Google on Monday announced that it will be providing a $1 million grant to IIT Madras to set up a first-of-its-kind multi-disciplinary Center for Responsible AI to research various aspects of bias in artificial intelligence, especially from an Indian context.
Google.org will also offer a similar sum to independent non-profit Wadhwani AI to develop efforts to deploy AI models that help with “crop disease monitoring, predicting yield outcomes and bringing efficiencies to Kisan call centers,” Manish Gupta, Director of Google Research India, told indianexpress.com.
As part of Google for India 2022, the company also announced its move to contribute to the Agri stack defined by the Government of India with a “combination of remote sensing and AI to tackle problems arising in agriculture in a scalable manner”.
Gupta said Google will apply its AI models over satellite images to be able to identify things like farm boundaries, and locations of the farms and then “also develop an understanding of the agricultural landscape”. He said the plan is to go deeper and “apply additional analysis like what crop is being grown, estimate yield and so on” helping policymakers and financial institutions.
Google has already kicked off a pilot project on this with the Telangana government and has been working with a few startups on making some base layers available for the solutions. Elaborating on how the model will work, Gupta said: “We will describe the boundaries and how these are changing over some time.
One of our goals is to be able to detect specific activity like sowing, harvesting, and even stubble burning, relying on imagery over some time.” He added that within Google they will collaborate with colleagues already working on forecasting floods and pest outbreaks to bring all these capabilities together.
Then there is Google’s project Vani, which attempts to better understand India’s different languages, dialects, and accents. “With our partner, the Indian Institute of Science, we are collecting speech data from every district of India.
By collecting data in a region-wise manner, rather than based on language, we hope to cover the full linguistic diversity of India,” Gupta said, adding that this data will then be made available in open source. The idea with such an exercise, Gupta clarified, was to ensure they did not pre-manage anything and respondents could speak in any language they were comfortable with.
Along with this, Google is also improving its earlier model for Indic languages by expanding it to as many as a hundred languages where it can understand text and speech. “Again, we want it to be very inclusive.
If there’s any language that’s spoken by over one lakh people, we wanted to make sure that language is covered in our model. These are foundational capabilities that will help us again, ultimately, improve our products and democratize access to information,” Gupta added.
Google Research is also working to digitise handwritten medical prescriptions and capture the information sitting in them. Gupta said while such technology is not new, “we saw was that even when you take the state of the art systems that look at these images and do some of this handwriting recognition or optical character recognition, we found that the accuracy with which you can extract out things like even a medicine name from the medical prescriptions was very low”.
This, he said, was a fairly challenging problem that needed some research advances. “There is so much information that is sitting in the pharmacist’s heads that they used to decipher and what we have been doing is, instead of using any rules, looking at how could AI capture some of that same kind of information to improve the accuracy of your handwriting recognition or OCR models.”
