Update README.md

README.md

@@ -141,9 +141,9 @@ BlurPool2d introduces a low-pass filter before subsampling:
 
   - x → (low-pass filter) → subsample
 
-The filter is a normalized binomial kernel [1, 2, 1] ⊗ [1, 2, 1], applied depthwise: one filter per channel, no cross-channel mixing.
+The filter is a normalized fixed binomial kernel [1, 2, 1] ⊗ [1, 2, 1], applied depthwise: one filter per channel with no cross-channel mixing.
 
-This enforces approximate band-limitedness prior to resolution reduction, reducing aliasing artifacts and making feature extraction more stable under downsampling.
+This enforces approximate band-limitedness prior to resolution reduction. It reduces aliasing artifacts and makes feature extraction more stable under downsampling.
 
 ### Squeeze-and-Excitation (SE) blocks
 
@@ -153,8 +153,7 @@ SE blocks perform channel-wise reweighting.
 
   - x → s ⊙ x
 
-The bottleneck dimension is max(C // r, 1) where r = 16, keeping the recalibration 
-lightweight relative to the feature dimension.
+The bottleneck dimension is max(C // r, 1) where r = 16, keeping the recalibration lightweight relative to the feature dimension.
 
 The network learns to suppress channels that carry unstable high-frequency information 
 and amplify channels that carry structurally stable features.