convolution - Convolutional Neural Networks, Matrix of Convolutional (Kernel) -

good afternoon! in first stage on input of convolutional neural network (input layer) recieve source image (hence image of handwritten english letter). first of using nxn window goes left right scanning image , multiplication on kernel (convolutional matrix) build feature maps? written exact values kernel should have (in other words on kernel values should multiply data retrieved n*n window ). suitable multiply data on convolutional kernel intended edge detection? there numerous convolutional kernels (emboss, gaussian filter, edge detection, angle detection, etc.)? written exact kernel need multiply data detecting hand written symbols.

sample of edge detection 3*3 kernel

convolutional operation multiplication on kernel

in addition, if size of entire image 30*30, possible use window of 5*5 building feature maps? enough sufficient reaching optimal precision of letter detection?

on exact kernel best multiply area of entire image maximum precision of letter recognition? or values in kernel equaled 0? ask, formula or rule applied detect overall needed amount of built feature maps? or if task in letter recognition of english language, in each stage of feature maps building process there must exact 25 feature maps? thank reply!

in cnn, convolutional kernel shared weight matrix, , learned in similar way other weights. initialized in same way, small random values, , weight deltas propagation summed across features receive output (i.e. "pixels" in output of convolutional layer)

a typical random kernel perform little edge detector.

after training, first cnn layer can displayed , have learned kernels can interpreted if familiar image processing

there nice animated view of kernel features being learned here: http://cs.nyu.edu/~yann/research/sparse/

in short answer this: there no need correct kernels use. instead cnn library set params such number of convolutional layers, , research way view kernels learn - cnn libraries have documented way visualise them.