C*-algebra Net: A New Approach Generalizing Neural Network Parameters to C*-algebra

Authors: Yuka Hashimoto, Zhao Wang, Tomoko Matsui

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply our framework to practical problems such as density estimation and few-shot learning and show that our framework enables us to learn features of data even with a limited number of samples.Then, in Section 5, we discuss practical applications and show numerical results.
Researcher Affiliation Collaboration 1NTT Network Service Systems Laboratories, NTT Corporation, Tokyo, Japan 2Institute for Disaster Response Robotics, Waseda University, Tokyo, Japan 3Department of Statistical Modeling, the Institute of Statistical Mathematics, Tokyo, Japan.
Pseudocode No The paper does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes The source code of this paper is available at https://www.rd.ntt/e/ns/qos/ person/hashimoto/code_c_star_net.zip.
Open Datasets Yes We used the mini Image Net dataset (Vinyals et al., 2016), which is composed of 100 classes, each of which has 600 images, and considered the 5-way 1-shot task in the same manner as Subsection 4.2 in (Rusu et al., 2019).
Dataset Splits No No explicit mention of a validation set split or its details was found.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions 'masked autoregressive flow' and 'Adam' but does not specify their version numbers or other software dependencies with versions.
Experiment Setup Yes We set the learning rate so that it decays polynomially starting from 0.001 with decay rate 0.5.We set the hyperparameter λ in Eq. (4) as 0.3 and set the distribution D on Z as the uniform distribution on S9 i=1{z Z | z zi 0.05}. We set the learning rate so that it decays polynomially starting from 0.001 with decay rate 0.5.In this experiment, we set µ = 0.1.For a new task Tnew, we set the finite-dimensional subspace V of AN as N1 j=1 Span{vj,1, . . . , vj,l}N0. Here, l 10, vj,i(z) = e 10 zj,i pj(z) 2, zj,1 = Z(Tnew)64(j 1)+1:64j, and zj,i for i = 2, . . . , l are randomly drawn from the normal distribution with mean zj,1 and standard deviation 0.01. µ = 0.05 for Task 1 and µ = 0.1 for Task 2.