Nankai University International E-Forum on Artificial Intelligence and Robotics

（第68期）

2025年南开大学人工智能与机器人国际学术讲坛

College of Artificial Intelligence, Nankai University

报告时间：2025年11月27日（周四）9:00-10:40

报告嘉宾：Tatsuya Akutsu

zoom会议：411 672 9907(Passcode:1111)

报告题目：On the Compressive Power ofAutoencoders With LinearThreshold Activation Functionsand ReLU Activation Functions

报告摘要：

An autoencoder is a layered neural network whose structure can be viewed as consisting of anencoder, which compresses an input vector to a lower dimensional vector, and a decoder, whichtransforms the low-dimensional vector back to the original input vector (or one that is very similar). in this talk, we explore the compressive power of autoencoders with linear threshold activation functions by studying the numbers of nodes andlavers that are required to ensure that eacvector in a given set of distinct input binary vectors is transformed back to its oriainal. we theo.retically show that for any set of n distinct vectors, there exists a seven-layer autoencoder withthe optimal compression ratio,(i.e... the size of the middle laver is logarithmic in n), but that thereis a set of n vectors for which there is no three-layer autoencoder with a middle layer of logarithmic size. We also show several related results, including upper and lower bounds of the numbeof nodes reguired in the decoder part of an autoencoder. Furthermore, we extend the problemand methodology for analysis of autoencoders consisting of RelU activation functions in whichinput and output data consist of real valued vectors.

报告人简介：

Tatsuya Akutsu received the B.E. and M.E. degrees in aeronautics and the D.E. degree in information engineering from the University ofTokyo, Tokyo, Japan, in 1984, 1986, and 1989,respective.Iy. From 1989 to 1994, he was with Mechanical Engineering Laboratory, Japan. From 1994 to1 996, he was an associate professor in Department of Computer Science at Gunma UniversityFrom 1996 to 2001, he was an associate professor in Human Genome Center, Institute of MedicaScience, University of Tokyo. Since 2001, he has been a professor in Bioinformatics Center, insti.tute for Chemical Research, Kyoto University. He is a fellow of information Processing Society otJapan. His research interests include bioinformatics, complex networks, neural networks, anddiscrete algorithms.