Speaker: Dr. Yan Yan
From: Illinois Institute of Technology
Abstract
In this talk, we will discuss the efficient cross-modal generation and neural tracing. (i) Diffusion probabilistic models (DPMs) have become a popular approach to conditional generation, due to their promising results and support for cross-modal synthesis. A key desideratum in conditional synthesis is to achieve high correspondence between the conditioning input and generated output. Most existing methods learn such relationships implicitly, by incorporating the prior into the variational lower bound. In this talk, we will take a different route—we explicitly enhance input-output connections by maximizing their mutual information. We demonstrate the efficacy of our approach in evaluations with dance-to-music generation. (ii) The generation process of current DPMs is notoriously slow due to the lengthy iterative noise estimations, which rely on cumbersome neural networks. It prevents the diffusion models from being widely deployed, especially on edge devices. Previous works accelerate the generation process of DPMs via finding shorter yet effective sampling trajectories. However, they overlook the cost of noise estimation with a heavy network in every iteration. In this talk, we will talk about accelerating generation from the perspective of compressing the noise estimation network. Due to the difficulty of retraining DMs, we exclude mainstream training-aware compression paradigms and introduce post-training quantization (PTQ) into DPMs acceleration. (iii) Neuron tracing is to reconstruct neuron anatomy in the 3D image stack, i.e., capturing the trajectory of neurons traversing in a 3D volume, which is an essential step in neuroscience. In this talk, we will discuss a novel methodology for multi-spectral neuron tracing from the perspective of online object tracking with uniquely designed modules. Our method is efficient training-free, where it learns online to adaptively update an enhanced discriminative correlation filter to conglutinate the tracing given only a starting bounding box. This distinctive training-free schema differentiates us from other training-dependent tracing approaches since we do not need annotated data.
For more info, please follow this link.
Read More