Christoffer Löffler

I am a profesor asociado at the school of informatics of the Pontificia Universidad Católica de Valparaíso (PUCV) in Chile.

My research is on machine learning with few labeled data and broadly on time series. I also contributed to optical positioning, complex event processing, and large-scale virtual reality.

Recently, I got my Dr.-Ing. at Björn Eskofier's MaD Lab, while working at Fraunhofer IIS in the ADA Lovelace Center as a senior scientist. I also did my Bachelors and Masters at Friedrich-Alexander University Erlangen-Nürnberg (FAU), where I was advised by Christopher Mutschler, back then researching at Michael Philippsen's Programming Systems lab. These and other collaborations led to a Dijkstra number of four and an Erdös number of fivefour!

Email  /  CV  /  Google Scholar  /  ResearchGate  /  ORCID  /  Github  /  Twitter  /  LinkedIn  /  Youtube  /  Blog

I'm interested in machine learning with few labeled data (like active learning) and building interactive systems for the real-world.

This dissertation provides streamlined information retrieval, cost-effective annotation through Deep Active Learning, and efficient learning of similarity functions. Three key publications contribute to the thesis: a metric learner for enhanced information retrieval, an imitation learning approach (IALE) optimizing Deep Active Learning, and a method combining fine-tuning with active learning, facilitating adaptive similarity search for unstructured data and cost-effective annotation of complex datasets.

A framework of six orthogonal metrics for gradient- or perturbation-based post-hoc visual interpretation methods. Its designed for time series classification and segmentation tasks.

This work improves information retrieval in a football trajectory dataset using deep metric learning and an entropy-based active learning method, and analyzes the effectiveness of sampling heuristics through a user study.

IALE uses imitation learning to select informative data samples in deep active learning by imitating the best-performing expert heuristic, and outperforms state-of-the-art imitation learners and heuristics on well-known image datasets.

We propose an imitation learning approach that learns a policy for active learning from an ensemble of deep active learners.

We enable interactive search in unordered sets of trajectories, with a focus on team sports.

We show that post-training quantization (done greedily) benefits from an optimal global bit-width and evaluate this on VGG, ResNet, UNet and our tool tracking FCN.

We combine inertial sensors, mounted on work tools, with AR-headsets to enrich modern assistance systems with a sense of process quality, powered by machine learning.

We describe an AutoML system for our custom hardware and classify multivariate data using deep and shallow methods.

Learning to fuse absolute poses (6 degrees of freedom) with optical flow (e.g., FlowNet) to improve a mobile agent's self positioning.

An evaluation of the big AR systems for real use-cases.

Using reference positioning systems, we record a multi-camera dataset with exact labels and propose criteria for evaluating indoor positioning.

How to build a cheap but reliable optical positioning system with Raspberry Pi with active LED markers and continuously identify them.

Event-Based Systems (EBS) can efficiently analyze large streams of sensor data in near-realtime. But they struggle with noise or incompleteness that is seen in the unprecedented amount of data generated by the Internet of Things

We've won the DEBS 2014 Grand Challenge with our hidden Markov model approach.

We model a smartphone playing back music and its energy management.

When running a distributed low-latency event processing system, you may want to optimize latency - heuristically.

Positioning system with combined optical and a radio-based determination of a position of a tracker and a tracker with an active light source.

Method for improving the performance of an event-based distributed analysis system that uses meta heuristics.

This website is forked from here, thank you Jon Barron.