Efficient Learning of Discrete-Continuous Computation Graphs

Authors: David Friede, Mathias Niepert

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

Reproducibility Variable Result LLM Response
Research Type Experimental With an extensive set of experiments, we show that we can train complex discrete-continuous models which one cannot train with standard stochastic softmax tricks. We also show that complex discrete-stochastic models generalize better than their continuous counterparts on several benchmark datasets. 4 Experiments The aim of the experiments is threefold.
Researcher Affiliation Collaboration David Friede1,2 david@informatik.uni-mannheim.de 2University of Mannheim Mannheim, Germany Mathias Niepert1 mathias.niepert@neclab.eu 1NEC Laboratories Europe Heidelberg, Germany
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes The implementations are in Py Torch and can be found at https://github.com/nec-research/dccg.
Open Datasets Yes Unsupervised Parsing on List Ops The Listops dataset contains sequences in prefix arithmetic syntax such as max[ 2 9 min[ 4 7 ] 0 ] and its unique numerical solutions (here: 9) [22]. ... Multi-Hop Reasoning over Knowledge Graphs Here we consider the problem of answering multi-hop (path) queries in knowledge graphs (KGs) [10]. We evaluate various approaches on the standard benchmarks for path queries [10].2 ... End-to-End Learning of MNIST Addition The MNIST addition task addresses the learning problem of simultaneously (i) recognizing digits from images and (ii) performing the addition operation on the digit s numerical values [20].
Dataset Splits No The paper mentions using specific datasets and generating test examples for extrapolation but does not provide explicit numerical details (percentages or counts) for train/validation/test dataset splits, nor does it refer to specific predefined splits with quantitative information.
Hardware Specification Yes All experiments were run on a Ge Force RTX 2080 Ti GPU.
Software Dependencies No The paper mentions 'Py Torch' but does not specify its version number or the version numbers of any other software dependencies.
Experiment Setup Yes All models are run for 100 epochs with a learning rate of 0.005 and we select τ {1, 2, 4}. We choose a dimension of 256, a batch size of 512 and learning rate of 0.001. We further train 1vs All with the cross-entropy loss for 200 epochs and with a temperature of τ = 4. We use a learning rate of 0.0001 and a temperature of τ = 8.