≡ Menu

Enterprise deep learning with TensorFlow

An interesting course offered by openSAP is Enterprise deep learning with TensorFlow, which is currently running in its last week. I found this a great insight in the current state of machine learning possibilities.

It was a very hands-on training where it was possible to play with TensorFlow Applications, an open-source library for numerical computation. For SAP, TensorFlow is a key element in the SAP Leonardo Machine Learning architecture. With SAP Leonardo, SAP aims to make machine learning euasy to use for businesses.

Deep learning is a sub-field of neural networks, machine learning, and artificial intelligence. It is inspired by the architecture of the human brain and consists of neural networks with many layers.

Deep learning is a promising approach when:

  • there is a large amount of training data available
  • it concerns solving an image/audio/natural language problem
  • the raw input data has little structure and it is needed for the model to learn meaningful representations (e.g., pixels in an image)

One of the topics in the course was about convolutional networks. Convolutional networks are used to classify objects on pictures. The complexity to do this is enormous, but with combining several techniques and doing smart optimizations it becomes possible.

Also some examples of use cases were given. One of them was a Medical Image Segmentation with Fully-Convolutional Networks. In this example images retrieved from an MRI scanner are processed with a fully convolutional network to construct a new image that points out possible cancer cells.

I found the explanation on how to deal with unsupervised and reinforcement learning very informative as well. To explain:
Machine learning applications fall into three broad contexts:

  • Supervised learning; in this case there is dataset with labels or annotations. Usually this dataset is not too big, because it is costly to label all the data. Most machine learning is done with these.
  • Unsupervised learning; in this case there is a data without labels or annotations. Typically this data is generated with machines or software, in an internet of things kind of way. With machine learning there are techniques to identify anomalies and outliers of the data. Making good use of this data. An example can be a financial pattern that is monitored. When an anomaly occurs, this can be due to fraud.
  • Reinforcement learning; In reinforcement learning there is no initial dataset. The dataset is accumulated with experience. The machine learning agents interact with the environment in an trial an error kind of way. An example is a robot learning a task. It performs actions, and when the action is correct it is rewarded, when the action is incorrect, there is no reward and a penalty.

Another inspiring example was the generating of new images using GANs (Generative Adversarial Networks). In this example a generator generates images, and this is combined with an discriminator that determines if the pictures is a real or fake image (i.e. blurry). This approach gives impressive results.

To conclude: another very inspiring course from the Open SAP learning environment. Very useful machine learning techniques for businesses were presented.