DevDiary

I could check out another deep learning meetup. This time it was hosted at Google's Isar Valley here in Munich. The three interesting walks were about the following topics: Visual Sentiment Analysis with Deep Convolutional Neural Networks (by Dr. Damien Borth, DFKI) Strategies for AI Deployment (by Henrik Klagges, TNG Consulting GmbH) DGX-1 and SATURNV: The World’s Most Efficient Supercomputer for AI and Deep Learning (by Ralph Hinsche, Nvidia) In the third talk of Nvidia, we were also able to hold a test sample of the latest Tesla P100 in our hands, which is one of the building blocks of Nvidia's deep learning super computers called DGX-1 . This is a nice super toy that every AI-researcher would like to have under the Christmas tree. Unfortunately, a single device costs more than 100.000 US-Dollar.

In the course of the development of my Master's Thesis "Deep Learning Approaches to Predict Future Frames in Videos" at TUM, I realized that the high flexibility of TensorFlow has its price: boilerpate code. Many things that are needed in almost every neural network training or evaluation script have to be implemented over and over again. To that end, I started to implement a high-level API for Google's machine intelligence library, called TensorLight . TensorLight comes with four guiding principles: Simplicity:  Straight-forward to use for anybody who has already worked with TensorFlow. Especially, no further learning is required regarding how to define a model's graph definition. Compactness:  Reduce boilerplate code, while keeping the transparency and flexibility of TensorFlow. Standardization:  Provide a standard way in respect to the implementation of models and datasets in order to save time. Further, it automates the whole training and validatio...

I finally finished my Master's Thesis in the Computer Vision chair at TUM. In the course of this thesis, I analyzed existing deep learning approaches to predict future frames in videos. Based on these findings and other modern deep learning practices, such as batch normalization, scheduled sampling to improve recurrent network training or ConvLSTMs, we were able to reach or event outperform state-of-the-art performance in future frame generation. So far, many people asked me about the practical application of frame prediction. Unfortunately, it won't tell us the end of any cliff-hanger movie such as Inception, but the main purpose of such a system is not to generate a perfect forecast of the long-term continuation of any movie clip. This completely impossible in my opinion, since there is not always a wrong or right in many situations. A neural network cannot be able to predict every decision made by all objects inside the scene. Furthermore, the pose of the camera or the ...