Parameter Sharing

Parameter sharing means the same kernel weights are used at many positions.

shared kernel3 weights reusedfully connectedseparate weights per linksharing reduces parameters and encodes the same detector across positions
Parameter sharing lets the same detector be used at many positions.

This reduces parameter count and encodes a useful assumption: the same local pattern may matter wherever it appears.

Compare a 1D input of length 100:

fully connected detector: 100 separate weights
convolution kernel size 5: 5 shared weights

The convolutional detector is much smaller. It can scan the input using the same pattern everywhere.

This is not always the right assumption. But for grids and sequences with local structure, it is often powerful.

DL-C14-T07-001Exercise: Parameter count comparison

A fully connected detector uses 20 weights. A convolution kernel uses 5 shared weights. How many fewer weights does the kernel use?

Compute it first, then check your number.

DL-C14-T07-002Exercise: Sharing meaning

Enter 1 if parameter sharing reuses the same weights across positions, or 2 if it creates new weights at every position.

Compute it first, then check your number.