Representing Hyperbolic Space Accurately using Multi-Component Floats

Authors: Tao Yu, Christopher M. De Sa

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

Reproducibility Variable Result LLM Response
Research Type Experimental Theoretically and experimentally we show our model has small numerical error, and on embedding tasks across various datasets, models represented by multi-component floating-points gain more capacity and run significantly faster on GPUs than prior work.
Researcher Affiliation Academia Tao Yu Department of Computer Science Cornell University tyu@cs.cornell.edu Christopher De Sa Department of Computer Science Cornell University cdesa@cs.cornell.edu
Pseudocode Yes Algorithm 1: Two-Sum Input: p-bit floats a, b, where p 3 x fl(a + b) bvirtual fl(x a) avirtual fl(x bvirtual) broundoff fl(b bvirtual) aroundoff fl(a avirtual) y fl(aroundoff + broundoff) Return: (x, y)
Open Source Code No We will release our code in Py Torch publicly for reproducibility, and in hopes that practitioners can use MCF to reliably compute in hyperbolic space.
Open Datasets Yes We conduct embedding experiments on datasets including the Gr-Qc and commonly used Word Net Nouns, Verbs and Mammals dataset whose statistics are shown in Table 2. ... Table 2: Datasets Information Datasets Nodes Edges Word Net[12] 74374 75834 ... Gr-QC[22] 4158 13422
Dataset Splits No The paper mentions using datasets from WordNet and Gr-QC but does not specify the train/validation/test splits (e.g., percentages, counts, or specific pre-defined split references).
Hardware Specification Yes For all timing results within this section, we use GPU: Ge Force RTX 2080 Ti and CPU: Intel(R) Xeon(R) Gold 6240 @2.60GHZ.
Software Dependencies No The paper mentions deep learning frameworks like PyTorch and TensorFlow, and Julia, but does not specify their version numbers or the versions of other software dependencies required to replicate the experiments.
Experiment Setup Yes Herein, we fix the negative sampling size to be 50 and only vary the batchsize in 32, 64, 128.