Generalizability of Neural Networks Minimizing Empirical Risk Based on Expressive Power

Authors: Lijia Yu, Yibo Miao, Yifan Zhu, XIAOSHAN GAO, Lijun Zhang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we give some simple experiments to validate our theoretical conclusions. Our experimental setup is as follows. We used MNIST data set and two-layer networks with Re LU activation function. When training the network, we ensure that the absolute value of each parameter is smaller than 1 by weight-clipping after each gradient descent. Two experiments are considered: About size and accuracy: For networks with widths 100,200,. . . ,900,1000, we observe their accuracy on the test set after training. The results are shown in Figure 1. About data and precision: Using training sets with 10%, 20%, . . . , 90%, 100% of the original training set to train a network with widths 200, 400 and 600. The results are shown in Figure 2.
Researcher Affiliation Academia Lijia Yu1, Yibo Miao2, 3, Yifan Zhu2, 3, Xiao-Shan Gao2, 3 , Lijun Zhang1, 3, 4 1 Key Laboratory of System Software of Chinese Academy of Sciences Institute of Software, Chinese Academy of Sciences 2 State Key Laboratory of Mathematical Sciences Academy of Mathematics and Systems Science, Chinese Academy of Sciences 3 University of Chinese Academy of Sciences 4 Institute of AI for Industries, Chinese Academy of Sciences
Pseudocode No The paper describes methods using mathematical formulations and natural language, but does not include any explicit pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements about releasing source code or links to a code repository.
Open Datasets Yes In this section, we give some simple experiments to validate our theoretical conclusions. Our experimental setup is as follows. We used MNIST data set and two-layer networks with Re LU activation function.
Dataset Splits Yes About data and precision: Using training sets with 10%, 20%, . . . , 90%, 100% of the original training set to train a network with widths 200, 400 and 600.
Hardware Specification No The paper describes experimental setup and results but does not specify any hardware details like GPU/CPU models or other computing resources.
Software Dependencies No The paper describes experimental setup but does not mention specific software names with version numbers (e.g., programming languages, libraries, or frameworks).
Experiment Setup Yes Our experimental setup is as follows. We used MNIST data set and two-layer networks with Re LU activation function. When training the network, we ensure that the absolute value of each parameter is smaller than 1 by weight-clipping after each gradient descent. Two experiments are considered: About size and accuracy: For networks with widths 100,200,. . . ,900,1000, we observe their accuracy on the test set after training. The results are shown in Figure 1. About data and precision: Using training sets with 10%, 20%, . . . , 90%, 100% of the original training set to train a network with widths 200, 400 and 600. The results are shown in Figure 2.