Gradiend-based learning Applied to Document Reconition點擊打開鏈接
http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf
輸入層 32X32
C1
6@28X28 Convolutions 核大小 6@5x5 ,pad=0,slide=1; 156個參數,122304(28x28*6(5x5+1)個連接。
S2
6@14X14 Subsampling 6@2*2,slide=2; 5580連接,14x14x6x(4+1)個鏈接
C3
16@10X10 Convolutions 16@5*5,slide=1; 參數個數:6*10*25+16,連接個數
6*10*(25+1)此層的核連接不是全連接,是稀疏連接,打破對稱性,減少參數個數。
S4
16@5X5 Subsampling 16@2*2,slide=2; 5x5x16x(4+1)
C5 120 Full conection 1*1*120*16(5*5)+120
F6 84 Full connection 120*84+84
OUTPUT 10 Gaussian connections