Earlier this year, Google’s AI researchers recently released a paper introducing Natural Questions (NQ), a new dataset for QA research, along with methods for QA system evaluation.
Also, there is the Bidirectional Encoder Representations from Transformers or BERT, which was open sourced last year, offers a new ground to embattle the intricacies involved in understanding the language models.
Whereas, Facebook demonstrated how serious they are about AI by open sourcing their NLP toolkit LASER.
One thing common with these tech giants is their willingness to open source their innovations. Their belief in accelerated innovation through transparency has started to see fruition in the form of diversified real world applications .
This week too, AI research powerhouses Google and Facebook continued their tradition of open sourcing in-house innovations to the public.
Facebook’s BoTorch And Ax
At the recently concluded 2-day event F8 Developer Conference held in San Jose,California, Facebook dished out more updates for the machine learning developers. BoTorch and Ax are two tools which open sourced this week.
BoTorch is a research framework built on top of PyTorch to provide Bayesian optimization, a sample-efficient technique for sequential optimization of costly-to-evaluate black-box functions. BoTorch advances the state of the art in Bayesian optimization research by leveraging the features of PyTorch, including auto-differentiation, massive parallelism, and deep learning.
Ax is an ML platform for managing adaptive experiments. It enables researchers and engineers to systematically explore large configuration spaces in order to optimize machine learning models, infrastructure, and products. At Facebook, Ax has been broadly applied by engineers who do not have extensive experience with machine learning, as well as by AI researchers.
Ax provides easy-to-use APIs to interface with BoTorch, along with the management necessary for production-ready services and reproducible research.
The above figure is a glimpse at how BoTorch and Ax can be deployed in an ML ecosystem for a variety of applications.
Google Releases Landmark Dataset
Google’s AI research has always been directed towards reaching AGI. Be it the tossing bot or the convolutional networks used on its Pixel phones, there is always a push to reach human level understanding of the surroundings. In a similar effort, Google experiments with landmark recognition. Humans have a smarter way to remember things or associate places with objects for faster retrieval. Hearing the word Delhi might remind us of Taj Mahal or looking at a pyramid can remind one of Egypt.
Google-Landmarks-v2 is a newer, larger version of last year’s landmark dataset. This new dataset has over 5 million images of more than 200 thousand different landmarks.
To prepare this dataset, Google crowdsourced the landmark labeling through the efforts of a world-spanning community of hobby photographers, each familiar with the landmarks in their region.
Along with releasing the dataset Google also launched Kaggle competitions based on the same task of landmark recognition and retrieval. This is an attempt to improve the machine learning models in the field of instance recognition.
To supplement their own research and that of the enthusiasts, Google is now open sourcing new technique called Detect-to-Retrieve. A method where extra weight is given to specific regions in an image for improving accuracy.
Know more about this dataset here