A Survey of the Interpretability Aspect of Deep Learning Models

Abstract

Deep neural networks have attained near-human degree of quality in images, textual, audio, and video recording categorization and predictions tasks. The networks, on the other hand, are still typically thought of as black-box functional probabilistic models that transfer an input data to a trained classifier. Integrating these systems into mission-critical activities like clinical diagnosis, scheduling, and management is the next stage in this human-machine evolutionary change, and it necessitates a degree of confidence in the technology output. Statistical measures are often employed to estimate an output's volatility. The idea of trust, on the other hand, is dependent on a human's sight into a machine's inner workings. To put it another way, the neural networks must justify its outputs in a way that is intelligible to humans, leading to new insights into its internal workings. "Interpretable deep networks" is the name we give to such networks. The concept of interpretability is not one-dimensional. Indeed, the variability of an interpretation owing to varying degrees of human comprehension necessitates the existence of a plethora of characteristics that together define interpretability. Furthermore, the model's interpretations may be expressed in terms of low-level network variables or input properties. We describe several of the variables that are helpful for model interpretability in this study, as well as previous work on those dimensions. We do a gap analysis to determine what remains to be improved to increase models interpretability as step of the procedure.