<ul><li>Existing works on Binary Neural Network (BNN) have focused on model weights and activations, neglecting considerations on input raw data.</li><li>This article introduces the Generic Learned Thermometer (GLT) encoding technique to enhance input data representation for BNN by learning non-linear quantization thresholds.</li><li>GLT involves multiple data binarizations to replace conventional Analog to Digital Conversion (ADC) using natural binary coding, improving model efficiency.</li><li>A compact network topology with lightweight grouped convolutions trained through block pruning and Knowledge Distillation is proposed, resulting in small, fully-binarized models suitable for efficient inference tasks.</li></ul>

End-to-end fully-binarized network design: from Generic Learned Thermometer to Block Pruning

Discover more