Adaptive methods for accelerating convolution operators in different numeric precision settings and sparsity conditions
AMACA studies supervised learning techniques to accelerate the convolutional layers of CNNs. AMACA allows for multi-objective optimization, leveraging algorithmic setting to balance accuracy, performance and power consumption.
At the moment, the project focuses on the ARM Compute Library (ARMCL) and its four convolution implementations.
ROADMAP
- Part 1: Design and implementation
- Part 2: Dataset Generation
- Part 3: Dataset Analysis
- Part 4: Single-objective Classification
- We are here!
- Part 5: Multiple-objective Classification
- Part 6: Finalization