Learning discriminative features via weights-biased softmax loss

Highlights：

• We determine the minimum number of units in FC layer by rigorous theoretical analysis and extensive experiments for various classes tasks, which reduces CNNs’ parameter memory and training time.

• We present a new W-Softmax loss to make CNNs learn more discriminative features, and it can effectively improve the classification performance by avoiding premature convergence.

• The size of decision margins can be optionally adjusted by a positive real-value paremeter α. By increasing the value of α, CNNs can maximize inter-class variance and minimize intra-class variance. Extensive experiments on benchmark datasets show the effectiveness of W-Softmax loss.