Post by Mastercard AI Garage

We’re excited to announce that a research paper from Mastercard AI Garage titled ‘Learning to Doubt: Forgetting-Aware Learning for Neural Networks’ has been accepted at KDD 2026! 🎉 Think about how a young person learns to be careful, not everywhere, but precisely where they've stumbled before. That earned self-awareness is exactly what our proposed regularizer teaches a neural network to do. We track moments when a model gets a prediction right one step, then wrong the next defined as forgetting events. FAL uses this instability to lower confidence on uncertain examples while leaving reliably learned ones untouched. We show both empirically and theoretically that this aligns the model's confidence with the underlying event it was built to predict. FAL is a lightweight, architecture-agnostic regularizer, under 1% training overhead, works on top of any existing objective. Any real-world model where confidence drives downstream decisions such as in fraud detection, clinical triage, credit risk, becomes more trustworthy when it genuinely knows what it doesn't know. Congratulations to the authors: Akshita Sawhney, Awanish Kumar, Soumyadeep Ghosh, Ph.D., Rahul Gupta. #MastercardAIGarage #KDD2026 #MachineLearning #AI #ModelCalibration #TrustworthyAI