|
|
|
     |
|
|
|
|
|
|
|
| |||||||||
|
|
Heuristic Network Initialisation The initialisation of the weights is a crucial process in designing neural
networks, if the weights are too large with respect to the incoming data, the
values of the local receptive field jeopardizes the activation function to work
in the saturated zone. If they are too small, the activation function
starts operating hard and fast in its linear area and
the final results likely trap into a local error minima with results close to multi linear regression.The heuristic network initialisation uses some statistical properties of all the data dedicated to training for finding an optimal range for each of the weights. Propagating the statistical distributions throughout the network prior to training generates for each weight an optimal range. It ensures therefore a much faster anticipation of the input data at the training process, the networks encompass from their initial state a perfect shape. The capability to learn a particular task is extremely large, the training is much faster and more accurate. Heuristic network initialisation is the default in cVision. |
|