Efficient Low Rank Gaussian Variational Inference for Neural Networks

Authors: Marcin Tomczak, Siddharth Swaroop, Richard Turner

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We find that adding low-rank terms to parametrized diagonal covariance does not improve predictive performance except on small networks, but low-rank terms added to a constant diagonal covariance improves performance on small and large-scale network architectures. In our experiments, we focus on demonstrating the following: (i) ELRG-D-VI improves over MF-VI for small networks, but not on larger models, (ii) ELRG-VI has better predictive performance than MF-VI, (iii) ELRG-VI scales up to large CNNs and provides better predictive distributions than MAP, MF-VI and MC Dropout, (iv) sharing variational samples as in [34, 41, 43] leads to poor predictive performance.
Researcher Affiliation Academia Marcin B. Tomczak University of Cambridge Cambridge, CB2 1PZ, UK mbt27@cam.ac.uk Siddharth Swaroop University of Cambridge Cambridge, CB2 1PZ, UK ss2163@cam.ac.uk Richard E. Turner University of Cambridge Cambridge, CB2 1PZ, UK ret26@cam.ac.uk
Pseudocode No The paper discusses algorithmic details and computational costs but does not contain any structured pseudocode or an explicitly labeled algorithm block.
Open Source Code Yes We open-source the implementation of the algorithm derived in this paper at https://github.com/marctom/elrgvi.
Open Datasets Yes We consider a two dimensional synthetic classification dataset... classify vectorized MNIST [26] images... We experiment with common simple computer vision benchmarks: MNIST, KUZUSHIJI [5], FASHIONMNIST [47] and CIFAR10 [23]... We consider 4 data sets: CIFAR10, CIFAR100 [23], SVHN [32] and STL10 (10 classes, 5000 images 96 96) [7].
Dataset Splits No We plot the learning curves in Figure 2... avg. valid neg. log likelihood... Using K > 0 improves held out log likelihood by a visible margin... The paper uses a validation set but does not explicitly provide specific split percentages or sample counts for training, validation, and test sets, only referring to performance metrics on them.
Hardware Specification No The paper does not provide any specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running its experiments.
Software Dependencies No The paper mentions using the 'ADAM optimizer [20]' but does not provide specific version numbers for any software dependencies or libraries.
Experiment Setup Yes We use the default ADAM optimizer [20], batch size of 256, 1 variational samples per update, and run optimization for 500 epochs. We train all models for 500 epochs (except MAP, which is run for 50 epochs) using a batch size of 512 using the ADAM optimizer [20], and do not use data augmentation. We train all algorithms for 200 epochs using a batch size of 256 and the ADAM optimizer [20], with data augmentation.