This Week in AI, February 1, 2018
Deep Learning best practices, Galileo’s Telescope, OpenAI’s newly-released unsolved problems, and more!
Here’s a really great list of best practices for improving your deep learning models. Of particular interest is the bit on learning rate optimization and stochastic gradient descent restarts. Both of these methods are generally applicable and are likely to improve models you’re working on right now. Also interesting, the list includes examples using the fastai package, a high-level framework developed by Fast.ai on top of PyTorch.
Ali Rahimi wrote a great article comparing the current state of deep learning with the understanding of optics when Galileo built his telescope 400 years ago. Back then, scientists had general ideas of how light bent through glass, but didn’t understand the principles well enough to make true optical devices. Over time as we learned more about optics, we were able to develop mental models that allow us to construct complex stacks of lenses.
We don’t have these mental models for deep learning yet, and for the most part, experts don’t understand why deep learning works so well. As more research is done in this field, we should start seeing better abstractions and language for describing what each part of a deep learning models is doing.
You’ve trained your network but how do you let other people actually use it? A great first option is to build a REST API with a simple package like Flask (one of my favorite Python packages). This way your model is available for training and making predictions using URLs and common HTTP methods. Learn how to serve your model with this great tutorial using Keras and Flask.
OpenAI has released a new set of unsolved problems in deep learning. These problems span several domains such as using reinforcement learning to learn the classic Snake game, and training an autoencoder to generate new data for augmenting datasets. This looks like a great set of problems for deep learning beginners—and experienced practitioners—to tackle.
Here’s a really cool project by Jeff Zito where video footage was generated by a deep learning model based on music from Lord Over. The videos can be somewhat disturbing, but I love seeing people making art with deep learning. Zito’s model is based on a deep learning model used to generate realistic videos from audio alone. It’s truly impressive (and somewhat worrying) how well the mouth movements match the audio.
Stay tuned for new updates as we continue to review all that’s new in the world of AI! And if you’re interested in mastering these transformational skills, and building a rewarding career in this amazing space, consider one of our Nanodegree programs:
Source: Deep Learning on Medium