The fourth edition of the Machine Learning Developers Summit (MLDS) aims to bring together India’s leading experts in the machine learning domain. In an interesting tech talk on the first day of the summit, titled “Move over ML, It’s Time for XL” by Soumendra Mohanty,  Chief Strategy Officer & Chief Innovation Officer at Tredence, stressed on how energies should be spent on “experience learning”, rather than just on building machine learning algorithms.

“The intent is not to forget machine learning concepts and the whole thing about algorithms, data, and cloud. Over the last two decades or so, there has been so much advancements and experimentation in this area of machine learning, but an area that has been neglected is what and for whom we are doing all of this,” Mohanty said.

Divide between IT & data science departments

Mohanty pointed out around 90 percent of machine learning models never make it to production. “ML engineers, data scientists, data engineers, IT professionals, etc don’t speak the same language. There is a major disconnect between what they can do together. In many companies, there is a fundamental divide between the IT and data science departments. IT tends to prioritise making things work and keeping them stable. Data scientists, on the other hand, like experimenting and breaking things. This does not lead to effective communication,”Mohanty said.

The challenges in implementing ML models include:

• Scaling up is harder than you think
• Efforts get duplicated
• Executives don’t always buy in
• Lack of cross language and framework support
• Versioning and reproducibility remain challenging

Mohanty raised three critical points to improve the user experience. 

How quickly is machine learning moving along?

He said the industry is moving forward very fast: Six months is like six years in other fields. A lot of emphasis is put on building more powerful machines to crunch more data and run algorithms faster. The emphasis now is more on machines than on learning, he added.

Biggest machine learning challenge at the moment 

The modeling part of machine learning has evolved much faster and better. But is there a holistic framework that everybody is subscribing to? “Lack of tooling and methodology around production ML is the biggest thing holding back the real world impact of the field,” said Mohanty.

The imminent game changer for machine learning

In terms of solving problems from the end-user perspective, tech giants have stepped up and are involving mechanisms to solve the challenges users can face. Google is championing human centered machine learning framework and 

Microsoft Research’s new Human-AI eXperience toolkit helps users to choose how they want to work with AI. 

“At one end is machine learning impact and at the other end is the user impact where we are doing something for a user who can apply these things to solve real-world problems. When I talk about problem solving, I look at it from a user perspective as to what might help you to do something that you were not able to do today. The starting point of the problem statement should be this. Then, the stakeholders should bring the capabilities of algorithms and models into play, not the other way around,” he said.

Goal-driven systems

Of the seven patterns of AI, the “goal-driven systems” has not made much progress, Mohanty said. Goal -Driven systems essentially means  to derive goals, learn from one step to another, do adjustments and keep solving the problem. 

On a positive note, Goal-driven systems and reinforcement learning are making inroads into gaming, military simulations, and surgeries.

“Unless and until we design AI solutions with a better Human-AI interaction, we will not get more involved users and the cycle will not improvise over and over again,” said Mohanty.

Human-AI interaction design woes

The difficulties in designing Human-AI interaction systems include:

• It is difficult to articulate what AI can/ cannot do
• It is difficult to foresee the potential effects of AI
• We often do not know whom to hold accountable for AI errors
• It is not easy to explain AI behaviour to users

What is XL?

XL is a set of best practices designed to help AI creators, including UX, AI, project management, and engineering teams to design, develop, and govern AI solutions keeping the outcome and end-user experience at the centre of everything.

Design thinking paradigm

Mohanty expanded on the major guidelines to keep in mind while building a product that enhances user experience:

• Who is my user? What matters to this person?
• What are their needs?
• Brainstorm creative solutions
• Rapid prototype to test your solutions
• What works and what doesn’t?

ML solution as a product

Human-AI interaction centric products put the user at the centre of the initiative instead of technology. “It takes a holistic view of the user experience, needs, interactions, usage, change management and adoption and then lays the technology underpinnings to deliver a solution that is always on Beta eventually reaching the final stage as a product,” he said.

Next generation AI

Mohanty wrapped up his talk touching on the major developments lining up in the field of artificial intelligence:

• Unsupervised learning as the next great frontier in AI
• Privacy-preserving AI-The most promising approach could be federated learning.
• Transformers’ great innovation will make language processing parallelised