Pushing the Frontiers in AI For Billions Around the World

November 24, 2022

iMerit CEO Radha Basu and Meta’s Manohar Paluri discuss the ideal framework for pushing the frontiers in AI research at the ML DataOps Summit.

At the iMerit MLDataOps Summit 2022, iMerit CEO Radha Basu hosted a panel with Meta Platform’s Director of AI, Manohar Paluri, about a framework for pushing the frontiers in AI research, while advancing technology that is impactful to the product endgame. The main topics covered during this session include:

  • The structure for developing AI applications including scaling ML models
  • Adopting a multi-modal understanding
  • Pairing tools and human intelligence to accelerate AI
  • Power of human intelligence to improve AI

AI @ Billion Scale

Meta currently serves 3.75 billion people today across all of its apps, demonstrating the reality of AI @ Billion Scale. When looking at the AI landscape, Manohar presents recent changes related to text-to-image generation, speech-to-speech translation, and text-to-video generation. 

A multilingual model that translates 200 languages powers translations on Facebook. Another striking example of the innovation experienced over the past decade is Meta’s real-time speech-to-speech translation tool. This tool has the power to take the spoken language of Hokkien, which has no standard writing system, and translate it to other languages to improve communication abilities.

Finally, Manohar presented the Blenderbot conversational AI prototype, which responded to his questions related to how to prepare for his keynote speech with practical advice.

Introducing the “SMART” Framework

The SMART framework for AI is simple, yet effective. The first component is Scale. Whether a company needs to scale up its project 10x or 100x, teams must work on engineering systems with scale in mind. This includes scaling up models, data efforts, and the computing power needed to train sophisticated AI models.

The next part of the framework focuses on breaking the artificial walls between different modalities. By bridging the gaps between NLP, computer vision, speech, and audio, engineering teams are better equipped to build recommendation systems, self-driving cars, search systems, assistants, and more. Understanding that AI is multimodal is a necessity.

“You need to break the old barriers that don’t help you.”

– Manohar Paluri, Meta Platform

A key phrase used by engineering and data science teams is, “data is the new code.” Therefore, experts need to spend enough energy thinking about the data used. There are instances of the same model with the same framework improving by 2-5x in accuracy by utilizing better data and annotations. While data annotation may seem simple, the challenge is that throughout the iterative process additional problem formulation is revealed.

It’s not possible to work towards applying AI models in today’s world without considering the potential negative outcomes. While AI was predominantly worked on in the lab, the past decade proved that moving to production yields significant positive and negative impacts on society. Researchers must ask, “How can I carry out my work responsibly?”

“The tradeoffs that you make will dictate the outcomes that you will achieve.”

– Manohar Paluri, Meta Platform

Finally, the barrier to entering the field of AI is lower as a result of the tools and frameworks available today. New tools empower data scientists to act as catalysts in a much faster way. At the same time, proven practices act as safeguards by preventing unnecessary changes to learning rates or mixing training and testing data.

Where is AI Going?

The future of AI promises to emphasize the importance of internalizing the allocation of resources spent on examining data, collecting data in the right way, and making sure the distribution of data is respectful of the outcomes desired. There are multiple ways to go about this:

  • Education – focus on educating the team about the importance of responsible AI, spending enough energy on these efforts, and building the right tools at scale.
  • Agile Development – Using best practices from software engineering and applying them to data engineering is the key to Agile development. This involves iterating on a daily or even hourly basis to train a model and improve it in small steps. 
  • Cross-functional Teams – Each individual working on AI models should go through training in stats, mathematical models, and other specialties. This results in a well-rounded team of statisticians, policymakers, product managers, researchers, data scientists, analysts, and more. 

As companies adopt a multi-modal understanding of artificial intelligence, it’s crucial to develop a structure for model deployment and scaling. Manohar Paluri’s keynote speech focused on how companies can go about achieving this, including following the SMART framework. With further innovations in the field approaching quickly, adopting software engineering best practices can make the difference between successful AI applications and those that fail. To enhance the quality of your data and continue to improve your model, consider working with a data annotation company like iMerit.

If you’d like to learn more about our services, contact us today to talk to an expert.