We present a cognitively-inspired salience technique that utilizes our knowledge of human information processing mechanisms to determine the extent to which each feature is responsible for making a given decision in a given context. While measures of model-based feature importance are prevalent in ML packages (e.g., the fitted attribute feature importances and permutation importance function in the sklearn package), there are fewer off-the-shelf techniques for class-based and decision-based importance. Our measures of cognitive salience result from deriving the output of the cognitive mechanisms with respect to specific features or instances in the current situation. Applying this technique to intrusion detection datasets (e.g., UNSW-NB15, CICIDS2017) we can compute the salience of each feature for any given classification, including different feature salience for correct vs incorrect decisions. By tracing (i.e., clamping) our model’s decisions to that of a specific classification technique or individual decision maker, we are able to provide measures of model-agnostic salience. While extant feature importance techniques (e.g., selectKbest, random forest) focus on which features improve overall model performance, cognitive salience instead considers how much a given feature influences the current decision. This allows for more fine-grained analyses of model performance and has been used previously to introspect over the decisions of deep reinforcement learning algorithms and explain why RL agents failed at a StarCraft 2 task or which strategies they discovered in grid world multi-agent adversarial tasks, as well as introspecting over features in an insider attack scenario to identify how (and when) humans learn the relative value of each feature. We report a comparison between cognitive salience and extant ML-based techniques.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.