|January 25 · Issue #72 · View online |
And as we just crossed 7000 subscribers, we would like to thank you once again for all of your support. As always, if you want to help us grow this great community of deep learning enthusiasts, simply share this issue with friends and colleagues.
See you next week!
| Yann LeCun is stepping down as Facebook's head of AI research and into a new role |
After founding FAIR in 2013 and leading the research group since then, Yann LeCun is succeeded by Jérôme Pesenti and will now serve as FAIR’s chief AI scientist. This change is apparently lead by Facebooks attempt to put more AI into its News Feed and other products.
| What are the Top 10 problems in Machine Learning for 2017? |
Although targeting the past year, this question is still up to date and the answers contain valuable information on all the things we haven’t solved yet. This includes memory networks, reasoning, image understanding and much much more.
| Mustafa Suleyman: The liberal activist ensuring Google DeepMind benefits all of humanity |
A quite personal, but nonetheless interesting look at the history of Mustafa Suleyman, one of three founders of DeepMind, now owned by Google. The article focuses on DeepMinds history and the overall intention behind its incorporation, as well as recent events and mysteries around ethics.
| Understanding Learning Rates and How It Improves Performance in Deep Learning |
However you step into the wonderful world of deep learning, the learning rate will always appear on your journey. This article distills a series of previous works and describes why we need a learning rate, how to choose one, it’s role during training, and how to handle it when using pretrained weights.
| How to do machine learning efficiently |
A quick series of tips, one more valuable than the other. Starting off with his ‘10 second rule’, Radek describes how to reduce friction when experimenting by ensuring low execution times, testing on subsets of your data, measuring timing and testing everything. Definitely worth a look!
| Faster R-CNN: Down the rabbit hole of modern object detection |
The folks at Tryolabs sat down to collect all the insights into the Faster R-CNN architecture they gained while working on their own deep learning toolkit for computer vision
. The resulting article explains the whole system in detail and should make it easy to understand all the nitty-gritty details, thanks to lots of visualizations.
| Reprojection Losses: Deep Learning Surpassing Classical Geometry in Computer Vision? |
Once again thinking about geometry, Alex Kendall shares his thoughts about the role of reprojection losses for unsupervised learning in computer vision. He describes how such losses can be used to learn from stereo images and why this allows self-supervised learning. He then takes a deep look at some short-comings and how the approach may be improved upon.
| Speedy Neural Networks for Smart Auto-Cropping of Images |
Twitter rolled out an optimized cropping algorithm, which generates better image previews for our tweets
in your timeline. They describe how they use saliency to detect interesting image areas and how they then reduced their model using distillation and pruning. The corresponding paper has been released as well.
| Detectron: FAIR's research platform |
Facebook has open-sourced the software system used at FAIR. Detectron, obviously based on Caffe2, is a high-quality, high-performance codebase for object detection research. It’s designed to be flexible in order to support rapid implementation and evaluation of novel research. Detectron includes implementations for a series of recent systems (RetinaNet
, Faster R-CNN
| Deep Neuroevolution |
| DeepLeague: leveraging computer vision and deep learning on the League of Legends mini map |
In this awesome series, Farza explains in detail how he used deep learning and computer vision to constantly track the position of heroes on the League of Legends minimap. The resulting code and an impressive dataset of 100k labeled LoL minimaps are available as well, so if you’re an esports fan, this should catch your attention.
| Inverted Residuals and Linear Bottlenecks: Mobile Networks for Classification, Detection and Segmentation |
Google has presented the next iteration of their MobileNet approach, once again trying to balance accuracy and model size using nifty architectural designs and exploration.
| DroNet: Learning to Fly by Driving |
Fascinating work on using a neural network to control a drone on city streets. To allow training of the network without actually having to manually fly drones through cities, all data was collected using cars and bicycles.
| Audio Adversarial Examples: Targeted Attacks on Speech-to-Text |
Very worrying work, where the authors managed to construct targeted audio adversarial examples on automatic speech recognition. They’re able to produce a matching waveform for any given sample, that’s 99.9% similar but transcribes something else and is able to fool Mozilla’s DeepSpeech implementation with 100% success rate.