A Decade of Progress in Sound Event Localization and Detection: Transforming Environmental Sound Analysis for Real-World Impact

Abstract

In this talk, I will outline the historical progression of computational Auditory Scene Analysis, with a particular emphasis on the sound event detection and localization (SELD) task. SELD aims to enhance the human auditory system by automating extended environmental listening, with applications spanning noise monitoring, surveillance, real-sound control for mobile devices, and biodiversity tracking. Over the past decade, substantial progress has been made in SELD, driven in large part by the Detection and Classification of Acoustic Scenes and Events (DCASE) challenge series initiated since 2010. This presentation will spotlight major breakthroughs in SELD, examine recent advancements and emerging research directions, and explore real-world applications. Additionally, I will share some of the latest research findings and proof-of-concept prototypes developed by my research group.

Exploring the Forward-Forward Algorithm: A Novel Learning Approach

Abstract

The Forward-Forward (FF) Algorithm is a recent learning procedure for neural networks based on Boltzmann machines and Noise Contrastive Estimation. This algorithm eliminates both the backpropagation's forward and backward passes and substitutes them with two forward passes, each on different data but with different goals. The positive pass employs actual data to raise “goodness” in the hidden layers, and the negative pass employs “negative data” to lower it.
The FF algorithm is especially useful when specifics of forward computation are unclear so that learning can take place without preserving neural activities or passing error derivatives. It takes nearly the same prediction time to perform as backpropagation and can be applied where reinforcement learning would normally be needed. While it may not be as good as backpropagation for specific problem types, it has prospects for use in low-power analog VLSI and as a model for learning in the cortex.
In this presentation, I will present some of the initial findings of the FF algorithm, apply it to small networks, and establish a basis for its application to larger networks. We will also talk about various measures of “goodness” and the general conclusion about the algorithm’s impact on further analysis of neural networks.

Eye-gaze-based Human-Intention Detection

Abstract

Recently, many robots have been working with humans in many places such as manufacturing factories, hospitals, department stores, airports, construction sites, and so on. Sometimes, accidents occur by robot and human crashes, since the robot has no function to sense human intention, while a man has some senses to estimate the intention of another man by detecting eye gaze. Eye gaze detection becomes an important basic function, which would be extended to detect intention and other applications such as intention reading for stroke patients, customer intention estimation, worker intention estimation for collaborative robots, game entertainment, and so on. This talk will update the technologies related to eye-gaze-based human-intention detection and discuss the trend in future research.

From GPT Evolution to Enterprise Deployment: Key Trends in Generative AI

Abstract

This presentation explores the latest trends in Generative AI, focusing on key topics such as the evolution of GPT, domain-specific optimization in large language models (LLMs), micro-sizing LLMs, and the deployment of Generative AI in Taiwan's Industrial Technology Research Institute (ITRI) and its potential expansion to Taiwanese enterprises. The talk delves into the advancements in GPT models, highlights the significance of expert-level optimization techniques in enhancing LLM performance, and discusses strategies for scaling down large models. Moreover, it showcases the current efforts by ITRI in implementing Generative AI and outlines the future prospects of adopting this technology within Taiwan's business landscape.

An Overview of Online Distributed Kernel Methods for Supervised and Unsupervised Learning

Abstract

This talk discusses nonlinear filtering algorithms using kernel methods. We start by discussing the problem for centralized online kernel methods using Random Fourier Features. We then discuss federated online learning algorithms discussing tradeoffs between performance, computational complexity, and communication costs. Finally, we present some examples for time series prediction and anomaly detection.

An AI-based Diagnostic-aid for Epileptic Electroencephalography

Abstract

This talk addresses an "AI-based diagnosis and treatment-aid" system that learns from electroencephalograms (EEG) of epilepsy patients and diagnoses made by clinicians. Japan has about one million epilepsy patients, but there is a very limited number (about 800) of clinicians who can properly read and interpret patient EEG. Epileptic seizures sometimes cause serious traffic accidents, highlighting the need for effective social measures to address this issue. We have developed AI-based automated algorithms that can learn from diagnoses made by medical specialists in epilepsy for EEG data measured in hospitals. I will report techniques to construct a dataset of EEG and to learn the interpretation of the data.

Machine Learning for Analytics Architecture: AI to Design AI Video

Abstract

Niklaus Emil Wirth introduced the innovative idea that Programming = Algorithm + Data Structure. Inspired by this, we advance the concept to the next level by stating that Design = Algorithm + Architecture. With concurrent exploration of algorithm and architecture entitled Algorithm/Architecture Co-exploration (AAC), this methodology introduces a leading paradigm shift in advanced system design from System-on-a-Chip to Cloud and Edge.
As algorithms with high accuracy become exceedingly more complex and Edge/IoT generated data becomes increasingly bigger, flexible parallel/reconfigurable processing are crucial in the design of efficient signal processing systems having low power. Hence the analysis of algorithms and data for potential computing in parallel, efficient data storage and data transfer is crucial. With extension of AAC for SoC system designs to even more versatile platforms based on analytics architecture, system scope is readily extensible to cognitive cloud and reconfigurable edge computing for multimedia and mobile health, a cross-level-of abstraction topic which will be introduced in this tutorial together with case studies in lightweight mobile edge for skin cancer detection with two layers CNN at 97% recognition rate. High level Synthesis (HLS) may also be presented should time allows.

