Amortised Learning by Wake-Sleep

Authors: Li Wenliang, Theodore Moskovitz, Heishiro Kanagawa, Maneesh Sahani

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate ALWS on a wide range of generative models. Details for each experiment can be found in Appendix C. The results are shown in Figure 8. According to FID, ALWS-A is the best ML method for binarised MNIST, Fashion, and CIFAR-10.
Researcher Affiliation Academia 1Gatsby Computational Neuroscience Unit. Correspondence to: Li K. Wenliang <kevinli@gatsby.ucl.ac.uk>.
Pseudocode Yes Algorithm 1: Amortised learning by wake sleep
Open Source Code Yes Code is at github.com/kevin-w-li/al-ws
Open Datasets Yes We chose six benchmark datasets: the binarised and original MNIST (Le Cun et al., 1998) (B-MNIST and MNIST, respectively), fashion MNIST (Fashion) (Xiao et al., 2017), natural images (Natural) (Hateren & Schaaf, 1998), CIFAR10 (Krizhevsky et al., 2009) and Celeb A (Liu et al., 2015).
Dataset Splits No The paper uses standard benchmark datasets like MNIST, CIFAR-10, etc., but does not explicitly state the train/validation/test splits (e.g., percentages or sample counts) used for its experiments, nor does it cite a specific methodology for splitting.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory specifications) used for running experiments were mentioned.
Software Dependencies No The paper mentions various algorithms and frameworks (e.g., DCGAN) but does not provide specific software version numbers for key dependencies (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For ALWS, we used a Gaussian kernel with a bandwidth equal to the median distance between samples generated for each b, and set λ = 0.01. Each algorithm is run for 50 epochs 10 times with different initialisations, except for SIVI where we trained for 1000 epochs with a lower learning rate for stability.