What if you were told that you can no longer use all this data you have collected to improve your AI practice? Well, if India’s first data protection bill is turned into law, you might as well be forced to do so. So, does this mean the end of AI projects in India?
“The success of any project, whether AI-based or otherwise, depends heavily on the data it collects, processes and its legitimacy. Thus, the Personal Data Protection Bill (PDP) will probably have an impact on such projects, as the underlying principle is that data should be collected and processed only for legitimate purposes ”, underlines Manish Sehgal, partner, Deloitte India.
Another pivot point specific to AI and ML projects is about automated decision making. It is important to understand the implication of the global and local regulations that surround it and to design the system accordingly. “Global regulations such as the GDPR give individuals the right not to be subject to decisions based solely on automated processing. We will know of similar provisions, if any, in India’s PDP bill once it is enacted, ”Sehgal added.
Industry executives believe companies that have topped the charts with AI projects so far may be worse off than those that have just started just because of this looming law.
“This is an inconvenience for early adopters that many organizations now face in light of data protection law,” said Sharad Jambukar, Head-IT & IS, Aadhar Housing Finance. “For almost a year we have been trying to minimize data entry points into our systems and we have tried to ensure the minimum information required, only the data that we collect. People who have done a lot of digital transformation will have to struggle to reduce the capabilities of their AI projects in order to comply with the law. Companies that have fallen behind in the digital transformation process will have the advantage of starting their AI projects with only the required data instead of collecting too much.
Make alternative arrangements
After investing millions of dollars in an AI project, it can be difficult to shut it down completely. Experts believe there might be ways out of the situation where you could continue to run your AI projects with all the data you’ve collected so far, while still sticking to the new bill.
“If the bill passes, but says you can’t use the data directly, you can train an AI model using existing use cases the company is currently working on in the regions.” current. However, you can transfer the learning from this model to another, which is known as a weight file in technical terms. These weight files can be used to help develop future models, ”explained Rahul Vishwakarma, co-founder and CEO of Mate Labs.
According to Vishwakarma, the government can only impose restrictions on the data side, which is owned by the people, but not on trained machine learning models, which are the intellectual property of the company. “If there is no application, companies can transfer this intellectual property to different regions, subsidiaries or companies and then assemble various models to form a combined machine learning,” he proposed.
Go from collection to deletion
Once you receive a request from a customer to delete all of their records, it can be a nightmare to even locate customer data in different systems, logs, databases, or test projects. And any breach can cost your organization not only heavy penalties, but also reputational damage, which can be very difficult to repair.
Instead, it’s important to know where the data is before you even have to delete it. According to Saurabh Saxena, Site Manager and Vice President of Product Development at Intuit India, data classification is an important aspect of data protection compliance.
“We have been running a data classification exercise within the company for quite some time. This has helped us comply with the GDPR and will now make it easier to comply with the Indian Data Protection Bill. Businesses should start the process of classifying data early, as it can be a long and arduous process. During our trip, we deleted all the data we didn’t need. The process also makes it easier for our AI projects to run without any complications. We ensure that the AI projects within Intuit use personally identifiable information, ”said Saxena.
Another important aspect that needs to be looked at is controlling the amount of data entering the systems. The less you have, the easier it will be to comply with data protection law. “I think we should change internal processes so that the data required for a certain process is used only for that particular process. Rather than making an effort to protect data, it’s important to minimize the amount of data we collect. For example, in loan processing, what are the minimum data required to process the loan. At Sriram Value Services, we aim to collect as little data as possible from the customer, ”said Prashant Deshpande, VP-IT, Sriram Value Services.
One shot AI
If all else fails, you can still continue pursuing new AI projects, albeit with a different approach. About a year ago, researchers at the University of Waterloo in Ontario, Canada, pointed out that AI models can be trained on a problem using data sets as small as 1 to 10 records instead. tens of thousands and in the case even millions of records that are currently in use. The process is called “less than one” learning or LO-shot learning. With the growing number of privacy regulations not only in India but around the world, LO-shot or one-shot AI could be your last shot in AI.
“AI as we know it today is also on the verge of a radical transformation. While today we typically need thousands and in some cases millions of images to train an AI model on how to identify an object, in the future you will need a lot less. . Compare it with humans. A child doesn’t have to see a million cats before they can create the ability to identify any cat. AI is moving in the same direction, ”said K Ananth Krishnan, CTO at TCS.
Saxena from Meta Labs believes that companies could also use learnings previously implemented from one region / demographics to another. “But with this law, they will need more time to serve new geographies and demographics until they can collect all the data. To get around this limitation, new modeling methods such as One-Shot Learnings, which can be trained on a small amount of data, can be used. “