Deep Learning Weekly - 🤖 - Issue #86: Catching Deepfakes, Amazon for Government Surveillance, Google Practica, ML for Artists, current state of AI, Semantic Search and more...

The Department of Defense has started a contest to generate the most convincing AI-generated fake video, imagery, and audio, while at the same time trying to create tools that are able to distinguish such counterfeits automatically. The article explores some of the existing techniques as well and gives nice insights into the possibilities they open up.

A new player on the market, Cambricon Technologies, who has previously created AI chips for Huawei smartphones is entering the datacenter AI chip market. 

The ACLU has acquired marketing materials and documents that hint at a quite worrying development: Amazon is actively marketing it’s facial recognition system ‘Rekognition’ to governments and especially for government surveillance. There even is a list of actual customers, a handful of cities, that currently deploys the said technology.

A nice opinion piece on the current state of AI and especially its limitations. The article discusses, what we initially expected or still do expect from AI and what’s actually available in the form of applications.

Deep Learning Studio is Free, Open and no-coding platform. Developers, Researchers, & Students love this platform. Try it out (It’s Free).

After the success of their ’Machine Learning Crash Course’, Google has decided to create an even more advanced version on image classification, that includes interactive coding as well. Dive right in if you’re eager to get started with deep learning!

An amazing article taking a look at the role of machine learning for artists and how current developments can be compared to historical parallels. Starting off in the mid-1800s the article goes all the way to the author’s thesis project, which eases ML development for artists.

Hamel Husain gives a very well end-to-end example of a system that’s able to search objects semantically. The system uses sequence-to-sequence models to allow searching through code. In theory, the system will then be able to find semantically similar code blocks.

A minimal implementation of Yolov3 in PyTorch.

In this paper, the authors take a closer look at data augmentation for images and describe a simple procedure called AutoAugment to search for improved data augmentation policies. Their key insight is to create a search space of data augmentation policies, evaluating the quality of a particular policy directly on the dataset of interest.

Batch Normalization (BatchNorm) is a widely adopted technique that enables faster and more stable training of deep neural networks (DNNs). In this work, the authors demonstrate that the distributional stability of layer inputs has little to do with the success of BatchNorm.