Deep Learning Weekly Issue #164

Analyzing medical scans with AI, an AI-powered weight training startup, TensorFlow Lite for NLP, and more

Hey folks,

This week in deep learning we bring you Microsoft's new feature that uses AI to make video chat less weird, and fundraising announcements from DeepCube for their software-based solution that accelerates AI on existing hardware, RapidAI to analyze medical scans with AI algorithms, and Tonal for their AI-powered weight training startup.

You may also enjoy learning about how AI planners in Minecraft could help machines design better cities, what’s new in TensorFlow Lite for NLP, array programming with NumPy, and more!

As always, happy reading and hacking. If you have something you think should be in next week's issue, find us on Twitter: @dl_weekly.

Until next week!


AI planners in Minecraft could help machines design better cities

A competition to see which AI produces the best settlements in the game is helping to explore new techniques for urban planning.

Photoshop’s sky replacement tool makes it easier to fake a perfect sunset

An influencer’s best friend.

DeepCube’s software-based solution accelerates AI on existing hardware

DeepCube, a startup developing a platform that reduces the computational requirements of AI algorithms on existing hardware, raised $7 million.

RapidAI raises $25 million to analyze medical scans with AI algorithms

RapidAI (formerly iSchemaView), a startup developing a portfolio of stroke imaging and assessment products, today closed a $25 million round.

AI-powered weight training startup Tonal raises $110 million amid home workout boom

AI gym equipment startup Tonal nabbed a $110 million round from new and existing investors.

Mobile + Edge

Create Homemade Recipes of Your Favorite Products on iOS Using Fritz AI Studio

Leverage Fritz AI to quickly generate a dataset and train an iOS-ready object detection model.

How Augmented Reality Is Becoming A Vital Tool In Style

As brand’s look to new and creative ways to engage a bigger pool of customers – namely Millennials and Generation Z kids – AR technology is being adapted to the fashion sphere.

What’s new in TensorFlow Lite for NLP

In this article, you’ll be introduced to end-to-end support for NLP tasks on TFLite. Learn about new features including pre-trained NLP models, model creation, conversion and deployment on edge devices.

Microsoft's new feature uses AI to make video chat less weird

How Microsoft's Surface Pro X uses driver-based software and dedicated AI hardware to make video chat more people-friendly.

This Raspberry Pi-powered AI helps robots to sort through your recycling

A team of engineers built an AI system based on a Raspberry Pi to automate recycling.


Advancing NLP with Efficient Projection-Based Model Architectures

Google AI describes an NLP model that achieves near BERT-level performance on text classification tasks, while using orders of magnitude fewer model parameters.

Animations of Neural Networks Transforming Data

In this post, the author shows how to use animations to develop an intuition for how a simple neural network transforms the input data.

Improving Sparse Training with RigL

RigL is a new algorithm for training sparse neural networks. Instead of pruning a pre-existing dense network, it dynamically builds one during training without sacrificing accuracy relative to traditional approaches.

Libraries & Code

[GitHub] DLR-RM/BlenderProc

A procedural Blender pipeline for photorealistic training image generation.

[GitHub] spellml/paint-with-ml

With a little help from deep learning, now you too can create your own happy accidents.

Papers & Publications

Array programming with NumPy

Abstract: Array programming provides a powerful, compact and expressive syntax for accessing, manipulating and operating on data in vectors, matrices and higher-dimensional arrays. NumPy is the primary array programming library for the Python language. It has an essential role in research analysis pipelines in fields as diverse as physics, chemistry, astronomy, geoscience, biology, psychology, materials science, engineering, finance and economics. For example, in astronomy, NumPy was an important part of the software stack used in the discovery of gravitational waves1 and in the first imaging of a black hole2. Here we review how a few fundamental array concepts lead to a simple and powerful programming paradigm for organizing, exploring and analysing scientific data. NumPy is the foundation upon which the scientific Python ecosystem is constructed. It is so pervasive that several projects, targeting audiences with specialized needs, have developed their own NumPy-like interfaces and array objects. Owing to its central position in the ecosystem, NumPy increasingly acts as an interoperability layer between such array computation libraries and, together with its application programming interface (API), provides a flexible framework to support the next decade of scientific and industrial analysis.

Decoupling Representation Learning from Reinforcement Learning

Abstract: In an effort to overcome limitations of reward-driven feature learning in deep reinforcement learning (RL) from images, we propose decoupling representation learning from policy learning. To this end, we introduce a new unsupervised learning (UL) task, called Augmented Temporal Contrast (ATC), which trains a convolutional encoder to associate pairs of observations separated by a short time difference, under image augmentations and using a contrastive loss. In online RL experiments, we show that training the encoder exclusively using ATC matches or outperforms end-to-end RL in most environments. Additionally, we benchmark several leading UL algorithms by pre-training encoders on expert demonstrations and using them, with weights frozen, in RL agents; we find that agents using ATC-trained encoders outperform all others. We also train multi-task encoders on data from multiple environments and show generalization to different downstream RL tasks. Finally, we ablate components of ATC, and introduce a new data augmentation to enable replay of (compressed) latent images from pre-trained encoders when RL requires augmentation. Our experiments span visually diverse RL benchmarks in DeepMind Control, DeepMind Lab, and Atari, and our complete code is available at this https URL.