Deep Learning Weekly - Issue #34: Google's AI strategy, Deep Learning Limits, Transfer Learning and more...

In our latest blog post, we explore Google’s current AI strategy and investigate the role of the recent Kaggle acquisition. Google seems to be paving its way into leading the rising AI market, by making potential future engineers and scientists comfortable with TensorFlow, expanding it’s cloud platforms and offering new services. And all of this due to the slowly decreasing importance of search ad placements and rising AI markets.

Ever wondered what that ‘AI’ everyone is talking about today is actually capable of? This article explains what possible and what are some of the current problems, like 'individually unreliable neural networks’ mean for their applications.

A detailed review of a recent paper on FPGAs from Intel, that comes to the conclusion that FPGAs may indeed be the future for running deep neural nets, as they offer more computational power while using less energy, when used for low precision computations.

In his article, Sebastian Ruder contrasts transfer learning with the well known supervised learning, highlights the advantages of transfer learning and explains its details and a technical definition using different scenarios.  He finishes with applications, practical methods and a look into the future.

Guim Perarnau gives an overview on generative adversarial networks in general, sums up their evolution over the last year and highlights the most important improvements and optimizations. The whole article is packed with links to code, more detailed explanations, and further reading material, so you should definitely take a look.

In this exceptionally detailed tutorial on neural networks, you learn the basic concepts and mathematical details using Python and the Numpy package. After learning about neural networks in general, their structure, gradient descent and backpropagation, you finally create your own network and assess its accuracy.

Weld is the product of an MIT and Stanford cooperation, that improves the performance of data-intensive applications. It optimizes across libraries and functions by expressing the core computations in libraries using a small common intermediate representation, similar to CUDA and OpenCL. See this article for more details.

An implementation of Cyclical Learning Rates for Keras, that allows you to cyclically increase your learning rate in order to make your model converge faster.

If you’re fed up with implementing and training your deep learning models, take a look at this tutorial series on the Google Cloud Machine Learning APIs. Learn how to use the different endpoints using Python.

Fascinating model from DeepMind, that allows direct transcription of a foreign spoken language into another language.

Carlos Perez lists some of the rejected papers from this years ICLR and explains, why they are worth a read.

Style transfer has become the omnipresent research topic and has been taken even further in this paper. Apply any style to an image in real-time.