倒计时1天 | 北京大学图灵学生科研论坛精彩预告

关键词：图灵学生科研论坛

01

论坛形式

本次学术论坛将于5月25日（星期六）在线下举行。

上午将举办亮点报告（lightning talk）和海报展览（poster session）环节。亮点报告中每位讲者限时8-10分钟进行报告（含问答环节）。每半场结束后，由图灵班科研导师进行点评，全部报告结束后，设听众投票环节。亮点报告结束后将进行海报展示，并由观众投票选出最佳海报。展示结束后进行颁奖。 

下午将开展不同主题的学术教程（tutorial）和围绕图灵班学生科研发展的圆桌论坛（panel）。学术教程中每位讲者限时25分钟进行报告，之后设有5分钟现场问答。学术教程结束后进行颁奖，随后进入圆桌论坛环节。

02

论坛日程

03

亮点报告概览

完整版请见秩序册，点文末“阅读原文”下载。

01

MaskClustering for Open-Vocabulary 3D Instance Segmentation

Mi Yan

Turing Class 2019 (U), 2023 (PhD)

Abstract:

Open-vocabulary 3D instance segmentation is cutting-edge for its ability to segment 3D instances without predefined categories. However, progress in 3D lags behind its 2D counterpart due to limited annotated 3D data. To address this, recent works first generate 2D open-vocabulary masks through 2D models and then merge them into 3D instances based on metrics calculated between two neighboring frames. In contrast to these local metrics, we propose a novel metric, view consensus rate, to enhance the utilization of multi-view observations. 

02

ProgressGym: Progress Alignment with Evolving Human Values

Tianyi Qiu

Turing Class 2022

Abstract:

Frontier AI systems, including large language models (LLMs), hold increasing influence over the beliefs and values of human users. Such influence reinforces the beliefs and values currently popular in society, contributing to the potential lock-in of societal values. Harms from the lock-in include the perpetuation of our currently ill-informed moral views, and consequently the perpetuation of problematic moral practices on a societal scale. We introduce progress alignment as a technical solution to this imminent risk, which is a new class of alignment techniques that learn to implement the mechanics of human moral progress. 

03

DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving

Shengyu Liu

Turing Class 2021

Abstract:

DistServe improves the performance of large language models (LLMs) serving by disaggregating the prefill and decoding computation. Existing LLM serving systems colocate the two phases and batch the computation of prefill and decoding across all users and requests. We find that this strategy not only leads to strong prefill-decoding interferences but also couples the resource allocation and parallelism plans for both phases, leading to over-provision compute resources to meet strict service level agreement (SLA)s. DistServe assigns prefill and decoding computation to different GPUs, hence eliminating prefill-decoding interferences. Given the application's SLA requirements, DistServe co-optimizes the resource allocation and parallelism strategy tailored for each phase. 

04

Understanding Adversarial Examples and Adversarial Training in Deep Learning: A Feature Learning View

Binghui Li

Turing Class 2019 (U), 2023 (PhD)

Abstract:

Adversarial training is a widely-applied approach to training deep neural networks to be robust against adversarial perturbation. However, although adversarial training has achieved empirical success in practice, it still remains unclear why adversarial examples exist and how adversarial training methods improve model robustness. In this paper, we provide a theoretical understanding of adversarial examples and adversarial training algorithms from the perspective of feature learning theory. 

05

Omni6DPose: A Benchmark and Model for Universal 6D Object Pose Estimation and Tracking

Weiyao Huang

Turing Class 2022

Abstract:

6D Object Pose Estimation is a critical yet challenging task in the field of computer vision, distinguished from more traditional 2D tasks by its lack of large-scale annotated datasets. To address this issue, we introduce Omni6DPose, a dataset characterized by its diversity in object categories, large scale, and variety in object materials. 

06

The Surprising Benefits of Base Rate Neglect in Robust Aggregation

Ying Wang

Turing Class 2019 (U), 2023 (PhD)

Abstract:

Robust aggregation integrates predictions from multiple experts without knowledge of the experts' information structures. Prior work assumes experts are Bayesian, providing predictions as perfect posteriors based on their signals. However, real-world experts often deviate systematically from Bayesian reasoning. Our work considers experts who tend to ignore the base rate. We find that a certain degree of base rate neglect helps with robust forecast aggregation.

07

Quantum Error Mitigation beyond Direct Expectation Value Estimation

Kecheng Liu

Turing Class 2021

Abstract:

Quantum error mitigation (QEM) has been pivotal in extending the computation reach of noisy quantum machines, but it is studied mostly for the task of direct expectation value estimation and a general theoretical framework for it is still missing. Our work explores how most of the other practical quantum algorithms, including single-shot algorithms with verifiable results like Shor's algorithm and those with non-verifiable results like quantum phase estimation, can all be transformed into a form whose errors can be suppressed byneiron quantum error mitigation, significantly expanding the range of applicability of quantum error mitigation.  

08

Robust Decision Aggregation with Second-order Information

Yuqi Pan

Turing Class 2020

Abstract:

We consider a decision aggregation problem with two experts who each make a binary recommendation after observing a private signal about an unknown binary world state. An agent, who does not know the joint information structure between signals and states, sees the experts' recommendations and aims to match the action with the true state.

