DMPCN: Dynamic Modulated Predictive Coding Network with Hybrid Feedback Representations

TL;DR

DMPCN tackles the limitations of traditional predictive coding networks by integrating local and global recurrent updates with input-dependent dynamic modulation and a tailored predictive consistency loss. The four-part loss, L_total = L_hybrid + μ L_SCT + L_SP + γ L_recon, guides precise prediction error minimization and fosters spatial and reconstruction coherence. Experiments on CIFAR-10, CIFAR-100, MNIST, and FashionMNIST show faster convergence and higher accuracy than BP and conventional PCN, along with improved calibration under noise and distribution shifts. This work advances context-aware predictive coding for vision, offering a practical framework for more efficient and reliable neural architectures.

Abstract

Traditional predictive coding networks, inspired by theories of brain function, consistently achieve promising results across various domains, extending their influence into the field of computer vision. However, the performance of the predictive coding networks is limited by their error feedback mechanism, which traditionally employs either local or global recurrent updates, leading to suboptimal performance in processing both local and broader details simultaneously. In addition, traditional predictive coding networks face difficulties in dynamically adjusting to the complexity and context of varying input data, which is crucial for achieving high levels of performance in diverse scenarios. Furthermore, there is a gap in the development and application of specific loss functions that could more effectively guide the model towards optimal performance. To deal with these issues, this paper introduces a hybrid prediction error feedback mechanism with dynamic modulation for deep predictive coding networks by effectively combining global contexts and local details while adjusting feedback based on input complexity. Additionally, we present a loss function tailored to this framework to improve accuracy by focusing on precise prediction error minimization. Experimental results demonstrate the superiority of our model over other approaches, showcasing faster convergence and higher predictive accuracy in CIFAR-10, CIFAR-100, MNIST, and FashionMNIST datasets.

Figures (3)

• Figure 1: Dynamic Modulated Predictive Coding Network (DMPCN). (a) the overall process of our proposed DMPCN, and (b) the computation process of individual blocks.
• Figure 2: Comparison of PCN and DMPCN overall prediction error acquired by using Eq. \ref{['Eq:loss_spatial']}. (a) prediction error convergence comparison on MNIST and FashionMNIST datasets with PCN and DMPCN, respectively; (b)prediction error convergence comparison on CIFAR-10 and CIFAR-100 datasets with PCN and DMPCN, respectively.
• Figure 3: Calibration error analysis of BP, PCN and DMPCN on CIFAR10 dataset using VGG-9 across varying levels of data corruption, including rotation, Gaussian blur, Gaussian noise, hue, brightness, and contrast. The DMPCN model shows superior calibration compared to BP and PCN, particularly under different degrees of data shift, demonstrating its robustness in maintaining accurate model calibration.