Local Response Normalization
Intermediate Google DeepMind, Meta+2
The problem, from first principles
Local Response Normalization (LRN) normalizes each activation using responses of neighboring channels. Useful in early CNNs for competitive inhibition. For each position, normalization is based on a local neighborhood across channels.
lrn(x,k,α,β,n)=xk+αi=max(0,c−n/2)∑min(C−1,c+n/2)xi2−βWhere the sum is over n neighboring channels (spatial location fixed).
From theory to code
Implement:
local_response_norm(x, k=2.0, alpha=1e-4, beta=0.75, n=5)
Constraints
- x shape: (N, C, H, W)
- k, alpha, beta, n are positive floats
- n is the size of neighborhood (odd).
- Return normalized x, same shape.
Keywords
deep-learningtrainingnormalization
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.