Generative artificial intelligence (AI) research projects are gaining traction in top engineering colleges in India, as researchers look to understand how technology can be utilized to create tools similar to OpenAI’s ChatGPT, but in Indian languages. With companies like Microsoft and Google incorporating generative AI platforms into their services, the Ministry of Electronics and Information Technology (MeitY) has also recognized the potential of AI as a catalyst for growth in India.
However, researchers at these academic institutes face several challenges in their generative AI projects. Sourcing ample data of Indic languages, which are diverse and complex, poses a significant hurdle. The cost of such projects, including data collection and processing, infrastructure, and computing power, can also be prohibitive. Additionally, the scale of computing power required for training large AI models can be a limitation in academic settings.
Despite these challenges, Indian researchers have been actively working on generative AI projects for over three years, recognizing the importance of creating AI tools that are tailored to Indian languages. These projects aim to leverage the potential of generative AI to address specific needs and challenges in the Indian context, and contribute to the development of AI technology that is relevant, inclusive, and impactful for the Indian population.
Tapas Kumar Mishra, an assistant professor of computer science engineering at the National Institute of Technology (NIT), Rourkela, highlighted that academia is utilizing techniques from language models, specifically the transformer architecture, for various tasks such as data classification, question answering, machine translation, and building chatbots. While global platforms predominantly work in English, Mishra’s team at NIT Rourkela is focusing on languages like Hindi, Bangla, and Kannada, creating models that can process questions in these languages and generate output in English. They have achieved impressive scores of 25 to 30 on Hindi to English and 19 on Bangla to English, as per the industry standard BiLingual Evaluation Understudy (BLEU) test.
NIT Rourkela recently published a research paper on translations from Hindi to English in collaboration with the Association for Computing Machinery, a renowned US-based scientific educational community that publishes research work on natural language processing (NLP). These efforts showcase the progress and potential of generative AI research in Indian languages, with academia actively contributing to the advancement of AI technology in the Indian context.
Researchers in academia in India are utilizing transformer architecture, such as the one used in OpenAI’s GPT models, for various tasks including classification, question answering, machine translation, and chatbot development. For instance, at the National Institute of Technology (NIT) Rourkela, researchers are working on creating language models that can take questions in languages like Hindi, Bangla, and Kannada, and generate output in English. They have achieved impressive scores of 25 to 30 on Hindi to English translation and 19 on Bangla to English translation, as measured by the industry standard BiLingual Evaluation Understudy (BLEU) test.
Similarly, students from the Indian Institute of Technology (IIT) Madras have taken up projects to develop better translated YouTube videos in Tamil, and compare their own research language models with GPT-4. Meanwhile, researchers at IIT Guwahati are working on creating affordable visual animation models that study eyes and facial movements from open-source visual databases. At IIT Delhi, a team of students and professors have created a language model called ‘MatSciBert’ specifically for the field of material science research, with the goal of discovering new materials with the help of AI.
However, one of the challenges faced by researchers in India is the requirement of high computing power. The cost of training large language models like GPT-3 can be prohibitive, with one training run estimated to cost $4.6 million, making it difficult for academic institutions and smaller companies to afford such resources. Access to India’s supercomputer infrastructure, owned by the Centre for Development of Advanced Computing (C-DAC), is also reportedly lacking clarity and progress. Additionally, availability of data, especially for low-resource languages, remains a challenge, although initiatives like ‘Bhashini’, an Indic language database launched by the Ministry of Electronics and Information Technology (Meity) in India, aim to address this issue.
Despite the challenges, researchers in academia in India are making significant progress in utilizing language models for various applications, and their work has the potential to contribute to the advancement of AI research in the country and beyond.
For more news click here!