09

Generating Lemma Automatically through Declarative Proof

Yi Fang

Turing Class 2021

Abstract:

Separation logic is an extension of Hoare logic, designed to describe memory storage in a more concise manner. It has become widely used for proving the correctness of imperative programs. A key component of proving assertions in separation logic involves proving separation entailments. In the classical style of proving separation entailments, defining & applying lemmas is the main challenge. Our work introduces a domain specific language (DSL) capturing the idea of ramify rule, in which user can write transformation of separation entailments freely with soundness guarantee.

10

Ad vs Organic: Revisiting Incentive Compatible Mechanism Design in E-commerce Platforms

Ningyuan Li

Turing Class 2019 (U), 2023 (PhD)

Abstract:

On typical e-commerce platforms, a product can be displayed to users in two possible forms, as an ad item or an organic item. Usually, ad and organic items are separately selected by the advertising system and recommendation system, and then combined by a content merging mechanism. Although the design of the content merging mechanism has been extensively studied, little attention has been given to a crucial situation where there is an overlap between candidate ad and organic items. Despite its common occurrence, this situation is not correctly handled by almost all existing works, potentially leading to incentive problems for advertisers and the violation of economic constraints. To address these issues, we revisit the design of the content merging mechanism.

04

教程概览

完整版请见秩序册，点文末“阅读原文”下载。

01

Eliciting High-Quality Information from Strategic Agents

Shengwei Xu

Turing Class 2017

Abstract:

In this tutorial, we will explore mechanisms to elicit high-quality information from strategic agents. We will begin with an introduction to proper scoring rules, which are fundamental for incentivizing truthful reporting. Following this, we will delve into peer prediction mechanisms that leverage peer reports to elicit subjective information, in which case there's no ground truth. The tutorial will also highlight recent advancements in eliciting information using large language models (LLMs). Finally, we will discuss the critical connection between information elicitation mechanisms and machine learning loss function.

02

High Performance Computing in Chip Design

Zizheng Guo

Turing Class 2018 (U), 2022 (PhD)

Abstract:

This talk introduces you to electronics design automation, which is the field of algorithms and software for integrated circuits design. Specifically, this talk focuses on an interdisciplinary area between EDA and high-performance computing. Modern chip designs are very large, requiring the solution of gigantic graph-based and AI-based problems. This process takes weeks to even months which slows down the time to market. HPC techniques like parallel and heterogeneous CPU/GPU computing have thus been introduced to solve EDA problems with order-of-magnitudes better efficiency. This talk reviews the frontier HPC applications in major EDA problems including logic gate placement, signal routing, static timing analysis, and more. We hope a better joint effort between the HPC and EDA fields can lead to better design quality and speed, and thus stronger computer chips.

03

Quantum Error Correction Codes: a Classical View

Huiping Lin

Turing Class 2018

Abstract:

Quantum error correction codes aims to protect quantum information from quantum noise. This talk will introduce quantum codes using a classical view, hoping to bring some insight of quantum codes for researchers in both quantum and classical computing fields.

04

Value Alignment: History, Frontiers, and Open Problems

Tianyi Qiu

Turing Class 2022

Abstract:

This session focuses on the critical issue of value alignment in AI, including within large language models, to ensure their long-term impact aligns with ethical and societal norms. As these technologies increasingly influence societal values and human decision-making, understanding and directing the intrinsic values they endorse is pivotal. We'll dive into the historical and theoretical foundations of machine ethics and the cutting-edge research on how language models can contribute to preference aggregation and moral progress. We will have a dual emphasis on both practical engineering pipelines (some of which already adopted by Anthropic and OpenAI) and open research questions in the field of value alignment. This discussion aims to articulate the complexities and the urgency of developing AI systems that not only adhere to but advance collective human values.

05

"AC" Automaton -- The Road to Automatic Algorithm Design

Ruyi Ji

Turing Class 2016 (U), 2020 (PhD)

Abstract:

Algorithm design is a fundamental way to achieve efficiency in practical software development but is also known as difficult and risky. Hence, the automatic design of algorithms has been a long dream of researchers. Many research efforts have been devoted to this target in the past decades, with many interesting techniques developed. This tutorial will review the progress of automatic algorithm design, introduce the key challenges and representative approaches, and show the remaining problems for future work.

05

论坛组织

论坛由北京大学信息科学技术学院、计算机学院联合主办，北京大学前沿计算研究中心承办。

论坛主席

Kecheng Liu

Turing Class 2021

06

论坛鸣谢

感谢北京大学 John Hopcroft 基金对本届论坛的大力支持。

感谢华为赞助本次活动奖品！

第三届北京大学图灵学生科研论坛

2024年5月25日，相约北京大学静园五院

↓↓会议秩序册↓↓

或点文末“阅读原文”跳转

北京大学图灵班

往届活动回顾

—   版权声明  —

本微信公众号所有内容，由北京大学前沿计算研究中心微信自身创作、收集的文字、图片和音视频资料，版权属北京大学前沿计算研究中心微信所有；从公开渠道收集、整理及授权转载的文字、图片和音视频资料，版权属原作者。本公众号内容原作者如不愿意在本号刊登内容，请及时通知本号，予以删除。

点“阅读原文”跳转论坛秩序册