Add to Calendar
2024-05-03 14:00:00
2024-05-03 15:30:00
America/New_York
Thesis Defense: Effective and Flexible Acceleration of Sparse Computations
Abstract: Sparsity is ubiquitous and abundant in many important application domains including deep neural networks (DNNs), big data analytics and scientific computing. Leveraging sparsity in these applications is a promising way to gain more efficiency in performance, energy and resource utilization when designing hardware accelerators. However, exploiting sparsity in hardware accelerators both effectively and flexibly is challenging. First, sparse computations often involve data with a wide range of sparsity ratios. Second, sparse computation demands a wide range of sparse data representations. And supporting the wide range of sparsity ratio and data representation efficiently in hardware is challenging. Third, the choice of sparse data representations limits the choice of efficient dataflows (compute schedules). Finally, sparsity causes dynamism which manifests as large variations of work and variable access latencies.To address these challenges, this thesis proposes codesigning data representation, dataflow, and hardware architecture to exploit sparsity effectively and flexibly. Leveraging this codesign approach, we propose three architectural techniques to accelerate important sparse applications including graph analytics, linear algebra and DNNs. SpZip is an architectural technique that accelerates a wide range of sparse data traversal and exploits data compression to address the data movement bottleneck of graph analytics and sparse linear algebra. ISOSceles is a hardware accelerator that dramatically reduces data movement of sparse CNN through inter-layer pipelining (fine-grain layer fusion). Trapezoid is a unified architecture that accelerates matrix multiplication effectively with a wide range of input sparsity ratio. All three designs achieve significant performance improvements over state-of-the-art accelerators for sparse computations.Zoom link: https://mit.zoom.us/j/97088952009
Seminar Room G882 (Hewlett Room)
Events
May 03, 2024
May 06, 2024
Thesis Defense: Towards Object-based SLAM
Yihao Zhang
MIT MechE
Add to Calendar
2024-05-06 10:00:00
2024-05-06 11:30:00
America/New_York
Thesis Defense: Towards Object-based SLAM
Abstract:Simultaneous localization and mapping (SLAM) is a fundamental capability for a robot to perceive its surrounding environment. The research area has developed for more than two decades from the original sparse landmark-based SLAM to dense SLAM, and now there is a demand for semantic understanding of the environment beyond pure geometric understanding. This thesis makes a number of contributions to help realize object-based SLAM, in which the map consists of a set of objects with their semantic categories recognized and their poses and shapes estimated. Such a map provides vital object-level semantic and geometric perception for applications such as augmented reality (AR), mixed reality (MR), mobile manipulation, and autonomous driving.In order to perform object-based SLAM, the sensor measurements have to undergo a series of processes. First, objects are semantically segmented in the sensor measurements. This step is typically done by a neural network. As robots are often required to bootstrap from some initial labeled datasets and adapt to different environments where labeled data are unavailable, it is important to enable semi-supervised learning to improve the robot performance with the unlabeled data collected by the robot itself. Second, after the objects are segmented, measurements for each object across different viewpoints have to be associated together for downstream processing. Lastly, the robot must be able to extract the object pose and shape information from the measurements without access to the detailed CAD models of the objects. This thesis studies these three aspects of object-based SLAM, namely semi-supervised learning of semantic segmentation in a robotics context, data association for object-based SLAM, and category-level object pose and shape estimation.For category-level object pose and shape estimation, we developed ShapeICP (ICP: iterative closest point), an algorithm that does not use pose-annotated data and generates meshes as the object shape representation. For data association, we developed DAF-SLAM (DAF: data association free) to estimate the associations in the back-end instead of relying on sensor-dependent front-end methods. For semi-supervised learning, we applied temporal semantic consistency inspired by the photometric consistency technique in the traditional SLAM methods. Each contribution is evaluated via experimental datasets, demonstrating improvements over previous techniques.Committee Members:John J. Leonard (Advisor), Department of Mechanical EngineeringFaez Ahmed, Department of Mechanical EngineeringNicholas Roy, Department of Aeronautics and Astronautics
32-G882 (https://mit.zoom.us/j/92202523862)
May 09, 2024
THESIS DEFENSE: On Physics-Inspired Generative Models
Yilun Xu
MIT-CSAIL
Add to Calendar
2024-05-09 17:00:00
2024-05-09 18:00:00
America/New_York
THESIS DEFENSE: On Physics-Inspired Generative Models
Abstract: Physics-inspired generative models such as diffusion models constitute a powerful family of generative models. The advantages of models in this family come from relatively stable training process and high capacity. A number of possible improvements remain possible. In this talk, I will discuss the design and enhancement of physics-inspired generative models. I will first introduce Poisson Flow Generative Models (PFGM), a new generative model arising from electrostatic theory, rivaling leading diffusion models. The extended version, PFGM++, places diffusion models and PFGM under the same framework and introduces new, better models. Secondly, I will present a sampling algorithm that combines the best of previous samplers, greatly accelerating the generation speed of text-to-image Stable Diffusion models. Finally, I will discuss a training framework that introduces learnable discrete latents into continuous diffusion models. These latents simplify complex noise-to-data mappings and reduce the curvature of generative trajectories. Several algorithms discussed in the talk are the state-of-the-art methods across standard benchmarks.Committee Members: Tommi Jaakkola (advisor, MIT), Phillip Isola (MIT), Karsten Kreis (NVIDIA)
32-D463 (Star). Zoom Link: https://mit.zoom.us/j/8869131067