Deep Recurrence for Dynamical Segmentation Models
Summary AI-generated
- TL;DR
- By adding a biological-style feedback loop to a standard image segmentation model, researchers made it much more robust to noise and highly data-efficient.
- Problem
- Most artificial vision systems process images in a single forward pass, unlike human vision which uses feedback loops to iteratively refine what we see. This feedforward limitation makes standard models struggle with noisy data and require large amounts of labeled training data to learn effectively.
- Method
- The researchers designed a feedback mechanism inspired by predictive coding, creating a loop that feeds the model's output back into its input to refine its predictions over time. They integrated this loop into a standard U-Net architecture, using two biological concepts—softmax projection and exponential decay—to keep the feedback loop stable.
- Results
- In controlled tests on a synthetic image segmentation task, the feedback model significantly outperformed traditional feedforward models in noisy conditions. Furthermore, it demonstrated extreme data efficiency, achieving above-random performance with only two training examples, whereas the feedforward model required at least four.
- Takeaways
- Introducing recurrent feedback loops allows neural networks to iteratively correct their own errors. This approach dramatically reduces the amount of training data needed and improves performance in noisy environments.
- For industry
- For businesses, this research introduces a way to build computer vision systems that work reliably even when images are blurry or low-quality. Because the model requires significantly less labeled data to train, it can drastically lower the time and cost associated with preparing datasets for new applications.
- Why it matters
- This work bridges the gap between biological and artificial vision, paving the way for more adaptive AI systems. Its high data efficiency and noise robustness have direct applications in fields like medical imaging and autonomous driving, where high-quality labeled data is scarce and reliability under challenging conditions is critical.
Abstract
While biological vision systems rely heavily on feedback connections to iteratively refine perception, most artificial neural networks remain purely feedforward, processing input in a single static pass. In this work, we propose a predictive coding inspired feedback mechanism that introduces a recurrent loop from output to input, allowing the model to refine its internal state over time. We implement this mechanism within a standard U-Net architecture and introduce two biologically motivated operations, softmax projection and exponential decay, to ensure stability of the feedback loop. Through controlled experiments on a synthetic segmentation task, we show that the feedback model significantly outperforms its feedforward counterpart in noisy conditions and generalizes more effectively with limited supervision. Notably, feedback achieves above random performance with just two training examples, while the feedforward model requires at least four. Our findings demonstrate that feedback enhances robustness and data efficiency, and offer a path toward more adaptive and biologically inspired neural architectures. Code is available at: github.com/DCalhas/feedback_segmentation.