R4年度 人工知能研究センター 国際シンポジウム
AIST Artificial Intelligence Research Center International Symposium 2023

Trust in Human, Trust in AI Trust is indispensable in successful collaboration and cooperation between humans as well as in the constitution of societies. Rapid advance in AI technology heightens the need to both examine the concept of trust itself and to ensure people's trust in AI. Social psychological research based on Prisoner's dilemma situation concluded that trust is a kind of non-rational cognitive bias, at individual level, underlying human decision making. I will review types of conceptions behind common-sense notion of trust, emphasize the role of dialogue interaction in establishing and maintaining trust, and contrast trust in human and trust in AI.

From Graphic System Interactions to Embodied Social Experience in XR Shared Spaces Augmented Reality has developed greatly over the last 20 years. Nevertheless, current applications are often still limited to simple overlays such as 2D annotation or 3D objects inserted into the scene. However, it has long been clear that automated VR/AR-systems designed to assist or interact with human activity must have some understanding of human behaviour to be effective. Actions and reactions need to match our expectations and information needs to be presented in a way that reflects our own perceptions. Less well understood is how this understanding of behaviour is to be achieved. In this talk, we will introduce different technologies for human workflow capturing and adaptive workflow monitoring with AR. Then, a new concept for seamless remote collaboration in an XR space shared by humans and avatars will be presented and discussed.

The integration of GeoAI and streaming data in urban flood response: A real case study in Taoyuan City In recent years, many rivers have established water level stations to detect the water level. When a certain water level is reached, warning information will be sent to relevant units for subsequent measures such as bridge and road closures. However, with the recognition ability of AI images improved significantly, this research aims to monitor water level by image recognition through CCTV images which can calculate the water level height more intuitively and further replace the original radar water level gauge or pressure water level gauge. This study uses the current image segmentation of the image deep learning to segment the image of the water surface. Then it uses the virtual water ruler, which the user draws to obtain the water level height by converting the water level height at the intersection of the water surface and the virtual water ruler. In response to the concept of this study, it uses the Deeplab V3 algorithm provided by Google to obtain 500 images in the demonstration area in the morning, noon, and evening. It recognizes the image of the water level by every minute.

Application of Privacy Preserving Federated Learning in Biomedical Applications – Lessons Learned AI/ML models are known to be vulnerable to dataset shift and under specification because of the inability of current deep learning methods to learn the causal structure. This problem manifests when a model is deployed in a real test domain, where even simple changes in demographics or image formats could lead to unexpected poor performance, straining credibility. The solution for this is to train deep learning models on as many as real world datasets as possible. But access to biomedical datasets is governed by complex data usage agreements, time-consuming IRBs. One possible solution is to send models to data and develop frameworks to perform secure federated learning with privacy guarantees. In this talk I will present work done under DOE-ASCR funded PALISADE-X project where we developed Secure Computational techniques on Private enclaves, Argonne Privacy Preserving Federated Learning Framework (APPFL) and other strategies to solve challenges in biomedicine.

Towards Differentially Private Federated Learning with Untrusted Server Federated learning has received increasing attention in academia and industry as a new privacy-preserving machine learning paradigm. Unlike traditional machine learning, which requires data collection before training, in federated learning, the clients collaboratively train a model under the coordination of a central server. In particular, the clients only share model updates to the server, and all raw data are stored locally. However, recent studies showed that the model updates might reveal sensitive information to the server. In addition, federated learning itself does not guarantee formal privacy. This talk will review recent advances on differentially private federated learning under untrusted servers, introduce our attempts towards this goal by leveraging LDP, the shuffle model of DP and TEE, and discuss some open problems.

