Kernel based Deep Learning using multi-layer kernel machines(MKMs) was proposed by Y.Cho and L.K. Saul in saul. In MKMs they used only one kernel(arc-cosine kernel) at a layer for the kernel PCA-based feature extraction. We propose to use multiple kernels in each layer by taking a convex combination of many kernels following an unsupervised learning strategy. Empirical study is conducted on mnist-back-rand, mnist-back-image and mnist-rot-back-image datasets generated by adding random noise in the image background of MNIST dataset. Experimental results indicate that using MKL in MKMs earns a better representation of the raw data and improves the classifier performance.