Microsoft CNTK: An introduction with an example implementation of semantic segmentation

While TensorFlow, Caffe and Theano are popular frameworks in the machine learning community, Microsoft’s CNTK received limited attention after its first release in 2016. However, with the latest releases it evolved into a serious competitor of TensorFlow regarding functionality and examples. As a global, innovative player in the automotive industry, IAV already started to evaluate various ML toolchains including CNTK. One popular use case for ML in the context of driverless cars is semantic segmentation. This talk enables participants to be able to set up a prototyping workflow using CNTK in combination with Jupyter and Matplotlib, read data and use transfer learning with pre-trained Caffe models.

Required audience experience: Basic knowledge of ML frameworks.

Objective of the talk:

• Get a deeper understanding of the CNTK
• Set up a prototyping workflow using CNTK, Jupyter and Matplotlib
• Application of transfer learning with pre-trained models
• Example implementation of semantic segmentation (PASCAL VOC datasets)

Keywords: CNTK, CNNs, Semantic Segmentation, TensorFlow, Caffe