Implicit Posterior Variational Inference for Deep Gaussian Processes

Authors: Haibin YU, Yizhou Chen, Bryan Kian Hsiang Low, Patrick Jaillet, Zhongxiang Dai

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluation shows that IPVI outperforms the state-of-the-art approximation methods for DGPs. We empirically evaluate and compare the performance of our IPVI framework against that of the state-of-the-art SGHMC [18] and doubly stochastic VI [48] for DGPs based on their publicly available implementations using synthetic and real-world datasets in supervised (e.g., regression and classification) and unsupervised learning tasks.
Researcher Affiliation Academia Dept. of Computer Science, National University of Singapore, Republic of Singapore Dept. of Electrical Engineering and Computer Science, MIT, USA {haibin,ychen041,daiz,lowkh}@comp.nus.edu.sg, jaillet@mit.edu
Pseudocode Yes Figure 1: Best-response dynamics (BRD) algorithm based on our IPVI framework for DGPs. Algorithm 1: Main, Algorithm 2: Player 1, Algorithm 3: Player 2
Open Source Code Yes Our implementation is built on GPflow [41] which is an open-source GP framework based on Tensor Flow [1]. It is publicly available at https://github.com/Hero Killer Ever/ipvi-dgp.
Open Datasets Yes We empirically evaluate and compare the performance of our IPVI framework...using synthetic and real-world datasets...UCI Benchmark Regression...Large-Scale Regression...Year MSD dataset...Airline dataset...Frey Face dataset [47]
Dataset Splits No We have performed a random 0.9/0.1 train/test split. The paper does not explicitly describe a validation dataset split or provide specific percentages for it.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., CPU, GPU models, memory) used to run its experiments, only mentioning the use of TensorFlow.
Software Dependencies No Our implementation is built on GPflow [41] which is an open-source GP framework based on Tensor Flow [1]. (No specific version numbers for GPflow or TensorFlow are provided in the text).
Experiment Setup Yes the depth L of the DGP models are varied from 1 to 5 with 128 inducing inputs per layer. The learning rates are 0.005 and 0.02 for IPVI and SGHMC (default setting adopted from [18]), respectively. We utilize a 4-layer DGP model with 100 inducing inputs per layer and a robust-max multiclass likelihood [21]; for MNIST dataset, we also consider utilizing a 4-layer DGP model with 800 inducing inputs per layer to assess if its performance improves with more inducing inputs.