Data Phoenix Digest - ISSUE 36

Data Phoenix team looking for speakers, MLCommons presents free and open-source datasets for speech recognition, accelerating inference up to 6x, document understanding transformer without OCR, GradInit, NÜWA, videos, jobs, and more ...

Dmitry Spodarets
💡
Data Phoenix team plans to renew our "The A-Z of Data" webinars just after the holiday season. We’re looking for speakers to collaborate on this project. If you are looking for the platform with a relevant audience, and you have experience and knowledge to share, we’ll be glad to see you among the speakers.
For details and terms, contact us at [email protected]

NEWS

What's new this week?

  • Eastern European bank creates AI supercomputer in collaboration with the Hungarian government.
  • BrainBox AI innovates the technology to manage the mechanical side of the buildings, including ventilation, temperature, and C02 control.
  • MLCommons presents free, open-source datasets for speech recognition. They were created by an international group from the US, South America, and China.
  • Accurate AI prediction even with a small number of experiments. Learn how this new, versatile AI technique was developer by the Japanese researchers.
  • Scared about the threat of AI? It’s the big tech giants that enable the spread of extremism and social chaos need reining in! The case for regulating them is clear.

Funding News

  • Elementary, an AI startup that enables customers to more easily inspect manufactured goods, raises $30M in a series B funding round led by Tiger Global.
  • Wonder Dynamic, a developer of “blockbuster-level” visual effects technologies enabled by AI and cloud services, raises $10M in a series A funding round.
  • ZenML, a startup providing an extensible and open source MLOps framework to accelerate and simplify the delivery of ML models, raises $2.7M in a seed round of funding.

ARTICLES

Learning Rates for Deep Learning Models
How can you make your deep learning models as effective as possible? In this article, you'll learn about the effects of a learning rate on the convergence and performance of DL models.

Visualizing the Vanishing Gradient Problem
In this tutorial, you'll learn why vanishing gradient problem exists, including its 101, do's and don'ts, and all about configurations of neural networks susceptible to vanishing gradient.

Meta-Learning for Keyphrase Extraction
This article explores how to build a keyphrase extractor that performs on in-domain data and in zero-shot scenarios where keyphrases need to be extracted from data with different distribution.

Accelerating Inference Up to 6x Faster in PyTorch with Torch-TensorRT
Torch-TensorRT is the new integration of PyTorch with NVIDIA TensorRT, which accelerates the inference with one line of code. Learn how you can start using it today!

A Beginner’s Guide to End to End Machine Learning
In this comprehensive guide, you'll learn how to train, tune, deploy, and monitor models, from collecting and prepping data to monitoring your models in production.

PAPERS

NÜWA: Visual Synthesis Pre-training for Neural Visual World Creation
NÜWA is a unified multimodal pre-trained model  that can generate new or manipulate existing visual data (i.e., images and videos) for various visual synthesis tasks.

GradInit: Learning to Initialize Neural Networks for Stable and Efficient Training
GradInit is an automated and architecture agnostic method for initializing neural networks. It improves the stability of the original Transformer architecture for machine translation.

End-to-End Referring Video Object Segmentation with Multimodal Transformers
This paper presents Multimodal Tracking Transformer (MTTR) that models the RVOS task as a sequence prediction problem. It simplifies the RVOS pipeline compared to existing methods.

Donut: Document Understanding Transformer without OCR
Donut is a novel VDU model that is end-to-end trainable without OCR framework designed to pre-train the model to mitigate the dependencies on large-scale real document images.

VIDEOS

NeX: Real-time View Synthesis with Neural Basis Expansion
In this video, you'll learn about an approach that uses a modification of MPI to model view-dependent effects by parameterizing each pixel as a linear combination of basis functions.

JOBS

Looking to feature your open positions in the digest? Kindly reach out to us at [email protected] for details. We'll be proud to help your business thrive!

Digest