MicroNet Challenge is a benchmark competition that challenges researchers to design the most efficient neural networks for specific tasks under extreme parameter and computation budgets — pushing the limits of model compression, efficient architecture design, and neural network efficiency.

Challenge Constraints

• Parameter Budget: Strict maximum number of parameters (e.g., <1M parameters for CIFAR-100).
• FLOP Budget: Strict maximum computation (e.g., <12M multiply-adds for CIFAR-100).
• Scoring: Models are scored on accuracy relative to a baseline at the given budget — higher is better.
• Tasks: Typically image classification benchmarks (CIFAR-10, CIFAR-100, ImageNet).

Why It Matters

• Efficiency Research: Drives innovation in model efficiency — pruning, quantization, efficient architectures.
• Real-World: Extremely small models are needed for MCU-class edge devices (kilobyte-scale memory).
• Benchmarking: Provides a standardized comparison framework for model efficiency techniques.

MicroNet Challenge is the efficiency Olympics for neural networks — competing to build the most accurate models under extreme size and computation constraints.