‘Race on to build bigger and better pre-trained generalised AI models’

The specialised models using this approach can be built at a fraction of cost and time, and require relatively less data.
Companies are now exploring specialised artificial intelligence models.
Companies are now exploring specialised artificial intelligence models. (Representational image)

BENGALURU: Technology has been playing a critical role in driving the country’s economic growth. If digital public infrastructure helped launch societal and e-governance initiatives, now AI will further accelerate the economy by unlocking higher levels of productivity, says Mohammed Rafee Tarafdar, chief technology officer at Infosys in an interaction with TNIE. Excerpts:

Companies are now exploring specialised artificial intelligence models. What are its advantages and what are the challenges in building them?

Today, there is a race to build bigger and more efficient pre-trained generalised models. However, I believe specialised models are more suitable for enterprise needs. These can either be pre-trained and built grounds up for a specific domain or task, or they could be developed using an open-source model as base and then by fine-tuning them with industry or organisation data. The cost and time for building a grounds-up specialised model is very high. It also requires a significant amount of high quality data.

Mohammed Rafee Tarafdar
Mohammed Rafee Tarafdar

Can you explain specialised models that are built at Infosys?

At Infosys, we have been building specialised models using a narrow transformer approach, wherein we use a suitable open-source model as base model and then fine-tune it with organisational data for a specific task, operation, or knowledge. The specialised models using this approach can be built at a fraction of cost and time, and require relatively less data. Since the specialised models are built for a specific task or purpose, they are a lot more effective for the task and perform much better. The other key challenge with both specialised and generalised models are high running costs and the need to minimise hallucinations and being responsible by design.

Do we have enough talent to explore AI and other tech spaces?

We have enough talent but not all of them may be experts on AI and other tech spaces like Quantum, Metaverse and Green computing. If you look at AI, we have talent that has already been working on data science, analytics, machine learning, and deep learning. Now, they have to be further skilled on generative AI, foundation models, and responsible AI. With the AI enablement programs launched by every tech company, we will see a lot of AI aware, AI builder, and AI masters in 2024.

Can you also explain how Infosys is evaluating its AI use cases considering regulatory and security aspects of AI?

One of the core principles of our AI-first approach is to be ‘responsible by design’. To implement this, we have enhanced our existing frameworks to cover 12 critical areas. These include regulatory compliance for AI, reproducibility and interpretability, among others. We are using the framework to evaluate all AI use cases and ensure that risks and mitigation strategies are planned and discussed with all stakeholders. In addition, we have started working on codifying the policies into the AI engineering lifecycle to ensure automation and compliance.

Related Stories

No stories found.
The New Indian Express
www.newindianexpress.com