On Feature Saliency and Deep Neural Networks

Technological advances have allowed for the assimilation of a variety of data, driving a shift away from the use of simpler and constrained patterns to more complex and diverse patterns in retrieval and analysis of such data. This shift has inundated the conventional techniques and has stressed the need for intelligent mechanisms that can model the complex patterns in the data. Deep neural networks have shown some success at capturing complex patterns, including the so-called attentioned networks, have significant shortcomings in distinguishing what is important in data from what is noise. This dissertation observes that the traditional neural networks primarily rely solely on gradient-based learning to model deep features maps while ignoring the key insight in the data that can be leveraged as complementary information to help learn an accurate model. In particular, this dissertation shows that the localized multi-scale features (captured implicitly or explicitly) can be leveraged to help improve model performance as these features capture salient informative points in the data.

This dissertation focuses on “working with the data, not just on data”, i.e. leveraging feature saliency through pre-training, in-training, and post-training analysis of the data. In particular, non-neural localized multi-scale feature extraction, in images and time series, are relatively cheap to obtain and can provide a rough overview of the patterns in the data. Furthermore, localized features coupled with deep features can help learn a high performing network. A pre-training analysis of sizes, complexities, and distribution of these localized features can help intelligently allocate a user-provided kernel budget in the network as a single-shot hyper-parameter search. Additionally, these localized features can be used as a secondary input modality to the network for cross-attention. Retraining pre-trained networks can be a costly process, yet, a post-training analysis of model inferences can allow for learning the importance of individual network parameters to the model inferences thus facilitating a retraining-free network sparsification with minimal impact on the model performance. Furthermore, effective in-training analysis of the intermediate features in the network help learn the importance of individual intermediate features (neural attention) and this analysis can be achieved through simulating local-extrema detection or learning features simultaneously and understanding their co-occurrences. In summary, this dissertation argues and establishes that, if appropriately leveraged, localized features and their feature saliency can help learn high-accurate, yet cheaper networks.