Research

Mnemosyne is a speculative memoization framework that exploits temporal redundancy across consecutive LiDAR frames. A lightweight inter-frame similarity checker and motion compensation enable reuse of detection outputs or intermediate activations.

Across five model-dataset benchmarks, Mnemosyne achieves 5.8× average speedup, up to 10.2× peak speedup, and a 2.39× energy reduction with approximately 10% accuracy loss.

AutoProbe is a multi-level profiling framework combining ROS 2 pipeline tracing, LLVM-based architecture-agnostic instrumentation, and hardware-specific binary profiling.

The framework reconstructs Autoware task graphs and critical execution paths, measures latency, communication, instruction mix, and memory traffic, and supports roofline characterization across heterogeneous platforms and sensor configurations.