatic Surgery (IEEE ISMAR 2024 (https://www.researchgate.net/publication/385116946_Design_and_Clinical_Evaluation_of_ARAS_An_Augmented_Reality_Assistance_System_for_Open_Pancreatic_Surgery_Omid_Ghamarnejad) 
4. ARAS: LLM-Supported Augmented Reality Assistance System for Pancreatic Surgery, ISWC/UbiComp 2024 (https://doi.org/10.1145/3675094.3677543) 
5. Media coverage for the project:
   1. https://www.dw.com/en/artificial-intelligence-saving-lives-in-the-operating-room/video-68125878
   2. https://www.dw.com/de/k%C3%BCnstliche-intelligenz-im-op-saal-rettet-leben/video-68125903
   3. https://www.saarbruecker-zeitung.de/app/consent/?ref=https%3A%2F%2Fwww.saarbruecker-zeitung.de%2Fsaarland%2Fsaarbruecken%2Fsaarbruecken%2Fsaarbruecken-winterberg-klinik-international-im-tv-zu-sehen_aid-106311259
   4. https://www.saarbruecker-zeitung.de/app/consent/?ref=https%3A%2F%2Fwww.saarbruecker-zeitung.de%2Fsaarland%2Fsaarbruecken-mittels-ki-erfolgreiche-operation-an-82-jaehriger-v29_aid-104053203
   5. https://m.focus.de/gesundheit/gesundleben/da-gibt-es-keinen-raum-fuer-fehler-kuenstliche-intelligenz-im-op-saal-rettet-leben_id_259629806.html 

Contact person: Samuel Kaski,, Aalto University (Samuel.kaski@aalto.fi) 

Internal Partners:

1. Aalto University, Samuel Kaski, samuel.kaski@aalto.fi
2. Delft University of Technology (TU Delft), Frans Oliehoek. F.A.Oliehoek@tudelft.nl

This micro-project contributes to developing methodologies that allow humans to be interactively involved “in the loop”. Here, the loop is a cooperative Bayesian optimization game where the goal is to optimize a 2D black-box function. At each iteration, the AI chooses the first coordinate, and then the user observes and opts for the second. Finally, the function is queried and the result is shown to both parties. The researcher can control agents’ characteristics, making it suitable for studying confirmation bias and imperfect knowledge. The project investigates how a planning AI agent can alleviate BO regret due to the human agent’s biases and imperfect information allocation. The aim is to build a planning AI agent to aid the user in the optimization task, where no single party has full decision-making power.

Results Summary

In this mini-project, we conducted an experiment with a synthetic user for various scenarios. We assumed the user decision process comprises two hierarchical steps:

updating the belief and taking action based on the belief. In the user model, these steps are regulated by model parameters, determining the conservatism in belief updates and exploration in opting for actions. Regarding that AI-assistant has a well-specified user model, we formulate the assistance decision problem as an MDP with unknown parameters ? and ?. We adjusted the Bayes-adaptive Monte Carlo planning methods to our problem to find the best policy for AI.

The reward in our planning AI is a weighted sum of two parts. Intuitively, one is responsible for ensuring that the AI favours actions for which the user can choose a promising second coordinate, and the other is responsible for reducing the risk that the user will act suboptimally, especially when the user model fails to predict the user well.

We compared our planning AI’s performance to a greedy AI (GreedyAI), which only tries to optimize the function based on its updated knowledge without considering the user. We also considered random query optimization (RandOpt) and a BO algorithm with the same acquisition function as GreedyAI uses but with full access to the data as logical lower and upper bounds for the optimization performance, respectively. The results demonstrate that the planning AI can assist the user in optimizing the function significantly better (measured as BO regret) than the GreedyAI and RandOpt, and relatively close to the logical upper bound under some conditions, even with its imperfect information.

Interestingly, a well-designed reward makes the cooperation effective even when the user follows a relatively high explorative policy. Investigating in-depth reveals that the planning AI lets the user explore the function adequately and reduces the chance of getting stuck at local optima.

