|December 29 · Issue #21 · View online |
Unsurprisingly this week was somewhat of a slow news week. Non-the-less, we still bring you interesting industry surveys, new public data sets, a course on practical deep learning, a hand-written text recognition library and Apple’s first research paper.
Enjoy and use the time to catch up with some paper reading ;)
As always, if you’d like to support deep learning weekly share this issue with friends and colleagues.
| Deep Patient: An Unsupervised Representation to Predict the Future of Patients from Electronic Health Records |
Fascinating nature article about the use of deep learning on electronic health records (EHRs) to advance clinical research and better inform clinical decision making.
| 2016: The Year That Deep Learning Took Over the Internet |
Interesting article that shows the world’s growing awareness of the impact of deep learning on the consumer internet. A case in point would be the recent massive improvement of Google’s machine translation capabilities which most of their users took notice of and were delighted by.
| Maluuba Releases World's Largest Human Created Question Answering Dataset to Advance Artificial Intelligence Research |
Maluuba, a Canadian Canadian deep-learning company, releases two sophisticated natural language understanding datasets. In making these resources available, the company seeks to further advance and facilitate breakthrough innovation in AI research.
| Facebook Shares Large Data Sets For Text Understanding and Reasoning |
The bAbI Project is organized towards the goal of automatic text understanding and reasoning.
| Deep Learning Race: A Survey of Industry Players’ Strategies |
An enlightening survey of the different research approaches taken by Google, facebook, Microsoft, OpenAI & co.
| Practical Deep Learning For Coders—18 hours of lessons for free |
Great, hands on deep learning MOOC by fast.ai’s. You’ll learn about CNNs, RNNs, computer vision, NLP, recommendation systems, keras, theano, and how to set up and run an Amazon P2 instance to train your models.
| Deep Learning Enables You to Hide Screen when Your Boss is Approaching |
Office Space meets Deep Learning.
| Is Deep Learning Overhyped? - Quora |
Interesting answer by François Chollet, tldr: Yes, somewhat. Deep learning doesn’t mean we have almost ‘solved’ AI and current hype runs risk of ushering in new AI winter when the tide goes out and investors sober up.
Gartner Hype Cycle
| Deep Learning at GILT |
Interesting example of how the fashion industry employs deep learning for detecting similar products and identifying facets in products (e.g. sleeve length or silhouette types in dresses). This is the sort of narrow, well-defined task which hitherto required human cognition that deep learning excels at.
| GitHub - Laia: A deep learning toolkit for HTR based on Torch |
Laia: A deep learning toolkit to transcribe handwritten text images based on Torch
| GitHub - TensorFlow implementation of the Value Iteration |
An implementation of Value Iteration Networks in TensorFlow which won the Best Paper Award at NIPS 2016
| Learning from Simulated and Unsupervised Images through Adversarial Training |
Roughly a month after Apple’s announcement that it would be publishing its AI research, their first deep learning paper has been published. When learning from synthetic images performances may be impeded by the gap between real and synthetic images. To overcome this gap the authors present a method based on Simulated+Unsupervised (S+U) learning, where the task is to learn a model to improve the realism of a simulator’s output using unlabeled real data, while preserving the annotation information from the simulator.