Graphs have long been used in various problems, such as analyzing social networks, machine learning, network protocol optimization, or image processing. In the last few years, a growing body of work has been developed to extend and complement well-known concepts in spectral graph theory, leading to the emergence of Graph Signal Processing (GSP) as a broad research field. In this talk, we summarize recent results that lead to a GSP perspective of machine learning problems. The main observation is that representations of sample data points (e.g., images in a training set) can be used to construct graphs, with nodes representing samples, label information resulting in graph signals, and edge weights capturing the relative positions of samples in feature space. We will first review how this perspective has been used in well-known techniques for label propagation and semi-supervised learning. Then, we will introduce the non-negative kernel regression (NNK) graph construction, describe its properties, and introduce example applications in machine learning areas such as i) model explainability, ii) local interpolative classification, and iii) self-supervised learning.

Image alignment/registration is an essential step in applications where two or more images are compared, factoring out the possible geometric distortions between them (e.g., medical imaging, remote sensing, computer vision). We present a novel algorithmic framework for estimating such geometric transformations, by re-interpreting their local effect---shifting---as a signal processing operation: a convolution with an all-pass filter. The key advantage of this reformulation is that no restrictions are placed on the amplitude of the deformation or on the spatial variations of the images. Moreover, by identifying all-pass filtering with a forward-backward filtering relation, our solution to the estimation problem boils down to solving a linear system of equations, which leads to a highly efficient implementation. The effectiveness of this algorithm is demonstrated on a variety of synthetic and real deformations that are found in applications such as image registration and motion estimation. In particular, the Local All-Pass (LAP) algorithm obtains very accurate results for significantly reduced computation time when compared to a selection of image registration algorithms and is very robust to noise corruption. When the transformation has a parametric representation, this robustness can be exploited to deal with large intensity differences between the two images.

The required data size and computation throughput of the recently proposed transformer models rapidly increase, thus causing memory bottlenecks to determine the overall system performance. High bandwidth memory (HBM) has been popularly used with GPUs and accelerators to overcome the bottlenecks. However, data transfer from memory to computing devices is still problematic from the viewpoints of computation throughput and power consumption. One of the challenging approaches to resolving the problem uses the "Processing-in-Memory (PIM)" device, especially DRAM, by designing a processing unit inside the device, performing the computation near memory banks, exploiting internal bandwidth, and eliminating the data transfer to a memory controller. Major DRAM makers have introduced the PIM samples, executing all banks' computations simultaneously to maximize the internal DRAM bandwidth to achieve the highest performance. However, their commercialization is still problematic since we must reconsider the interaction between all the architectural layers, from programming models to memory devices, and we would require system component modifications. In this talk, we first review the PIM performance opportunities in the recent transformer models. Then, we discuss the PIM design issues in all the architectural layers and propose possible solutions from recent research targeting PIM commercialization.

Wasserstein distance, which measures the discrepancy between distributions, shows efficacy in various types of natural language processing (NLP) and computer vision (CV) applications. One of the challenges in estimating Wasserstein distance is that it is computationally expensive and does not scale well for many distribution comparison tasks. In this talk, I propose a learning-based approach to approximate the 1-Wasserstein distance with trees. Then, I demonstrate that the proposed approach can accurately approximate the original 1-Wasserstein distance for NLP tasks. Moreover, I introduce the problem of self-supervised learning (SSL) utilizing the 1-Wasserstein distance on a tree structure (a.k.a., Tree-Wasserstein distance (TWD)).

The metaverse is a blend of “meta” which means "beyond" in Greek and “universe”, first coined in 1992 science fiction novel Snow Crash, which describes a vast virtual reality network where real people use their avatars to interact with each other in the virtual world. In October 2021, the term metaverse has become globally important and received significant attention when Mark Zuckerberg announced the rebranding of his Facebook to Meta and made commitment to create the metaverse. The metaverse has many potentials span across various sectors, including entertainment and gaming, education and training, real estate and architecture, healthcare and telemedicine, retail and E-commerce, workforce collaboration, events and conferences, art and creativity, social networking and communication, tourism and hospitality. These are just a few examples of sectors where the metaverse can be applied, but the possibilities are vast and continually expanding as technology advances and innovation continues. Ultimately, the success of metaverse applications depends on their ability to address specific industry needs, deliver value to users, and drive adoption in the marketplace. In this presentation, we introduce our ongoing metaverse activities in the education sector, referred to as MANGOs (Metaverse of Academic Nexus for Global Opportunities), which may shed some light on the true potential of the metaverse.