Dean's Speaker Series Day on Computer Science and Engineering
Guest: İrem Boybat, IBM Research Europe
Title: Analog In-Memory Computing-Based Deep Neural Network Inference Acceleration
Abstract: The advent of deep neural networks (DNNs) has revolutionized numerous fields, including computer vision and natural language processing. These powerful models have showcased remarkable capabilities in solving complex problems, but their training and inference procedures often require significant computational resources. To address this challenge, notable activity was centered around specializing or developing digital hardware accelerators for DNNs, such as graphics processing units (GPUs) and tensor processing units (TPUs). However, this talk will go beyond digital acceleration and instead focus on the emerging field of analog in-memory computing (AIMC). More specifically, it will delve into devices, circuits, and architectures for building AIMC-based inference accelerators for DNNs. AIMC blurs the distinction between memory and processing; however, sustaining energy and area efficiency gains across the system necessitates a holistic approach at hardware and software levels. Recent prototype chips will be presented as demonstrations showcasing a notable level of maturity in the field.
Bio: Dr. Irem Boybat is a Research Scientist at IBM Research Europe, Zurich, Switzerland. She received her Ph.D. degree in Electrical Engineering from Ecole Polytechnique Federale de Lausanne (EPFL), Switzerland, in 2020. Previously, she had obtained an M.Sc. degree in Electrical Engineering from EPFL, Switzerland, in 2015, and a B.Sc. degree in Electronics Engineering from Sabanci University, Turkey, in 2013. Her research is primarily centered around analog in-memory computing for accelerating deep neural networks using phase-change memory devices. She has co-authored over 50 scientific papers in journals and conferences, received four best conference presentation/paper/poster awards and holds 8 granted patents. She was a co-recipient of the 2018 IBM Pat Goldberg Memorial Best Paper Award and 2020 EPFL PhD Thesis Distinction in Electrical Engineering.
--
Guest: Kaan Kara, Oracle
Title: Scaling In-Memory Relational Data Processing to Thousands of Cores
Abstract: Relational data processing and SQL databases continue to be as relevant as ever. In this talk, we will discuss how in-memory database engines scale query processing workloads to thousands of cores. In particular, we will discuss the data partitioning problem from a computer architecture perspective. We will explore ideas about how reconfigurable hardware can be used to accelerate such operations and also how software optimizations can go a long way in helping the scaling efficiency of relational operators such as joins and aggregation to many thousands of cores.
Bio: Kaan is a developer lead for MySQL HeatWave at Oracle, with a focus on query processing. Before joining Oracle, he obtained a PhD at ETH Zurich in Systems Group, part of the Computer Science Department. His research focused on using reconfigurable hardware as a compute acceleration platform for data analytics and machine learning workloads. His work has been published in top database venues such as SIGMOD and VLDB.
--
Guest: Fatma Güney, Koç University
Title: Perception and Prediction in Self-Driving
Abstract: In this talk, I will start by introducing common paradigms to self-driving, namely modular and end-to-end approaches. As examples of our efforts in the perception module, I will talk about our work on 3D scene understanding. We showed that modeling the motion of independently moving vehicles in addition to the ego-motion of the self-driving vehicle not only improves the performance of monocular depth estimation but also jointly produces a segmentation of the scene. In addition to the driving domain, we recently worked on self-supervised learning for multi-object segmentation from generic videos. Previous work showed that the camera motion lies in a low-dimensional subspace given the 3D structure of the scene. Building on that idea, we extended it to moving objects and found that when we model the 3D geometry, we can obtain a more accurate segmentation of the scene. This line of work uses motion for supervision at train time but it may not always be possible to extract accurate motion in unconstrained videos. In parallel work, we developed a method to decompose the scene into objects using spatial and temporal slot attention without any supervision other than the frames in a video. As examples of our efforts in the prediction module, I will talk about future prediction in different representation spaces. In earlier work, we showed that video prediction performance can be improved significantly when we predict the future in motion space in addition to pixel space. Motion history especially helps with repetitive motion patterns such as walking or driving. While predicting frames in pixels, a common problem is that the model needs to predict all aspects of the scene which may not always be relevant to driving, quoting Yann LeCun's example, e.g. leaves moving. Therefore, in our recent work, we focused on predicting the future in coordinate space and Bird's Eye View (BEV) space. The coordinate space is commonly used in trajectory prediction to simplify the problem by assuming perfect perception input to be able to focus on the complex dynamics of the traffic environment and the multi-modality of future predictions. While previous work focuses on predicting the trajectory of a single agent, we extended it to all agents in the scene efficiently with dynamic weight learning. In BEV, we address the multi-modality with stochastic latent variables, leading to diverse predictions. As a new direction, we explored out-of-distribution detection by segmenting unknown objects. While analyzing a commonly used segmentation method, we found that object queries behave like one-vs.-all classifiers, each focusing on a single class. Then, we formulate finding unknowns as identifying pixels that are rejected by all queries.
Bio: Fatma Güney is an Assistant Professor at Koc University in Istanbul. She received her PhD from the Max Planck Institute in Germany. Her research focuses on computer vision problems related to autonomous driving. In the last few years, she published papers on monocular depth estimation, unsupervised object segmentation, and future prediction in different representations. She is a recipient of the ERC Starting Grant as well as prestigious fellowships including the Newton Fund Advanced Fellowship and the Marie Curie Individual Fellowship. She regularly serves as a reviewer with multiple outstanding reviewer awards and more recently as an Area Chair in top-tier Computer Vision conferences.
