As deep learning moves closer to the edge, organizations face a critical trade-off: how to reduce model size and inference time without sacrificing the accuracy needed for real-world applications. Resource-constrained devices such as wearables, IoT devices, smartphones, and embedded systems have limited compute power and memory, making conventional deep learning models difficult to deploy efficiently.
This whitepaper explores how TinyML optimization and compression techniques can help overcome these constraints. Through experiments with Custom, VGG16, and MobileNet models across fashion, radiology, and dermatology datasets, the study evaluates techniques including quantization and pruning to understand their impact on accuracy, inference time, and model size. The findings demonstrate the importance of selecting and tailoring optimization strategies to achieve the right balance between model fidelity and efficiency, providing practical insights into deploying lightweight deep learning models on resource-constrained edge devices