drawering – our newest idea for improving the performance of neural networks in regression tasks will be presented during The 2017 International Joint Conference on Neural Networks in in Anchorage, Alaska, USA.
In our case the improvement is very clear (see the accuracy-loss chart below), but we strongly believe the method should work well in most of the regression tasks. Let us know how this idea works on your model!
Here’s the abstract:
The method presented extends a given regression neural network to make its performance improve. The modification affects the learning procedure only, hence the extension may be easily omitted during evaluation without any change in prediction. It means that the modified model may be evaluated as quickly as the original one but tends to perform better.
This improvement is possible because the modification gives better expressive power, provides better behaved gradients and works as a regularization. The knowledge gained by the temporarily extended neural network is contained in the parameters shared with the original neural network.
The only cost is an increase in learning time.
And here’s the full paper.