Machine Learning Demos

Notebooks and code snippets demonstrating various machine learning techniques:

8. PyTorch implementation of SNGP (Spectral-normalized Neural Gaussian Process) and comparison against VBLL

• Implementation of SpectralNormalization wrapper
• Implementation of RandomFeatureGaussianProcess
• Implementation of DeepResNet SNGP
• Test of replacing classification head with VBLL layers

7. Bias of predictive mean calculated with averages of Dropout layers

• Compare predictive means calculated by averaging samples from a BNN vs output calculated for Dropout averages.

6. Comparison of REINFORCE vs Gumbel-Softmax vs MDNF gradients and convergence for a simplified objective

• Optimization using REINFORCE vs reparametrization gradients (with GradientTape)
• Gumbel-Softmax relaxation for discrete variables - an illustration of a bias
• Mixture of Discrete Normalizing Flows relaxation for discrete variables

5. Illustration of how entropy of the relaxed categorical distribution can be estimated and utilized for VI

• Comparison (and discussion of gradients) of three estimates of the entropy/KL-term in ELBO

4. Variational Autoencoder using Relaxed Categorical distribution

• Sampling from Gumbel softmax with and without straight-through
• Implementation of different approaches to estimation of KL divergence
• Training with Mnist data
• Reconstruction of digits and unconditional sampling latent codes

3. A demonstration of Discrete Flows: Invertible Generative Models of Discrete Data

• Arithmetic on one-hot encoded vectors
• Trainig simple discrete transformation
• MLE-training of an autoregressive flow with masked autoencoder to match a target distribution.

2. Probabilistc Matrix Factorization model with mean-field variational inference.

• Probabilistc Matrix Factorization implementation
• estimating ELBO using MC
• training using pyTorch automatic differentiation
• simple evaluation of RMSE on test subset

1. Framing multi-output Bayesian optimization with GPyOpt

• fitting individual GPs
• fitting multi-task GPs using coregionalization
• BO optimization of single function
• BO optimization of 2-task problem
• An implementation of a custom acquisition function
• Extensive visualization of the optimization process