batchnorm_slideaaaaaaaaaaaaaaaaaaaa.pptx

Batchnormalization

1. Batch Normalization on Images (N×C×H×W) For each channel c: Mean: μ_c = (1 / N·H·W) · Σ x[n,c,i,j] Variance: σ²_c = (1 / N·H·W) · Σ (x[n,c,i,j] - μ_c)² Normalize: x̂[n,c,i,j] = (x[n,c,i,j] - μ_c) / sqrt(σ²_c + ε) Scale & Shift: y[n,c,i,j] = γ_c · x̂[n,c,i,j] + β_c