I'm Tim a PhD Student

This project was submitted to the ICLR Reproducibility Challenge 2019. The paper we reproduced is “A Resizable Mini-batch Gradient Descent based on a Multi-Armed Bandit”. Its aim is to combine SGD and Multi-Armed Bandit to chose the size of the mini-batch used. Throughout the optimization procedure, the current estimate oscilates between zones where the loss surface is rather flat, and zones where the loss surface is much more complicated. This adaptive size of the mini-batch combines fast descent rates when possible and variance reduction when needed. This is the Exploration-Exploitation Dilemma. Below, you can see a heamap showing the chosen batch size at each epoch for MNIST optimization.

It is beautiful to see that large batch sizes are chosen at the end of the optimization since variance reduction is needed approaching the local minimum.

Full paper and code on my github

The goal of this project is to get familiar with CUDA technology while following all the best practices w.r.t. optimization such as coalsced memory access, usage of shared memory, no bank conflicts etc. The Conjugate-Gradient algorithm for solving linear system was implemented as it contains a lot of paralellizable like a matrix multiplication for instance.

The outcome from the GPU version of the code (NVIDIA K40) was a speedup of about 30 compared to the sequential CPU version.

Read more on my github

The goal of this project was to make a self built, fully autonomous quadcopter that has to navigate along a path of QR code it detects on the way. The first part of this project was the theoretical implementation of linear-quadratic regulators (LQR) for 3 differents state space systems (attitude, altitude, navigation). For each of these systems, we also developped Kalman Filters (KF) for estimating the state vectors instead of purely relying on the sensor measurements.

The second part of this project was embedded development for running our high level code on ZYBOs. This was the most tricky part of the project; we had to run the navigation controller at 5-10 Hz because an image processing thread was using a big part of the CPU time. A video of the fully autonomous drone can be found here.

Video of the autonomous quadcopter on Google drive

See the code on my github

This project aims at understanding how Deep Learning frameworks like Pytorch and Tensorflow work under the hood. We have implemented a modular mini Deep Learning framework for training neural networks. It currently contains linear layers, MSE loss, optimizers and different activation functions. The goal here is not to build a robust framework for practical use but to get a feel with the inner workings of modular backpropagation.

See the code on my github

Have you ever missed a train because your previous train was late? Stochastic planners give an overview of how feasible a certain trip is. This gives the user the possibility to take alternative routes when the "fastest route" is not very likely. The historical data used here comes from the Swiss network and is made available by SBB.

See the code on my github