Full-Stack AI Systems Development with Murali Akula - #563

Federated Learning

Federated learning systems are transforming how AI models are trained by leveraging local data on devices. Murali Akula explains that traditionally, AI training was centralized, but now, training on devices is becoming more prevalent for personalization purposes. This approach allows models to be tailored to individual users without data leaving their devices, enhancing privacy and efficiency.

We implemented a federated learning system for distributed training completely on the devices. Nothing is happening on the server except for getting the incremental updates from various devices and combining them.

--- Murali Akula

This system supports both PyTorch and TensorFlow, enabling scalability across thousands of devices 1.

   

Decentralized Training

Decentralized training represents a significant shift from traditional centralized AI training methods. Murali Akula highlights the challenges and breakthroughs in implementing parallel processing on devices, which allows for more efficient AI model training. By exploiting the parallelism in neural networks, Qualcomm's team has managed to overcome bottlenecks in entropy encoding and decoding, enabling real-time processing on Snapdragon platforms.

We were able to implement the parallel entropy decoder on the CPU and match the parallelism which is happening on the accelerator.

--- Murali Akula

This advancement allows AI models to be trained and deployed more effectively on resource-constrained devices 2.