Multiscale Adaptive Representation of Signals: I. The Basic Framework

Authors: Cheng Tai, Weinan E

JMLR 2016 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we discuss some examples of applications of the multi-scale adaptive frames, the Ada Frames. A thorough comparison of the proposed model and other existing models will be postponed to future publications. ... In this section, we discuss some examples of applications of the multi-scale adaptive frames, the Ada Frames. ... In section 8, we discuss applications to image processing and image classification.
Researcher Affiliation Academia Cheng Tai EMAIL PACM, Princeton University Princeton, NJ 08544, USA Weinan E EMAIL School of Mathematical Sciences and BICMR Peking University and Department of Mathematics and PACM Princeton University Princeton, NJ 08544, USA
Pseudocode Yes Algorithm 1 Adaptive construction of frames 1: Input: x. 2: Initialize k = 0, B = 0, A = A0, D = WA0x. 3: while not converge do 4: Dk+1 arg min D D 1,1 + η 2 D WAkx Bk 2 F 5: Ak+1 arg min A WAx Dk+1 + Bk 2 F s.t. A Q. 6: Bk+1 Bk + WAk+1 Dk+1. 7: k k + 1 8: return Ak
Open Source Code No The software we use is downloaded from http://www.cs.technion.ac.il/~ronrubin/software. This refers to the software for the K-SVD algorithm, not the authors' own implementation for Ada Frame.
Open Datasets Yes Our next example is the fingerprint dataset (Maltoni et al., 2009). ... We tested this idea on the extended Yale human face dataset B (Lee et al., 2005). ... The dataset we used is MNIST (Le Cun et al., 1998).
Dataset Splits Yes The dataset we used is MNIST (Le Cun et al., 1998). It contains 70000 28 28 images of digits from 0 to 9, 60000 for training and 10000 for testing.
Hardware Specification No In our laptop with the same setup, the K-SVD algorithm takes 25s to train a dictionary with 256 atoms of support size 8 8 and 6.5s to denoise the image. The Ada Frame takes 3.7s to train 64 filters with support size 8 8 and takes 0.6s to denoise. This mentions a 'laptop' but lacks specific details (CPU, GPU, RAM, etc.).
Software Dependencies No The paper mentions 'MATLAB syntax' but does not specify a version. It also mentions 'K-SVD algorithm' and a URL to download its software, but does not provide version numbers for K-SVD or any other libraries/frameworks used for their own implementation.
Experiment Setup Yes We use Algorithm 1 to learn the filters with η = 102. ... The filters are learned using Algorithm 2 with parameters η = 102, λ = 103. ... 4 filters of support size 6 6 are learned using Algorithm 1 with η = 102. ... Initialization is done using the Daubechies filters db3. ... 7 levels of decompositions are performed. ... we train the filters both from the noisy image and the clean image with m = 36, r = 6, η = 102, λ = 103. A two-level decomposition is performed. The soft thresholding parameter is set to be 0.14. Initialization is done by setting the filters to be random orthogonal vectors.