ENCORE Compression: Exploiting Narrow-width Values for Quantized Deep Neural Networks