Compression of Large AI Models

Abstract

With the unprecedented effort by academy and industry recently, large AI models have demonstrated impressive achievement and great potential in diverse areas; however, such models demand rapidly increasing resources in energy, computation and storage. To make them green and economical is the key toward sustainable, responsible and affordable solutions. In this talk, we will discuss into issues and opportunities for pruning and compressing large deep neural networks, in order to improve implementational efficiency, hardware design, power consumption, and so on, without jeopardizing system performance.

Introduction to Multi-Camera Systems and 3D Quality Assessment

Abstract

As the boundaries between Computer Vision and Graphics continue to blur, the use of 3D Human Avatars through Vision technology is gaining increasing attention. With these technological advancements, the ability to capture high-quality 3D data has opened up new possibilities for creating more refined 3D Avatars using Deep Learning. Setting up a camera system plays a key role in this process, offering a more efficient and essential method for acquiring 3D data. In this presentation, I will introduce the camera system setup, discuss 3D Human Avatars, and explore the technologies used to evaluate them. Defining mesh quality features is challenging due to the irregular topology of meshes, which are defined on vertices and triangles. To address this, we propose a novel 3D projective structural similarity index (3D-PSSIM) for meshes that is robust to differences in mesh topology. Additionally, to evaluate the quality of the generated human motion, we introduce various metrics for human pose and motion analysis.

Highlight of New Image Generative Models and Applications to Image Manipulations

Abstract

In this overview, we are going to highlight the essences some recent generative models and give an introduction to our new domain transfer generative model. We subsequently will discuss briefly the ways to make various applications, such as face generation, image inpainting, image/video super-resolution, face manipulations and video coding.

Overview of Source Camera Identification Techniques

Abstract

One of the key problems in image forensics is the source camera identification which aims to uncover the origins of digital photos. Traditionally, signal processing (SP) techniques constructed camera signatures to uniquely characterize capturing devices. With the surge of deep learning methodologies across diverse research fields, these techniques have been used for source camera identification tasks. Architectures such as convolutional neural networks (CNNs) and Residual Networks (ResNets) have been explored. However, studies found that these approaches are effective up to the model level, i.e., the camera model used to capture the photo can be identified. Accurately identifying the specific camera device remains a challenge compared to traditional SP methods.
SP-based and DL approaches require either the source camera to produce a camera fingerprint or substantial photos captured by the device for model training. This requirement posts hurdles in practical real-world scenario. To address this limitation, the source camera identification problem can be reformulated as an image similarity matching problem in which contrastive learning can be used. It offers a promising solution to enhance performance and address practical constraints. In this talk, an overview of source camera identification is given. We will highlight the principles and results of different methods, including traditional SP approaches, deep learning approaches and the recent contrastive learning approaches. Various challenges will also be discussed.

Recent Advances in Complete Quality Preserving Data Hiding

Abstract

Data hiding aims to insert data into content to achieve specific purposes, including claiming ownership through watermarking, validating the integrity of the content in authentication, and providing extra/paid features, to name a few. Data hiding concerns three conflicting requirement aspects, namely, quality, capacity, and robustness. For many years, researchers have aimed to insert as much data as possible while minimizing quality deterioration and maximizing robustness. However, in a typical data hiding method where the content is modified to encode data, the quality of the host is always degraded.
In this overview, the concept of complete quality preservation is introduced, and the need for such feature is discussed. Subsequently, recent techniques designed to hide data into PDF file and animated image while completely preserving their qualities are introduced, followed by some potential research directions and use cases.

Real or Fake? Frontiers of Countering Fake Media in the Age of Infodemics

Abstract

AI can now learn from human-derived information such as the face, voice, body, and natural language to generate synthetic media that can be mistaken for the real thing. Synthetic media are used in a variety of applications in the fields of communication and entertainment. On the other hand, as a negative aspect of synthetic media, there are cases in which fake media, such as fake video, fake audio, and fake text, are generated and distributed for fraud, inducement of thought, and manipulation of public opinion. There is a fear of an "infodemic," a flood of uncertain information that causes anxiety and confusion in society. This talk will give an overview of the threats posed by such fake media and introduce the latest results of the research project "Social Information Technologies to Counter Infodemics" (CREST FakeMedia), which the speakers are currently working on under the Strategic Basic Research Program (CREST) of the Japan Science and Technology Agency (JST). Ref: CREST FakeMedia, http://research.nii.ac.jp/~iechizen/crest-e.html

User Preference Modeling and Analysis in Choice Problems

Abstract

We make decisions from many choices with conflicting attributes every day. Although these decision-making problems are diverse, a user's decision is often influenced by three factors: the inter-competition competition, that is, the competition among available choices; the context effect where a user’s preference for a seller is affected by the market where the seller is in; and the projection bias where users are biased when estimating their future preference. A challenging issue here is to model a user's personal preference and to understand his/her decision process, and this is especially important in the information era, where users are often overwhelmed by the avalanche of information available online. In this talk, we will talk about our recent works on model-based methods to analyze the impact of the above factors on users’ choices in e-commerce. This investigation can help address the information overload problem, offer personalized services, and provide essential guidelines on seller pricing strategies, market demand analysis, etc.