The Freezes dialog is used to set the node freeze levels and the way they are decremented while the network is learning.
After the initial number of training cycles are run the freeze levels are applied in the reverse order of the node importance. All nodes are considered, not just the inputs. The least important node has the lowest freeze level. The most important node has the highest freeze levels. Any nodes with their freeze level above zero is excluded from the learning process and their activation does not change. Frozen input nodes have their activation set to the median value of the input range. The freeze levels are reduced as learning progresses so that the nodes are progressively added to the learning process. When learning has finished the input importance and sensitivity measures are more accurate.
Check to remove the freezes that are still set from previous runs.