Machine Learning and Transformers

Machine Learning at the Edge: “Edge AI” or TinyML brings on-device intelligence to microcontrollers and IoT devices. By running neural networks locally (often at <1 mW and <1 MB memory), edge ML offers low latency and privacy (no cloud needed) for applications like autonomous drones, wearables, and smart sensors. Research focuses on co-design of tiny models and hardware: techniques include model compression (quantization, pruning, compact architectures like MobileNet/EfficientNet-lite) and specialized edge accelerators (e.g. Edge TPUs). Current challenges revolve around fundamental tradeoffs in accuracy/adaptability vs. energy/memory vs. latency/speed, on computing resource limited platforms.

Analog Computing and Transformers: This area explores non-digital computing paradigms for AI. For example, neuromorphic/analog hardware (memristors, photonic synapses) can implement neural networks by physical dynamics, achieving very low-power inference. Optoelectronic transformer chips exploit optical interference to perform transformer operations with hundreds-fold better energy efficiency than electronics. Key questions include how to co-design algorithms and physical training protocols considering analog devices limitation. Transformers on edge devices have significant potential for compression of embedded protocols by eliminating power and memory overhead associated with traditional software solutions.