Neural Networks Framwork for Embedded Plattforms
Neural Networks have had a huge hype in the last decade, still very few implementations have been presented on embedded platforms, as neural networks are to compute-intense. Though, recent works have shown that correctly trained reduced-precision networks reach similar performance than their high-precision equivalents. In previous projects, we have successfully developed neural network implementations on a general purpose microcontroller platform and on our own PULP multicore processor platforms. But as applications and research is changing fast, these implementations need always to be adapted. In this thesis, you use common Machine learning framework like Torch or Tensorflow to evaluate and train the network and export it in a compilable form for efficient inference on the previously mentioned platforms. The work may also include implementation and evaluation of hardware-based special-purpose instructions on the PULP platform to accelerate the computing task.
- Knowledge of C/C++
- Interest Machine learning
- 20% Theory / Literature Research
- 80% software development