Data Phoenix Digest - 22.07.2021

AI in gaming and energy, an overview of the CVPR 2021, CI/CD for ML online serving and models, BlenderBot 2.0, JupyterLite, TensorRT 8, Color2Style, videos from LOGML summer school, an introduction to Active Learning, datasets, jobs, and more...

Dmitry Spodarets


What's new this week?

Faster AI inferences. Robots that can touch you. AI in gaming and energy. AI that flips veggie burgers and more.

AI evolves, and so do models. They are growing increasingly complex to meet the demands of real-time DL applications. The release of TensorRT 8, the latest version of NVIDIA's SDK for AI/ML inferences that power search engines, ad recommendations, and chatbots. TensorRT 8 can cut inference time in half for language queries compared with the previous release of TensorRT.

AI isn't just about AI brain. In many ways, robotics provides a more serious engineering challenge than compute and inferences. For example, it's much harder than it seems to build a robot that can safely touch you. You don't want it to break your neck, and so do the researchers and engineers from MIT. Maybe one day robots like that will be able to cook a tasty burger for you; for now, AI is just used to find the perfect balance of taste for veggie burgers.

No worries, though. If you are in the burger business, robots won't replace you. Same as they don't replace humans in the energy industry, or in any other industry that is riddled with too many mundane tasks that can (and should be) automated. (At least, that's what the industry wants us to believe. Spoooky!)

And finally, gaming news! Sony AI plans to collaborate with PlayStation developers to create intelligent computer-controlled characters. They'll be powered by reinforcement learning. The question now is, will these characters be diverse enough?


CVPR 2021: An Overview
At this page, you'll find materials and the in-detail overview of the 2021 CVPR conference, one of the main computer vision and machine learning conferences. Carve out some free time and make sure that you dig in — it's truly a collection of gems.

JupyterLite: Jupyter, WebAssembly, and Python
A succinct overview of JupyterLite, a JupyterLab distribution that runs entirely in the web browser, backed by in-browser language kernels. The article explains why it's a nice-to-have solution and how to use it. It offers several use cases to try it out online.

Blender Bot 2.0: An Open Source Chatbot that Builds Long-Term Memory and Searches the Internet
In this article by Facebook AI, you'll learn about BlenderBot 2.0, the first chatbot that can simultaneously build long-term memory, search the internet for timely information, and have sophisticated conversations on nearly any topic.

Alien Dreams: An Emerging Art Scene
In this article, you’ll learn about new methods of creating art with AI. Specifically, the author explores the evolution of CLIP-based generative art. CLIP is Open AI’s NLP tool, allowing artists to create all sorts of interesting visual art merely by inputting some text.

Choosing the Right Machine Learning Approach for Your Application
In this article, Lak Lakshmanan, Director of Analytics & AI Solutions at Google Cloud, puts out a framework for choosing the right ML approach for your application, from working with datasets to navigating the intricacies of specific technology stacks.

Continuous Integration and Deployment for Machine Learning Online Serving and Models
In this article, the team of Uber Engineering shares their vision on resolving some of MLOps challenges (e.g. a large volume of model deployments). As Uber’s ML infrastructure keeps evolving to support new ML use cases, they see new MLOps challenges emerge.

State of Computer Vision — CVPR 2021
Another detailed overview of the 2021 CVPR conference featuring recent trends, CV learning examples, new CV learning methods, vision language models, and more. The article ends with several specific use cases of learning on limited data.

Tesla AI Chief Explains Why Self-Driving Cars Don’t Need Lidar
In this article, Andrej Karpathy, the director of artificial intelligence and Autopilot Vision at Tesla, explains why Tesla relies on the pure vision-based approach to autonomous driving. The talk was given at this year’s Conference on Computer Vision and Pattern Recognition.


Color2Style: Real-Time Exemplar-Based Image Colorization with Self-Reference Learning and Deep Feature Modulation
In this paper, Hengyuan Zhao et al. present Color2Style, a deep exemplar-based image colorization approach to resurrect grayscale image media by filling them with vibrant colors. The algorithm design is much more lightweight and intelligible, achieving appealing performance with real-time processing speed. The model does not require multifarious loss functions and regularization terms like existing methods, but only loss functions.

Per-Pixel Classification Is Not All You Need for Semantic Segmentation
MaskFormer is a simple mask classification model designed to predict a set of binary masks, each associated with a single global class label prediction. The proposed method simplifies effective approaches to semantic and panoptic segmentation tasks and shows excellent empirical results.

MobileDets: Searching for Object Detection Architectures for Mobile Accelerators
In this paper, the researchers investigate the optimality of the design of inverted bottleneck layers, to discover that regular convolutions are a potent component to boost the latency-accuracy trade-off for object detection on accelerators.

Deep Learning based Food Instance Segmentation using Synthetic Data
This paper proposes a food segmentation method applicable to real-world through synthetic data. The team generates synthetic data using the open-source 3D graphics software Blender placing multiple objects on meal plate and train Mask R-CNN for instance segmentation.


LOGML Summer School [Videos]
Here you'll find a collection of videos from LOGML Summer School (London Geometry and Machine Learning). The videos encompass a wide range of topics, from deep 3D generative modeling to tropical support vector machines and graph ML.


An Introduction to Active Learning - Jennifer Prendki, PhD | ODSC West
A talk from ODSC West featuring Jennifer Prendki, PhD. You'll learn about the secrets to training the ultimate dataset, starting with data labeling, common challenges of data, and finishing with an overview of active learning for data.


WebFace260M is a new million-scale face benchmark designed to help the research community to more efficiently close data gap and provide an easier access to various datasets. Make sure that you check out a preprint of WebFace260M benchmark paper at arxiv.

FakeCatcher: Detection of Synthetic Portrait Videos using Biological Signals
Deep Fakes Dataset is a dataset of “in the wild” portrait videos. They originate from various sources such as news articles, forums, apps, and research presentations. In total, it features up to 142 videos, 32 minutes, and 17 GBs.