Scientific activities
Since its creation, AxTRADE has been addressing the high computational and energy cost of large neural network (NN) models by developing a method that dynamically chooses, at runtime, the smallest pretrained model capable of correctly handling each input. Instead of compressing a single network, the approach uses multiple complementary pretrained models and an input-aware neural dispatcher trained through a custom multi-objective loss. A new evolutionary algorithm explores accuracy–efficiency trade-offs and identifies Pareto-optimal configurations.
The team is also studying how approximate hardware operators can be combined with runtime reconfiguration to further improve performance and energy efficiency
Additionally, AxTRADE is currently examining the fault tolerance of Mixture-of-Experts architectures, with the goal of selectively protecting vulnerable components to improve system reliability.
Exchanges
IIT Dharwad partners visited Inria Rennes from May 19 to June 20 2025.
A visit of Inria partners to IIT Dharwad is planned for February-March 2026