|November 26 · Issue #66 · View online |
Happy hacking and reading. As always if you enjoy receiving this newsletter you can help us by sharing it with friends and colleagues.
See you next week!
| High-fidelity speech synthesis with WaveNet |
DeepMind shares some details on how they managed to tune their WaveNet model performance to be production ready. Using distillation they were able to move to a more parallel network design and are now able to generate whole sentences at once, without having to wait for each individual sample.
| Wanted: AI That Can Spy |
Interesting article about the ongoing IARPA challenge on satellite image understanding. The article shows how many challenges exist in satellite imagery, why even imperfect results from deep learning may help the agencies and where this all may be going.
| AWS ramps up in AI with new consultancy services and Rekognition features |
With their big AWS conference coming up next week, Amazon has already started announcing a new AI consultancy service and some updates to their Rekognition API.
| When Creativity meets A.I. |
The image editing startup describes how they managed to bring deep learning and creativity together. And how one of your trusty curators moved from a general segmentation model to a fine-tuned architecture running in real-time on your iPhone.
| Artificial Intelligence Can Hunt Down Missile Sites in China Hundreds of Times Faster Than Humans |
More details on satellite imagery processing and how researchers managed to reduce the processing time from 60 hours, when done by humans, to 42 minutes using tuned off-the-shelf models.
| Uncovering the Intuition behind Capsule Networks and Inverse Graphics |
This is a series on the idea of ‘Capsule Networks’ that starts off with a look at CNN’s and their flaws according to Hinton. Very visual, very detailed and a pleasure to read. We can’t wait for the next part!
| Expressivity, Trainability, and Generalization in Machine Learning |
Eric Jang shares his approach on how to read papers. He explains his general categories, how he decides which papers belong into which category and demonstrates this approach on current research.
| Do the weights trained from a dataset also come under the same license terms as the dataset? |
An interesting discussion around licensing models or weights. We’d like to see some more opinions or even actual real-life experiences here, as this topic may become more and more important in the near future.
| Counting Crowds and Lines |
A great and practical application of multi-scale CNNs on a webcam image. The author nicely explains his (rather obvious) motivation, how they approached the challenge and shows results. Great read!
| Installing TensorFlow 1.4.0 on macOS with CUDA support |
This is quite a niche problem, as Nvidia GPUs on MacOS have become quite rare. But with the newly introduced eGPU support in High Sierra and machine learning engineers that use such a setup, this reference may be helpful for anyone who is thinking about using macOS and e.g. a Titan Xp.
| Training on the device |
A nice article on the concept of on-device-training. In this case, this is done using iOS frameworks, but the general ideas, limitations, and advantages may be applied to any platform. If you want to enlighten your users without worrying about privacy and a backend, here you go.
| VoxelNet: End-to-End Learning for Point Cloud Based 3D Object Detection |
Interesting work from Apple researchers. They managed to transform sparse LiDAR sensor data into isolated 3D objects using deep learning. Leading to encouraging results in 3D detection of pedestrians and cyclists, based on only LiDAR.
| Coupled Ensembles of Neural Networks |
This paper investigates the architecture of deep convolutional networks. Building on existing state of the art models, the authors propose a reconfiguration of the model parameters into several parallel branches at the global network level, with each branch being a standalone CNN.