In summary, the findings indicate that integrating a user model into a planning AI agent can mitigate potential biases of the user, enabling the team to avoid local optima and achieve better planning outcomes in sequential optimization games. By anticipating the user’s future behaviour, the agent can better guide the user towards optimal query.

Tangible Outcomes

1. Khoshvishkaie, A., Mikkola, P., Murena, P. A., & Kaski, S. (2023, September). Cooperative Bayesian optimization for imperfect agents. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases (pp. 475-490). Cham: Springer Nature Switzerland. https://link.springer.com/chapter/10.1007/978-3-031-43412-9_28

Contact person: Uwe Köckemann (uwe.kockemann@oru.se; michele.lombardi2@unibo.it) 

Internal Partners:

1. Örebro University (ORU), Uwe Köckemann
2. Università di Bologna (UNIBO), Michele Lombardi 

Methods for injecting constraints in Machine Learning (ML) can help bridging the gap between symbolic and sub-symbolic models, and address fairness and safety issues in data-driven AI systems. The recently proposed “Moving Targets” approach achieves this via a decomposition, where a classical ML model deals with the data and a separate constraint solver with the constraints. Different applications call for different constraints, solvers, and ML models; this flexibility is a strength of the approach, but it makes it also difficult to set up and analyze. Therefore, this project relies on the AI Domain Definition Language (AIDDL) framework to obtain a flexible implementation of the approach, making it simpler to use and allowing the exploration of more case studies, different constraint solvers, and algorithmic variants. We used this implementation to investigate various new constraint types integrated with the Moving Targets approach (e.g., SMT, MINLP, CP).

Results Summary

The moving targets method integrates machine learning and constraint optimization to enforce constraints on a machine learning model. The AI Domain Definition Language (AIDDL) provides a modeling language and framework for integrative AI.

We have implemented the moving targets algorithm in the AIDDL framework for integrative AI. This has benefits for modeling, experimentation, and usability. On the modeling side, this enables us to provide applications of “moving target” as regular machine learning problems extended with constraints and a loss function. On the experimentation side, we can now easily switch the learning and constraint solvers used by the “moving targets” algorithm, and we have added support for multiple constraint types. Finally, we made the “moving targets” method easier to use, since it can now be controlled through a small model written in the AIDDL language.

Tangible Outcomes

1. Example Jupyter Notebooks (3 data sets) – Uwe Köckemann, Fabrizio Detassis, Michele Lombardi. https://gitsvn-nt.oru.se/uwe.kockemann/moving-targets 
2. Experiments Jupyter Notebooks (3 data sets) – Uwe Köckemann, Fabrizio Detassis, Michele Lombardi. https://gitsvn-nt.oru.se/uwe.kockemann/moving-targets 
3. Program/code: Python library: Moving targets via AIDDL – Uwe Köckemann, Fabrizio Detassis, Michele Lombardi. https://gitsvn-nt.oru.se/uwe.kockemann/moving-targets 
4. Moving targets tutorial – Michele Lombardi. https://gitsvn-nt.oru.se/uwe.kockemann/moving-targets 
5. presentation about the project. https://gitsvn-nt.oru.se/uwe.kockemann/moving-targets/-/blob/master/presentations/microproject_presentation_ORU-UBO.pptx 
6. Video presentation summarizing the project

Contact person: Mireia Diez Sanchez (mireia@fit.vutbr.cz) 

Internal Partners:

1. BUT, Brno University of Technology, Mireia Diez Sanchez, mireia@fit.vutbr.cz; cernocky@fit.vutbr.cz
2. TUB, TECHNISCHE UNIVERSITÄT BERLIN, Tim Polzehl, tim.polzehl@dfki.de;  klaus.r.mueller@googlemail.com

In this microproject, we pursued enabling access to AI technology to those who might have special needs when interacting with “AI: Automatic Speech Recognition made accessible for people with dysarthria”. Dysarthria is a motor speech disorder resulting from neurological injury and is characterized by poor articulation of phonemes. Within Automatic speech recognition (ASR), dysarthric speech recognition is a tedious task due to the lack of supervised data and diversity.

