Keynote Speakers
The following speakers have graciously accepted to give keynotes at AIST-2023.
Samuel Horvath

Towards Real-World Federated Learning: Addressing Client Heterogeneity and Model Size
Abstract: In this talk, I will introduce federated learning and discuss two recent approaches for addressing the challenges of client heterogeneity and model size in federated learning.
In the first part of the talk, I will introduce federated learning. I will discuss the motivation for federated learning, the key challenges, and some of the existing approaches.
In the second part of the talk, I will discuss the FjORD framework. FjORD is a framework for addressing the problem of client heterogeneity in federated learning. FjORD uses Ordered Dropout to gradually prune the model width without retraining, enabling clients with different capabilities to participate by tailoring the model width to the client’s capabilities.
In the third part of the talk, I will discuss the Maestro framework. Maestro is a framework for addressing the problem of model size in federated learning. Maestro uses a technique called trainable low-rank layers to compress the model without sacrificing accuracy.
I will conclude the talk by discussing the future of federated learning.
Hakim Hacid

Towards Edge AI: Principles, current state, and perspectives
Abstract: The artificial intelligence (AI) community has invested heavily in developing techniques that can digest very large amounts of data to extract valuable information and knowledge. Most techniques, particularly deep learning models, require large amounts of computing and storage power, making them suitable for cloud-based environments. The intelligence is therefore remote from the end user, raising concerns about, for example, data privacy and latency. Edge AI addresses some of the problems inherent in the cloud and focuses on best practices, architectures and processes for extending data AI outside the cloud. Edge AI brings AI closer to the end user and uses, for example, fewer communication resources, as processing is performed directly on the edge device. This presentation will introduce edge AI and give an overview of existing work and potential future contributions.
Artem Shelmanov

Safety of Deploying NLP Models: Uncertainty Quantification of Generative LLMs
Abstract: When deploying a machine learning (ML) model in practice, care should be taken to look beyond prediction performance metrics such as accuracy or F1. We should ensure also that it safe to use ML-based applications. This entails that applications should be evaluated along other critical dimensions such as reliability and fairness. The widespread deployment of large language models (LLMs) has made ML-based applications even more vulnerable to risks of causing various forms of harm to users. While streamline research effort has been devoted to the “alignment” via various forms of fine-tuning and to fact checking of the generated output, in this talk, we focus on uncertainty quantification as an effective approach to another important problem of LLMs. Models often “hallucinate”, i.e., fabricate facts without providing users an apparent means to discern the veracity of their statements. Uncertainty estimation (UE) methods could be used to detect unreliable generations unlocking the safer and more responsible use of LLMs in practice. UE methods for generative LLMs are a subject of bleeding-edge research, which is currently quite scarce and scattered. We systemize these efforts, discuss common caveats, and provide suggestions for the development of novel techniques in this area.
Narine Sarvazyan

Decoding Hyperspectral Imaging: From Basic Principles to Medical Applications
Abstract: Over the past few decades, the application of hyperspectral imaging (HSI) has significantly expanded, finding widespread use in areas such as satellite imaging, agriculture, the food industry, and medicine. What sets HSI apart is its capacity to acquire complete spectral data from every pixel of an image. Each HSI dataset is a collection of individual images across numerous spectral bands and/or varied lighting conditions. A distinctive element of HSI is that, unlike grayscale or color images – where each pixel contains one or three to four color channels respectively – HSI captures hundreds of spectral bands for every pixel. Hence, the output from HSI is essentially a three- or four-dimensional dataset, with two dimensions representing spatial axes and the rest providing spectral values. Each dimension typically encompasses hundreds of individual values, so the massive amount of information collected by HSI hardware presents a great opportunity to apply ML and AI tools for data analysis. During this keynote presentation, the speaker will overview the fundamental principles of HSI technology including examples from her own projects. The goal will be to illustrate both the immense promise of HSI in revealing previously unseen surgical targets, as well as challenges posed by the high-dimensionality of HSI data. The key considerations for automatic processing and analysis of HSI data for medical use will also be touched upon. These include preserving the original spectral detail of an image to prevent the loss of information, ensuring processing efficiency for real-time application in a clinical environment, and managing the demands on processing power to ensure broader implementation.