Artificial intelligence (AI) models can make highly accurate predictions, but understanding why they make certain decisions remains a challenge. One approach to uncovering their reasoning is mask-based explanation methods, which work by identifying the most influential parts of the input—such as pixels in an image or time points in a time series. However, these methods are computationally expensive and heavily influenced by how they start.

To address this, we introduce StartGrad, a smarter initialization technique specifically designed for mask-based explanations. Despite its simplicity, our experiments on vision and time-series data show that StartGrad enhances the optimization process of mask-based explanation methods by helping them to reach target metrics faster and, in some cases even improve their overall performance.

StartGrad is a simple yet powerful technique that enhances AI explanations, making them both faster and more accurate. This is a crucial step toward achieving trustworthy AI, particularly in critical areas like healthcare, where model transparency and model understanding are essential for practical, real-world application.