Daily Newsletter — 13th December 2020
Generating optimized robot structures, building a gigascale ML Feature Store and creating fake Pokémon using GANs in today’s Data Science Daily Newsletter 📰
RoboGrammar: Graph Grammar for Terrain-Optimized Robot Design
RoboGrammar is a fully automated approach for generating optimized robot structures to traverse given terrains. In this framework, each robot design is represented as a graph, and use a graph grammar to express possible arrangements of physical robot assemblies. Each robot design can then be expressed as a sequence of grammar rules. Using only a small set of rules the grammar can describe hundreds of thousands of possible robot designs.
Paper: https://people.csail.mit.edu/jiex/papers/robogrammar/paper.pdf
Project Page: https://people.csail.mit.edu/jiex/papers/robogrammar/index.html
Video Explanation: https://youtu.be/RAyQYCCYRP8
Generating Fake Pokémon with a Generative Adversarial Networks
Generative Adversarial Network (GAN) is a form of unsupervised learning, where two neural networks face off in conflict with each other. For the task of image generation, the first neural network tries to generate fake images using a seed of random numbers, or even starter images.
In this article, Justin Kleiber attempts to generate fake Pokémon using a Generative Adversarial Network (GAN).
He makes use of a Kaggle dataset, with images of over 800 Pokémon. For GANs, ~800 images is a small dataset; he also makes use of PyTorch transforms to create mirrored and different coloured training images. This tripled the size of the training data with much better results.
Building a Gigascale ML Feature Store for DoorDash
DoorDash Inc. is an American food delivery service. It launched in Palo Alto, California in 2012. As of January 2020, it had the largest food delivery market share in the United States.
Imagine building a machine learning model for a company such as DoorDash with millions of users, the amount of feature data can grow to billions of records with millions actively retrieved during model inference under low latency constraints. Overcoming these challenges and warranting a deeper look into selection and design of a feature store is what is being primarily discussed in this article.
Some of the immediate goals here are to prevent overrunning cost budgets, compromising runtime performance during model inference, and curbing model deployment velocity.
Article: https://doordash.engineering/2020/11/19/building-a-gigascale-ml-feature-store-with-redis
Contact Us
Reach out to us on community@jovian.ai to get featured here. Learn data science and machine learning with free hands-on data science courses on Jovian.
Follow us on Twitter, LinkedIn, and YouTube to stay updated.
