Normal Guidance is what Attention Needs

A new regularization method called Normal Guidance improves how attention-based models localize disease in 3D medical scans when trained only with whole-volume labels, researchers report on arXiv [1]. The technique addresses a persistent weakness in multiple instance learning (MIL) for medical imaging. In weakly supervised settings, where only a single binary label exists for an entire 3D volume, attention-based MIL models assign an importance score to each 2D slice [1]. However, recent work demonstrated that a simple baseline that focuses on the center of a scan, ignoring image content entirely, can outperform both attention-based and transformer-based MIL at slice-level classification of 3D brain scans [1]. The authors of the new paper show this same center-focused baseline also beats existing MIL methods on thoracic and abdominal CT scans [1]. Motivated by that finding, the researchers designed Normal Guidance, a regularization technique that pushes the learned attention distribution toward a bell-shaped curve [1]. The intuition draws on a broader understanding of attention as a selective process. In cognitive science, object-based attention describes how the brain enhances the sensory representation of a selected perceptual unit, improving the processing of its features [3]. By imposing a structured, Gaussian-like prior on slice-level attention scores, Normal Guidance steers the model away from scattered or degenerate focus patterns without requiring additional annotations [1]. The team evaluated the approach across three medical imaging datasets encompassing more than 4 million 2D slices [1]. With Normal Guidance applied, attention-based and transformer-based MIL methods delivered significantly better slice-level localization than the current state-of-the-art, while remaining competitive on whole-scan classification accuracy [1]. The work was posted on arXiv on 26 May 2026 and is hosted under the arXivLabs framework, which supports experimental projects developed with community collaborators who adhere to arXiv’s values of openness, community, excellence, and user data privacy [1]. The bell-shaped constraint echoes principles found in other domains where a normal distribution describes a continuous trait. For instance, executive functioning and self-regulation are understood to exist on a bell curve, with attention deficit hyperactivity disorder representing the extreme lower end of that distribution [4]. While unrelated to machine learning architecture, the parallel illustrates how a Gaussian assumption can serve as a useful prior when modeling phenomena that cluster around a central tendency. The Normal Guidance technique does not require changes to underlying network architectures, making it compatible with existing attention-based and transformer-based pipelines [1].