TLDR

When: 28th July, 2022

Time: 09:00 - 17:30

Where: room G300 (new room!), University of Konstanz and on Zoom (online only the talks)

Here, you can find a guide on How to Get To the Room

Registration: To help us with the organization, please, register HERE by Friday 15.07.2022.

Overview

This workshop is the closing event of the MA course "The Present and Future of AI Research" offered by the Department of Linguistics and the Department of Philosophy of the University of Konstanz. Five invited speakers will present their work on different aspects that are crucial in Artificial Intelligence. At the same event, our students will present their scientific posters on topics related to the various talks.

Program (updated schedule!)

09:00 - 09:15: registration & coffee break

09:15 - 09:30: Welcome and Introduction

09:30 - 10:30: Menna El-Assady, "Human-AI Collaboration for Mixed-Initiative Topic Model Refinement"

10:30 - 11:00: coffee break (room: G201- downstairs)

11:00 - 12:00: Philipp Kellmeyer, "Challenges in Human-AI Interaction in Medicine: Interpretability, Fairness and Ethics-by-Design"

12:00 - 12:15: short break

12:15 - 13:15: Melika Payvand, "Closing the Gap Between Devices, Circuits and Algorithms Towards Efficient AI Hardware" (zoom talk)

13:15 - 14:30: (not organized) lunch break

14:30 - 15:30: Laura Candiotto, "The Affective Dimension of Trust: Why Emotions Make Us Trust an AI"

15:30 - 16:00: coffee break (room: G201- downstairs)

16:00 - 17:00: Massimiliano Di Luca, "Computational Modelling of Perception and Action"

17:00 - 17:30: closing remarks

Speakers

Laura Candiotto, University of Pardubice

Title: "The Affective Dimension of Trust: Why Emotions Make us Trust an AI".

Abstract: After the publication of the now-classic "extended mind" paper by Andy Clark and David Chalmers (Clark & Chalmers 1998), a profuse debate on the conditions that should be met for granting the extension of the mind in the world has taken place at the crossroads of philosophy of mind, cognitive science, and artificial intelligence. Notably, Clark (2010) has claimed that some “glue and trust” criteria allow us to differentiate between genuine extended cognition and the mere instrumental employment of a technological tool. Unfortunately, the affective dimension of trust has been rarely studied in this regard. In this talk, I will argue that trust has a crucial affective component. We cannot have a phenomenologically accurate description of what trusting an AI means without considering it. But I will also advance a stronger hypothesis, namely that without affective trust, cognitive integration is not attainable most of the time. So, if my hypothesis is correct, affective trust is a necessary condition of extended cognition. I will take assistive technology (AT) as a case study.

The final upshot is that affective trust, understood as an affective habit, enables the subject to carve out new relationships with the environment through the employment of AT. In tailoring a new affective-cum-cognitive niche through trustful engagement with technological devices, agents can extend their action possibilities and thus increase their feeling of agency in the world.

Menna El-Assady, ETH Zurich

Title: "Human-AI Collaboration for Mixed-Initiative Topic Model Refinement".

Abstract: Applying topic modeling algorithms to analyze the content of text corpora is a prevalent task in the humanities and social sciences. However, as their results are typically subjective and domain-dependent, no single ground-truth segmentation of a corpus can be used to optimize such models. Hence, refining these models has relied on humans externalizing their knowledge to adapt the results to their domain understanding. To tackle this challenge, I develop human-centered machine learning approaches, contributing novel mixed-initiative paradigms to explain, diagnose, and refine topic models. In this talk, I present four different workflow and interface designs, each tailored to a different user group. First, data scientists can use the progressive learning approach to give feedback on the relation between inputs and outputs efficiently. Next, we visualize all considered model decisions while reconstructing conversational text data to enable analysts to understand the model's decision space. This workflow is designed for machine learning experts and uses an intelligible model. Based on live updates, the experts can investigate the model and see the impact of their interactions before applying them using speculative execution.

Lastly, for domain experts, I introduce Semantic Concept Spaces, a workflow for applying a machine teaching paradigm to capture the experts' knowledge. This series of approaches introduce different workflows but achieve comparable refinement results. This allows machine learning experts and non-experts alike to rely on tailored human-AI interactions and refine topic models for their data and tasks.

