Memory and Computation-Efficient Kernel SVM via Binary Embedding and Ternary Model Coefficients

Kernel approximation is widely used to scale up kernel SVM training and prediction. However, the memory computation costs of models are still too large if we want deploy them on memory-limited devices such as mobile phones, smart watches IoT devices. To address this challenge, propose a novel computation-efficient model by using both binary embedding coefficients. First, an efficient way generate compact data which can preserve similarity. Second, simple but effective algorithm learn linear classification with coefficients support different types loss function regularizer. Our achieve better generalization accuracy than existing works learning since allow coefficient be -1, 0 or 1 during stage removed inference. Moreover, provide detailed analysis convergence our inference complexity model. The shows that local optimum guaranteed much lower other competing methods. experimental results five real-world datasets have demonstrated proposed method build accurate nonlinear cost less 30KB.