|March 20 · Issue #101 · View online |
As always happy reading and hacking. If you enjoy this newsletter, please recommend us to your friends and colleagues.
Until next week!
| NVIDIA GTC 2019 Recap |
This week NVIDIA held their annual GTC conference. The GPU maker showed off a new tool called GauGAN that can create photorealistic landscapes from simple segmentation sketches They also unveiled interactive AI Playgrounds and a host of new chips and optimizations for machine learning.
| OpenAI LP |
In case you missed it, last week OpenAI made some big changes to its structure. The entity, originally created as a non-profit, is re-organizing into what it calls a “capped-profit” company. It will (presumably) productize some of its research to generate revenue. Returns are capped at 100X which has ignited a debate about how altruistic this move truly is.
| DeepMind and Google: the battle to control artificial intelligence | 1843 |
Demis Hassabis founded a company to build the world’s most powerful AI. Then Google bought him out. Hal Hodson asks who is in charge.
| Stanford Launches Institute for Human-Centered Artificial Intelligence |
Stanford has launched a new institute focused on human-centered AI. The new entity has two directors—AI pioneer Fei-Fei Li and philosopher John Etchemendy. Stanford aims to raise $1 billion to fund research initiatives that combine policy, technology, and design.
| RNN metadata for mimicking individual author style - Gwern.net |
Gwern taught GPT-2 to write poetry. The model is even capable of generating perfect iambic pentameter. The original post linked above has a lot of implementation details and you can find a nice summary of results here
| Checklist for debugging neural networks – Towards Data Science |
Machine learning code can be notoriously difficult to debug with bugs that are expensive to chase. This checklist is a great starting point for identifying and dealing with those bugs.
| Yann LeCun's talk on "The Power and Limits of Deep Learning" |
Facebook’s Yann LeCun posts slides from two talks he gave recently at Harvard on “The Power and Limits of Deep Learning”.
| GitHub - benedekrozemberczki/Splitter: A Pytorch implementation of "Splitter: Learning Node Representations that Capture Multiple Social Contexts" (WWW 2019). |
A Pytorch implementation of “Splitter: Learning Node Representations that Capture Multiple Social Contexts” (WWW 2019). - benedekrozemberczki/Splitter
| Introducing spaCy v2.1 · Blog · Explosion AI |
spaCy, a high performance NLP framework for Python, has released version 2.1. This version brings pretrained models, better pattern matching, and faster tokenization, processing 10,000 words per second on a single CPU core.
| [1901.09491] Stiffness: A New Perspective on Generalization in Neural Networks |
Generalization is one of the most important goals of machine learning. Models that are overfit on training data aren’t valuable in production. To help understand generalization better, researchers at Google introduce the concept of “stiffness”. Similar to work on the analysis of variance, stiffness measures how gradient steps on one example impact loss on other examples.
| [1903.00812] 3D Hand Shape and Pose Estimation from a Single RGB Image |
A new graph convolutional neural network can produce 3D meshes of hand shapes and pose from a single 2D image. Feature maps from images are extracted by traditional deep CNNs and are fed into GCNNs, which directly estimate a 3D mesh for the hand. Another interesting note is that models were primarily trained on synthetic data and only fine-tuned on real data.