Philipp Kellmeyer, University of Freiburg

Title: "Challenges in Human-AI Interaction in Medicine: Interpretability, Fairness and Ethics-by-Design".

Abstract: The enormous success of AI-based applications, especially artificial neural networks for deep learning, promises a substantial transformation of data-driven decision-making in medicine. However, the interaction between humans and AI-based decision support systems creates relevant technical, psychological, ethical, legal, and social challenges that need to be addressed to ensure the responsible use of AI-based decision support in health care. In his talk, Dr. Kellmeyer will highlight some of the most pressing challenges in human-AI interaction such as the interpretability of AI models, fairness and representativeness of data in medicine, and ethics-by-design approaches for medical AI applications.

Massimiliano di Luca, University of Birmingham

Title: "Computational Modelling of Perception and Action".

Abstract: In this talk, I will analyse how the brain processes sensory stimuli supporting perception and effective behaviour. The information we acquire from the environment is continuously varying: we reach out, explore, and interact with objects that can move unpredictably. So we have multiple sensory signals available contemporarily, these signals are dynamic, and the information that they carry is a function of our actions. Despite such variations and complexity, our brain is capable of picking up, combining, and using the information to create a percept and guide our behaviour. It is not entirely clear what are the computations that allow us to effectively process such information and what the properties of the neural mechanisms involved. To shed light on these processes, I use psychophysical experiments and neuroimaging methods to capture neural data and human responses. I then employ signal processing and machine learning to discover patterns in the interaction and user's movements that relate to perception. The leitmotiv of this research is to create computational models that constitute quantitative and testable theories about the underlying cognitive and neural processes. Such models can be used to understand the brain and the mind, for simulations (e.g., to be implemented in robots), rendering (e.g., in haptic devices), and prediction about the user movement, responses, states (e.g., to optimise the generation of sensory cues in VR system by using perceptual metrics).

Melika Payvand, ETH Zurich

Title: "Closing the Gap Between Devices, Circuits and Algorithms Towards Efficient AI Hardware".

Abstract: As Artificial Intelligence (AI) becomes an increasingly integrated part of our daily lives, our digital society is shifting to an era of pervasive specialized "edge-computing" systems for a wide variety of tasks. The stringent memory and power budgets at the edge of the sensors have been fueling the advances in AI accelerators, whose precision reduction and sparsity exploitation techniques are in agreement with the solutions found by the ‘natural intelligent’ systems through millions of years of evolution. Neuromorphic technologies aim to take further inspiration from the organizing principles of biological brains as a natural next step towards low-power intelligent edge devices. The advances in emerging memory technologies augment this opportunity in terms of area and power efficiency. Simultaneously, significant progress is being made with new neuroscientific discoveries and machine learning insights inching us closer to understanding the efficiency of natural intelligent systems. In this talk, I will go over a variety of design methodologies that we have developed for closing the gap between the new trends in devices, circuits, and algorithms, using a holistic approach, which is the foundation towards low-power real-time learning and processing on the edge.

Posters

Aikaterini Aivalioti, Aikaterini Antoniou & Costanza Presi: "Is the Interaction Between AI and Neuroscience Beneficial?"

Matteo Guida & Kornelia Nowinska: "Artificial Intelligence is WEIRD"

Andrea Ferreira: "Can You Trust AI?"

Ragnhild Mogedal & Dustin Richter: "Extended Cognition and Artificial Intelligence"

Organizers

Junior Professor Diego Frassinelli (dept. of Linguistics)

Dr. Caterina Moruzzi (dept. of Philosophy)

Carmen Afilipoaie (dept. of Linguistics)

Aikaterini Aivalioti (dept. of Linguistics)

Aikaterini Antoniou (dept. of Linguistics)

Andrea Ferreira (dept. of Linguistics)

Matteo Guida (dept. of Linguistics)

Ragnhild Mogedal (dept. of Philosophy)

Kornelia Nowinska (dept. of Linguistics)

Costanza Presi (dept. of Psychology)

Dustin Richter (dept. of Education)

Sponsor

This workshop has been funded by the "Transdepartmental Teaching Programm" of the Zukunftkolleg (University of Konstanz, Germany).

Contacts

For information, please contact:

Junior Professor Diego Frassinelli, diego.frassinelli@uni-konstanz.de

Dr. Caterina Moruzzi, caterina.moruzzi@uni-konstanz.de