Assessment of AI-systems Artificial intelligence is penetrating more and more areas of our lives, taking on increasingly responsible activities. At the same time, it is to be expected that this technology of the future can only develop its full potential if sufficient trust can be created in its use. AI assessments (by independent parties) can be an appropriate measure of promoting this trust. Such assessments are partly required by the currently evolving regulatory landscape of AI. They can also be beneficial on a voluntary basis, for example as a basis for insuring AI systems or for establishing a strong brand. At the same time, AI systems differ greatly from classically programmed software, so that established procedures for testing software fall short of systematically proving that AI systems meet the requirements demanded of them. The talk starts with a short overview of the planned European regulation of AI systems and then presents an AI assessment catalog that was developed at Fraunhofer IAIS and addresses important challenges in the evaluation of AI systems. Finally, some practical examples of AI assessments are presented which have been carried out based on the assessment catalog.

Quantum Computation and Artificial Intelligence – From the Lab to the Market In this talk I shall define the notion of quantum computing and its constituent parts. Moreover, I shall provide a succinct introduction to the history of quantum computing followed by a concise review of the mathematical and computational foundations of this discipline. I will then talk about quantum algorithms and present several results of my research group with an emphasis on the multidisciplinary nature of quantum computing and two key areas in this field: artificial intelligence and cybersecurity. I will finish by briefly addressing some key features of the emerging high-tech market of quantum technologies.

Location & Privacy: Past, Present and Future In this talk, I will review various types of mobile applications, from Location-Based-Services (LBS) and Ride-Sharing to pandemic risk-prediction applications, which collect, analyze and use location data to provide various services to users. I explain some of the main underlying technologies (e.g., kNN queries and contact-network analysis) that enable these applications from a spatial data management and analysis (aka GeoAI) perspective. I will also discuss some of the privacy concerns with these applications due to their location leaks and review several approaches to protect location privacy without sacrificing the utility of these applications. I will wrap up by presenting some new envisioned applications and their corresponding open problems.

GeoAI Applications: from land use detection to agricultural yield predictions Geospatial Artificial Intelligence sees a rapidly increasing amount and diversity of applications. Many of these include the automatic analysis of satellite imagery. While there are huge labelled datasets for traditional image analysis, we usually have to work with limited amounts of labelled data for satellite imagery. In this talk, we present some of our GeoAI applications such as agricultural yield predictions, land use land cover (LULC) classification and flood mapping. We highlight different approaches to making use of the vast amount of geospatial input data sources, while being limited in the amount of labelled data accessible.

Pre-training without Natural Images: Introduction of Formula-driven Supervised Learning Is it possible to use convolutional neural networks pre-trained without any natural images to assist natural image understanding? The presentation introduces a novel concept, Formula-driven Supervised Learning (FDSL). We automatically generate image patterns and their category labels by assigning fractals, which are based on a natural law. Although models pre-trained with the proposed Fractal DataBase (FractalDB), a database without natural images, do not necessarily outperform models pre-trained with human annotated datasets in all settings, we are able to partially surpass the accuracy of ImageNet/Places pre-trained models.

Large-scale Pre-training of Vision Transformers on Synthetic Datasets The true potential of vision transformers (and transformers in general) is realized only when pre-trained on extremely large datasets. Recent advances in self-supervised learning on open source text + image pairs such as LAION have solved the issue of labeling such large datasets. However, at such scale, it is infeasible to manually remove ethical issues in the dataset such as societal bias, copyright infringement, and privacy. In the present work, we scaled up formula-driven supervised learning (FDSL) to investigate the possibility of using synthetic datasets to pre-train vision transformers at large scale. We show that ViT-Base pre-trained on synthetic datasets can outperform ImageNet-21k on multiple downstream tasks.

Variational Deep Learning for Integration of Disjoint Longitudinal Patient Data Patient data of cellular measurements often comes from heterogeneous sources with many types of technical variability. Our method is a flexible variational deep learning framework for integration of disjoint cytometry data from large cohorts of patients, originating from multiple sources and acquisition batches. We benchmark various types of conditional modelling and data normalization, and subsequently apply our method to integrate longitudinal flow cytometry data collected from Covid-19 patents admitted to four hospitals in Manchester, UK.

Closed workshop invited speaker

Closed workshop invited speaker

Closed workshop invited speaker

Closed workshop invited speaker