Dennis Bautembach

Unpublished PhD Thesis

In my PhD thesis I summarize and elaborate on my research work of the last three years, including never-before-released (implementation) details. If you want to learn how to write high-performance, parallel SNN simulators from scratch on GPGPU hardware, this will be of interest to you.

Resources:

• Thesis

10.1109/HPEC49654.2021.9622805

We present two novel optimizations for SNN simulation. The first, called “shared memory-based spike delivery” is a blocking scheme that makes it possible to load neurons piecewise into shared memory and perform/deliver all updates/spikes there. The second is a variation of event-driven plasticity called “lazy event-driven plasticity” which performs fewer updates and in a more memory-efficient way compared to event-driven plasticity. Both optimizations combined sped up Spike by more than 2x, re-establishing ourselves as the fastest SNN simulator.

Resources:

• Paper
• Source code

10.1109/IJCNN52387.2021.9533921

We present the first SNN simulator to scale to 8 GPUs, allowing you to simulate larger models and in less time than ever before. Features include:

• State of the art performance (incl. lightning fast setup)
• Ability to define custom models in native C++
• Modern, user-friendly API
• No 3rd-party dependencies (except CUDA)

Resources:

• Paper
• Source code

10.1109/IJCNN48605.2020.9206752

“Spice” is a state of the art spiking neural network simulator written in C++ and CUDA. It

• is up to 3x faster than the competition,
• allows you to specify your own models through a modern and intuitive API (you simply inherit from a model base class and override callbacks),
• builds out of the box without any 3rd-party dependencies (apart from CUDA), and
• does not involve any proprietary compilation steps or domain-specific languages.

Resources:

• Paper
• Source code

github.com/denniskb/hy475

As part of my TA duties I designed a fully interactive online course on robotic navigation. It allows you to code and simulate robots directly in your browser using JavaScript.

Student testimonials:

• “I really felt that somebody cared about me as a student after all these years of studies.”
• “I have to say you make the most interesting assignments. Most of the assignments of other courses do not have any meaning.”
• “a big thanks for your help … no one had ever spend so much time … and actually helped us with immediate feedback, so please keep it up for the next years.”

10.3390/technologies6040097

We estimate human poses by treating them as missing/incomplete rows in a “pose database” matrix and calculating missing entries via matrix completion.

Paper

10.1145/2461912.2461940

KinectFusion is a well-known 3D scanner powered by Microsoft’s Kinect camera. It has recently found its way into the official Kinect for Windows SDK. While KinectFusion produces very high-quality scans in real-time, it is limited to relatively small spaces.

In this paper we extend KinectFusion with a sparse data structure and streaming, allowing us to scan areas of virtually infinite size while maintaining the quality and performance of the original algorithm.

Resources:

• Paper

Unpublished BSc thesis

Voxels date back 20 years. They have many advantages over polygon meshes. However, they always suffered form the lack of a hardware-accelerated rendering pipeline and the infeasiblity of animating them. I tackled this age-old computer graphics problem in my BSc thesis. The (GPU-accelerated) animation technique I developed enables:

• Skeletal animation of voxels models
• Rendering of mixed polygon- and voxel-based content.

Resources:

• BSc thesis
• In-depth video tutorial series
• Source code