This week in AI & Machine Learning: AlphaFold, AI balloons, transformers for computer vision, Yann LeCun’s deep learning course, Amazon’s ML compute chip, AI happy hour, and more!
Artificial Intelligence News:
AlphaFold: creates a solution to a 50-year-old grand challenge in biology
By far, the biggest news this week comes from the Google DeepMind team. They published the results from AlphaFold 2, their AI system, that solves the protein folding problem. Understanding protein shapes can potentially unlock treatments for diseases or development of other industry solutions that require the understanding of enzymes.
There is a lot of debate and potential hype over how impactful this breakthrough could be. Understanding the protein folding problem and its impact is complicated. So if you’re interested, I recommend watching the video below or reading this article. Both do a good job of breaking it down.
Autonomous balloons take flight with artificial intelligence
Before you get too excited, these are balloons running experiments or capturing data. It might be a while before someone creates a self driving hot air balloon taxi! But still, this application that uses reinforcement learning to make weather balloon actions autonomous is very cool. It could really help important equipment carrying balloons, like Google’s “project loon” overcome challenging weather conditions.
Google Reveals Major Hidden Weakness In Machine Learning
Google recently released a paper mapping out some areas of concern for bias in machine learning when deployed to real world applications. If you’ve been working in the field for a while many of the areas may not be “news” to you. But, I think it’s important that companies continue researching areas of weakness in AI models and formalize those findings from which everyone can learn.
Developer Tools & Education:
Transformers for Image Recognition at Scale
It’s likely you have at least heard of transformers by now and how they have greatly improved Natural Language Processing applications. Read how Google is using Visual Transformers to replace Convolutional Neural Networks (CNNs) with transformers for efficient computer vision applications.
NYU & Yann LeCun Deep Learning Course
Check out this new course on deep learning and PyTorch by none other than Yann Lecun and NYU. The format and material looks really good–a great blend of theory and application with PyTorch code.
Siamese networks with Keras, TensorFlow, and Deep Learning
This week’s PyImageSearch tutorial continues with the Siamese network theme showing how to implement them in Keras and Tensorflow.
Amazon introduces its custom ML chip: AWS Trainium
Every major tech company is getting into the machine learning computing chip game. Amazon launched its own chip designed to train and run machine learning models.
ML engineering for production ML deployments with TFX (TensorFlow Fall 2020 Updates)
Learn how Google uses TFX to put ML models in production and what’s new this fall.
Interesting Podcasts & Interviews:
Feature Stores for Accelerating AI Development | TWiML
Listen to this panel discussion to learn how to decrease time to market with MLOps, feature stores, and more.
Deep Learning for NLP: From the Trenches with Charlene Chambliss TWiML
Explore the current state of NLP, like BERT, HuggingFace, and much more.
Manolis Kellis: Meaning of Life, the Universe, and Everything | Lex Fridman
Manolis Kellis is a computational biologist at MIT. Much of this conversation is more philosophical than technical, but there may be some parts you find interesting in relation to the field of AI.
Notable Research Papers:
Some of the interesting machine learning papers published this week.
- 6.7ms on Mobile with over 78% ImageNet Accuracy: Unified Network Pruning and Architecture Search for Beyond Real-Time Mobile Acceleration