Skip to yearly menu bar Skip to main content


Poster
in
Affinity Workshop: Global South AI

Requirement for Machine Unlearning Techniques for Kannada Language

Yashaswini Viswanath · Vishwanath Hulipalled · Vanama Yaswanth

Keywords: [ Machine unlearning ]


Abstract:

In a Traditional Machine Learning where machine are capable of learning is what the world has seen in the last few years. Current research area of Unlearning researchers are defining that machine should be capable of Unlearning too, whichwill help us make sure we get good accurate content in the model's learning. LLMs have been trained on Kannada data that is out in the open. These Generative AI models which spew Kannada content while the temperature is varied, is not trained on curated Kannada literature and content. This results in wrong learning of Kannada language and it's context and nuanaces which are based on local culture. The case of going wrong is high and just as second language learners are corrected by native speakers, we as native Kannada speakers have a prerogative to have access to info on what content the Generative AI model was trained on. The main strength we need to excercise is that of Machine Unlearning when the mistakes are found. The Machine Unlearning techniques that are beingĀ  developed have catered to the North and predominently English. This abstract is about brining awarenes that we need research in the direction of Machine Unlearning in Kannada.Why does this needs to be showcased at NeurIPS?1. We are proposing language based unlearning2. We are bringing awareness to Global South languages

Chat is not available.