Lossless Compression of Structured Convolutional Models via Lifting

Authors: Gustav Sourek, Filip Zelezny, Ondrej Kuzelka

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To test the proposed compression in practice, we selected some common structured convolutional models, and evaluated them on a number of real datasets from the domains of (i) molecule classification and (ii) knowledge-base completion. The questions to be answered by the experiments are: 1. How numerically efficient is the non-exact algorithm in achieving lossless compression? 2. What improvements does the compression provide in terms of graph size and speedup? 3. Is learning accuracy truly unaffected by the, presumably lossless, compression in practice?
Researcher Affiliation Academia Gustav Sourek, Filip Zelezny, Ondrej Kuzelka Department of Computer Science Czech Technical University in Prague {souregus,zelezny,kuzelon2}@fel.cvut.cz
Pseudocode No The paper describes the two algorithms in narrative text but does not include any explicit 'pseudocode' or 'algorithm' blocks with structured formatting.
Open Source Code No The paper does not include an unambiguous statement or a link to open-source code for the methodology described.
Open Datasets Yes For structure property prediction, we used 78 organic molecule classification datasets reported in previous works (Ralaivola et al., 2005; Helma et al., 2001; Lodhi & Muggleton, 2005). For knowledge base completion (KBC), we selected commonly known datasets of Kinships, Nations, and UMLS (Kok & Domingos, 2007).
Dataset Splits Yes We then trained against MSE using 1000 steps of ADAM, and evaluated with a 5-fold crossvalidation.
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 software frameworks like 'Py Torch Geometric (Py G)' and 'Deep Graph Library (DGL)' and 'ADAM' but does not specify their version numbers.
Experiment Setup Yes We approached all the learning scenarios under simple unified setting with standard hyperparameters, none of which was set to help the compression (sometimes on the contrary). We used the (re-implemented) LRNN framework to encode all the models, and also compared with popular GNN frameworks of Py Torch Geometric (Py G) (Fey & Lenssen, 2019) and Deep Graph Library (DGL) (Wang et al., 2019). If not dictated by the particular model, we set the activation functions simply as CW = 1 / (1+e^(W x)) and A = avg. We then trained against MSE using 1000 steps of ADAM, and evaluated with a 5-fold crossvalidation. For a more fair comparison, we further increased all (tensor) dimensions to a more common dim=10.