The project studied the adaptation of automatic speech recognition (ASR) systems for impaired speech. Specifically, the micro-project focused on improving ASR systems for speech from subjects with dysarthria and/or stuttering speech impairment types of various degrees. The work was developed using German “Lautarchive” data comprising only 130 hours of untranscribed doctor-patient German speech conversations and using English TORGO dataset, applying human-in-the-loop methods. We spot individual errors and regions of low certainty in ASR in order to apply human originated improvement and clarification in AI decision processes.

Results Summary

Particularly, in this work, we have studied the performance of different ASR systems on dysarthric speech: LF-MMI, Transformer and wav2vec2. The analysis revealed the superiority of the wav2vec2 models on the task. We investigated the importance of speaker dependent auxiliary features such as fMLLR and xvectors for adapting wav2vec2 models for improving dysarthric speech recognition. We showed that in contrast to hybrid systems, wav2vec2 did not improve by adapting model parameters based on each speaker.

We proposed a wav2vec2 adapter module that inherits speaker features as auxiliary information to perform effective speaker normalization during finetuning. We showed that, using the adapter module, fMLLR and xvectors are complementary to each other, and proved the effectiveness of the approach outperforming existing SoTA on UASpeech dysartric speech ASR.

In our cross-lingual experiments, we also showed that combining English and German data for training, can further improve performance of our systems, proving useful in scenarios where little training examples exist for a particular language.

Tangible Outcomes

1. M. K. Baskar, T. Herzig, D. Nguyen, M. Diez, T. Polzehl, L. Burget, J. Černocký, “Speaker adaptation for Wav2vec2 based dysarthric ASR”. Proc. Interspeech 2022, 3403-3407, doi: 10.21437/Interspeech.2022-10896 https://www.isca-speech.org/archive/pdfs/interspeech_2022/baskar22b_interspeech.pdf
2. Open source tool for training ASR models for dysarthic speech: The repository contains: A baseline recipe to train a TDNN-CNN hybrid model based ASR system, this recipe is prepared to be trained on the TORGO dataset. And an end-to-end model using ESPnet framework prepared to be trained on UASpeech dataset. https://github.com/creatorscan/Dysarthric-ASR

Contact person: Dino Pedreschi (dino.pedreschi@unipi.it) 

Internal Partners:

1. University of Pisa – Department of CS, Dino Pedreschi (dino.pedrschi@unipi.it)   

External Partners:

1. University of Antwerp – Department of CS, Daphne Lenders (daphne.lenders@uantwerpen.be) 
2. Scuola Normale Superiore, Roberto Pellungrini (roberto.pellungrini@sns.it)   

Our project revolves around the topic of fair Artificial Intelligence(AI), a field that explores how decision-making algorithms used in high-stake domains, such as hiring or loan allocations, can perpetuate discriminatory patterns in the data they are based on, unfairly affecting people of certain races, genders or other demographics. Early attempts to address bias in AI systems focused on automated solutions, attempting to eliminate discrimination by establishing mathematical definitions of “fairness” and optimizing algorithms accordingly. However, these approaches have faced justified criticism for disregarding the contextual nuances in which algorithms operate, and for neglecting the input of domain experts who understand and can tackle discriminatory patterns effectively. Consequently, policymakers have recognized the pitfalls of solely relying on these approaches and are now designing legal regulations, mandating that high-risk AI systems can only be deployed when they allow for oversight and intervention by human experts. With our project, we investigate how to effectively achieve this human control by exploring the intersection between fair and explainable AI (xAI), whereas the latter is concerned with explaining the decision processes of otherwise opaque black-box algorithms. We develop a tool that provides humans with explanations about an algorithmic decision-making system. Based on the explanations, users can give feedback about the system’s fairness and choose between different strategies to mitigate its discriminatory patterns. By immediately getting feedback about the effects of their chosen strategy, users can engage in an iterative process, further refining and improving the algorithm. Since little prior work has been done on Human-AI collaboration in the context of bias mitigation, we took on an exploratory approach to evaluate this system. We set up a think-aloud study where potential end-users can interact with the system and try out different mitigation strategies. We analysed their responses and thoughts, to identify the tool’s strengths and weaknesses as well as users’ mental model of the tool. Additionally, we compared the systems’ biases before and after human intervention, to see how biases were mitigated and how successful this mitigation was.

