The parameter saliency approach can be naturally generalized to any architecture, not just vision. The changes that need to be made:
1. Aggregation. Currently, our code aggregates saliency of the model parameters by averaging on the conv filter level. We used that because in the literature filters have been shown to be interpretable. However, no aggregation can be used and the saliency profile can be computed on individual parameter level allowing for any architecture.
2. Loss. The loss is also not limited to classification losses, any other loss function can be used, e.g. metric learning or regression.
These should be fairly simple modifications of the code. Happy to help if needed!
The parameter saliency approach can be naturally generalized to any architecture, not just vision. The changes that need to be made: 1. Aggregation. Currently, our code aggregates saliency of the model parameters by averaging on the conv filter level. We used that because in the literature filters have been shown to be interpretable. However, no aggregation can be used and the saliency profile can be computed on individual parameter level allowing for any architecture. 2. Loss. The loss is also not limited to classification losses, any other loss function can be used, e.g. metric learning or regression.
These should be fairly simple modifications of the code. Happy to help if needed!