Machine-learning models are utilised in the real world to assist radiologists in identifying potential diseases in X-rays; however, these models are intricate and their prediction process remains elusive even to their creators. To address this, researchers employ saliency methods, techniques that seek to offer insights into the model’s behaviour and elucidate its decision-making procedure.
Researchers from the Massachusetts Institute of Technology (MIT) and a multinational technology company have collaboratively developed a tool with a new method to assist users in selecting the most suitable saliency method for their specific requirements. Therefore, they introduced saliency cards, providing standardised documentation summarising how a particular process of saliency operates, including its strengths, weaknesses, and explanations to aid users in correctly interpreting the method’s outputs.
The Co-lead Author, Angie Boggust, a graduate student in electrical engineering and computer science at MIT and a member of the Visualization Group of the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL), expresses the team’s aspiration that users equipped with this knowledge will be able to consciously select a suitable saliency method based on the specific machine-learning model being employed and the task it aims to accomplish.
Boggust explains that saliency cards are purposefully crafted to provide a concise and easily understandable overview of a saliency method while highlighting the essential attributes most relevant to human users. These cards are intended to be accessible to a wide range of individuals, including machine-learning researchers and even those unfamiliar with the field and seeking guidance in selecting a saliency method for the first time.
Choosing the “wrong” saliency method can have serious consequences. For instance, one saliency method known as integrated gradients compares the importance of features in an image to a meaningless reference point. Features with the highest priority compared to this reference point are considered the most meaningful for the model’s prediction. If an unsuitable saliency method is chosen, it can lead to incorrect or misleading interpretations of the model’s behaviour and predictions. Therefore, selecting a saliency method appropriate for the specific task requirements is crucial to avoid these consequences.
Saliency cards can assist users in avoiding choosing “the wrong method” by reducing the operational details of a saliency method into ten user-centric attributes. The attributes encompass the methodology for calculating saliency, the connection between the saliency method and the model, and how users interpret the outputs generated by the method.
The saliency cards can also serve as a valuable resource for scientists by revealing areas where further research is needed. For instance, the researchers from MIT encountered a challenge in finding a saliency method that was both computationally efficient and applicable to any machine-learning model. This highlights a gap in the research space that warrants further exploration and development.
In the future, the researchers aim to delve into the less-explored attributes of saliency methods and potentially create task-specific saliency techniques. They also seek to enhance their understanding of how individuals perceive saliency method outputs, with the potential for developing improved visualisations. Furthermore, they have made their work accessible through a public repository, inviting feedback from others that will contribute to future advancements.
Boggust is optimistic, envisioning these saliency cards as dynamic documents that will evolve as new saliency methods and evaluations emerge. Ultimately, this marks just the beginning of a broader discussion regarding the attributes of saliency methods and their relevance to different tasks. Boggust believes that in the future, there will be other researchers who will further develop this discovery.