# Rasmus Berg Palm

Machine Learning Researcher and Engineer

# Blog

I’m a machine learning researcher and engineer. I’m broadly interested in probabilistic machine learning, generative models, bayesian statistics, unsupervised and self-supervised learning and inductive priors.

I’m currently working at twig.energy, trying to remove some carbon from the power grid. Prior to that I did a PostDoc in the REAL group at the IT University of Copenhagen working with Sebastian Risi. Before joining ITU I did my PhD at the Cognitive Systems group at the Technical University of Denmark with Ole Winther. Prior to that I was a Staff Engineer and Machine Learning Team lead at Tradeshift.

“What I cannot create, I do not understand.” - Richard Feynman

## Software

I often create software libraries and tools and I love publishing them as open source. It lets me write clean code, think about abstractions and interfaces and ship working software to users, which I enjoy.

• nanograd - The simplest and smallest possible library for autograd. Great for teaching.
• shapeguard - A tiny library, which allows you to very succinctly assert the expected shapes of tensors in a dynamic, einsum inspired way. A great tool for avoiding bugs.
• EvoStrat - A library that makes Evolutionary Strategies (ES) simple to use. It has a flexible and natural interface for ES that cleanly separates the environment, the reinforcement learning agent, the population distribution and the optimization, which allows you to use the standard pytorch nn.Modules for policy networks and torch.optim optimizers for optimization.
• pymc3-quap A quadratic approximation package for PyMC3.
• pytorch-lgssm A clean Linear Gaussian State Space Model for pytorch, which supports sampling and inference using the Kalman filtering algorithm.
• Blayze A very fast and efficient Bayesian Naive Bayes classifier which perfectly incorporates new information in an online learning setting and support both Gaussian and Categorical/Multinomial features.
• DeepLearnToolbox - A now deprecated Matlab toolbox for Deep Learning. I was unwise enough to do my Masters on Deep Learning in Matlab, which didn’t have any support for it, so I made a library for MLPs, CNNs, DBN, etc, implementing everything from the ground up, including the gradients, backprop, etc. It became quite popular. I quickly moved on to Python though (yay autograd!), and deprecated it.

## Teaching

I thoroughly enjoy teaching. It challenges me to understand things at a much deeper level, and I find it very rewarding.

Hosted on GitHub Pages — Theme by orderedlist