Function-space Inference with Sparse Implicit Processes

Authors: Simon Rodrı́guez-Santana, Bryan Zaldivar, Daniel Hernandez-Lobato

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

Reproducibility Variable Result LLM Response
Research Type Experimental We have evaluated SIP in extensive regression experiments, including synthetic and real-world problems. When compared against other methods from the literature, SIP often leads to better generalization properties and it captures complex patterns in the predictive distribution. It can also adjust the prior parameters to explain the training data. Finally, in very large datasets, it also remains scalable, reaching good results in less time than other methods.
Researcher Affiliation Academia 1Institute of Mathematical Sciences (ICMAT-CSIC), Madrid, Spain. 2Institute of Corpuscular Physics, University of Valencia and CSIC, Spain. 3Escuela Polit ecnica Superior, Universidad Aut onoma de Madrid, Spain.
Pseudocode No The paper describes the proposed method using prose and mathematical equations but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes The tensor-flow code for SIP is included in the github repository.
Open Datasets Yes We compare each method and SIP with each prior (i.e., a BNN and a NS) on multivariate regression problems from the public UCI dataset repository (Dua & Graff, 2017).
Dataset Splits Yes We split the data 20 times into train and test with 80% and 20% of the instances, respectively.
Hardware Specification Yes Each of them employed on its own 2 CPU Intel(R) Xeon(R) Gold 5218 CPU at 2.30GHz (16 cores, 22 Mb L3 cache), [32 cores in total], with 192 GB RAM at 2,4 GHz.
Software Dependencies No The paper mentions using 'tensor-flow code' but does not specify a version number for TensorFlow or any other software libraries or dependencies.
Experiment Setup Yes Each method is trained until convergence using a mini-batch size of 10. For training, in SIP we use 100 samples to estimate (12) and its gradients. In test, 500 samples are used in (14). The exact number of inducing points used is specified in each experiment... Finally, we use α = 1 in SIP in the synthetic experiments (...) and α = 0.5 in the real-world ones.