Search results

Diffusion models excel at generating photorealistic images, yet their outputs betray a hidden signature: spectral artefacts — distinct anormalies in the frequency domain that deviate from natural image statistics. This webinar will unpack the mathematical underpinnings behind these artefacts, revealing why and how incorrect frequencies emerge during the generation process. We will discuss the extent to which frequency artefacts can be exploited as reliable markers to identify deepfakes, offering new tools for image forensics.

The demand for high sound quality is increasing in both entertainment and communications. Consequently, audio restoration algorithms play a critical role in mitigating distortions and interferences that originate from recording processes or arise from imperfect transmission pipelines. This webinar offers an in-depth examination of generative audio restoration algorithms, with a particular focus on diffusion-based techniques for speech enhancement.

Audio source separation is the task of extracting one or more constituent components, or composites thereof, from their mixture. Creatively produced audio signals, such as music and cinematic audio, present a unique challenge for source separation algorithms due to the sheer diversity of potential sound sources within a particular mixture. However, most state-of-the-art deep learning systems for source separation have often been either a collection of single-source separators or a tightly coupled system that cannot be easily adapted to support additional or unseen sound sources.

In this talk, the presenter will discuss the progress made over the last decade for the implementation of clinical AI solutions for digital pathology. She will draw upon both academic and industrial experiences to discuss the various stages of growth of AI for digital pathology, including the challenges of the past, where we are now, and some perspectives for the future.

Deep Neural Networks (DNNs) have had tremendous positive impact on performance in several tasks in image processing (e.g. image segmentation). In this seminar, I will present works done in my team for improving the performance of DNNs, and how efficient solutions based on DNN combined with Graphs, can be designed in applications such as automatic discovery and geotagging of objects.

In the rapidly evolving field of computational imaging, the coherent codesign of image sensors and signal/image processing unlocks numerous new capabilities. To retrieve sparse macroscopic impulse response functions (IRF) to modulated light we integrate advanced circuit design with signal processing techniques. We introduce a novel sensor design that enables the realization of a single-shot Fourier camera, alongside computationally efficient methods for sparse IRF recovery.

Graph Fourier transform (GFT) is one of the fundamental tools in graph signal processing to decompose graph signals into different frequency components and effectively represent graph signals with strong correlations using various modes of variation. The GFT on undirected graphs has been well-studied, with a conventional approach based on the eigen decomposition of the graph Laplacian. However, this method does not apply to directed graph settings.

In wideband integrated sensing and communications (ISAC) systems, conventional designs optimize dual-functional waveforms over the entire bandwidth. This approach significantly restricts beamforming flexibility and degrades communications performance under strict radar sensing constraints. This webinar presents a novel subcarrier allocation strategy to enhance the communications–sensing performance tradeoff in ISAC. The proposed method selectively assigns a subset of subcarriers for radar sensing while enabling communication across the full bandwidth.

Deep Learning (DL) methods currently obtain state-of-the-art performance in the tasks of positional Sound Source Localization (SSL) and acoustic Direction of Arrival (DOA) estimation. However, most DL methods require matched microphone array geometries between training and testing scenarios, requiring separate models to be trained for different devices. In this webinar, the presenter will present geometry-aware and geometry-agnostic DL approaches for SSL, comparing their advantages and drawbacks, and future research directions.

Section: Events

Date: 11-February-2026 Time: 10:00 AM ET (New York Time) Presenter: Dr. Hung-yi Lee Website link: Register