Results Summary

We developed an algorithm that can reject predictions both based on their uncertainty and their unfairness. By rejecting possibly unfair predictions, our method reduces error and positive decision rate differences across demographic groups of the non-rejected data. Since the unfairness-based rejections are based on an interpretable-by-design method, i.e., rule-based fairness checks and situation testing, we create a transparent process that can empower human decision-makers to review the unfair predictions and make more just decisions for them. This explainable aspect is especially important in light of recent AI regulations, mandating that any high-risk decision task should be overseen by human experts to reduce discrimination risks. This methodology allows us to essentially bridge the gap between classifiers with a reject option and interpretable by design methods, encouraging human intervention and comprehension. We produced a functioning software, which is available, and are working on a full publication with experiments on multiple datasets and multiple rejection strategies. A publication is planned out of the outcome.

Tangible Outcomes

1. The full software: https://github.com/calathea21/IFAC 

Contact person: Jesus Cerquides (j.cerquides@csic.es)

Internal Partners:

1. CSIC Consejo Superior de Investigaciones Científicas, Jesus Cerquides
2. CNR Consiglio Nazionale delle Ricerche, Daniele Vilone   

Many citizen science projects have a crowdsourcing component where several different citizen scientists are requested to fulfill a micro task (such as tagging an image as either relevant or irrelevant for the evaluation of damage in a natural disaster, or identifying a specimen into its taxonomy). How do we create a consensus between the different opinions/votes? Currently, most of the time, simple majority voting is used. We argue that alternative voting schemes (taking into account the errors performed by each annotator) could severely reduce the number of citizen scientists required. This is a clear example of continuous human-in-the-loop machine learning with the machine creating a model of the humans that it has to interact with. We propose to study consensus building under two different hypotheses: truthful annotators (as a model for most voluntary citizen science projects) and self-interested annotators (as a model for paid crowdsourcing projects).

Results Summary

We have contributed to the implementation of several different probabilistic consensus models in the Crowdnalysis library which has been resealed as a Python package.

We have proposed a generic mathematical framework for the definition of probabilistic consensus algorithms, and for performing prospective analysis. This has been published in a journal paper.

We have used the library and the mathematical framework for the analysis of images from the Albanian earthquake scenario.

We exploited Monte Carlo simulations to understand which can be the best way to assess group decisions in evaluating the correct level of damage in natural catastrophes. The results suggest that Majority rule is the best option as long as all the agents are competent enough to address the task. Otherwise, when the number of unqualified agents is no longer negligible, smarter procedures must be found out.

Tangible Outcomes

1. Program/code: Crowdnalysis Python package – Jesus Cerquides (cerquide@iiia.csic.es) https://pypi.org/project/crowdnalysis/ 

Contact person: Joao Gama (INESC TEC) (jgama@fep.up.pt)

Internal Partners:

1. INESC-Tech Joao Gama,
2. CNR, Giuseppe Manco,
3. ULEI, Holger Hoos 

The goal is to devise a data generation methodology that, given a data sample, can approximate the stochastic process that generated it. The methodology can be useful in many contexts where we need to share data while preserving user privacy. There are known literature for data generation based on Bayesian neural networks/hidden Markov models that are restricted to static and propositional data. We focus on time-evolving data and preference data. We will study essentially two aspects: (1) the generator to produce realistic data, having the same properties of the original one, and (2) we want to investigate how to inject drift within the data generation process in a controlled manner. The idea is to model the stochastic process through a dependency graph among random variables so that the drift can be simply modeled by changing the structure of the underlying graph through a morphing process.

Tangible Outcomes

1. available on github – https://github.com/fsp22/mcd_dds4rs 
2. implementation of the model presented in the paper “Modelling Concept Drift in Dynamic Data Streams for Recommender Systems” https://github.com/fsp22/mcd_dds4rs