Deep Learning in the Wild

Climate change and human exploitation of our planet has a significant influence on the habitat and existence of wild animals. The resulting biodiversity loss threatens ecosystems and the human development that depends on them. Protecting these habitats is based on delivering evidence by collecting data. This is usually labour intensive, since it depends on field work done by biologist and volunteers.

This master thesis tries to make a contribution to scale up this important process by using acoustic detection of animals using deep learning on embedded systems.

Goals

  • Develop deep learning algorithms for bioacoustic signal detection
  • Evaluate the portability of these algorithms on an embedded machine learning platform (e.g. Intel Myriad)
  • Benchmark both predictive and computational performance of the developed algorithms

Tasks

  • Literature research and familiarisation with bioacoustic signal detection using deep learning / machine learning
  • Dataset preparation (e.g. species selection for detection)
  • Feature selection and data representation
  • Neural Networks (NN) training on a GPU-platform
  • NN-Algorithm porting an low power embedded platform (for prediction)
  • Document work and provide a 1 minute video summarizing the master thesis

Possible Technologies

  • Python, Matlab, R
  • TensorFlow, PyTorch, Keras, MXNet
  • librosa, scikit-learn, C++

Kind of Work
50% Deep Learning Theory and Implementation, 30% Embedded Implementation, 20% Benchmarking

Requirements
Machine Learning & Deep Learning Basics, Signal Processing, Python

Time & Effort
Master’s Thesis, 1 Student

Contact
fabian.schenkel@scs.ch