FMCE-Net++: Closed-Loop Training Framework
- The paper introduces a closed-loop framework that integrates a frozen FMCE-Net auxiliary head to predict convergence scores and supervise backbone optimization.
- It employs a Representation Auxiliary Loss modulated by a tunable Representation Abstraction Factor to align unsupervised intermediate representations with global task labels.
- Evaluated on MNIST, CIFAR-10, FashionMNIST, and CIFAR-100, FMCE-Net++ enhances accuracy without requiring architectural changes or additional data.
FMCE-Net++ is a training framework for image classification deep neural networks that converts Feature Map Convergence Evaluation (FMCE) from a post hoc interpretability mechanism into a closed-loop optimization signal. It augments a standard classifier backbone with a pretrained, frozen FMCE-Net auxiliary head that predicts Feature Map Convergence Scores (FMCS), and uses those predictions together with task labels to supervise backbone optimization through a Representation Auxiliary Loss (RAL) modulated by a tunable Representation Abstraction Factor. In the reported formulation, the method targets the longstanding mismatch between global task supervision and unsupervised intermediate representations, and is evaluated on MNIST, CIFAR-10, FashionMNIST, and CIFAR-100, where it is reported to improve accuracy without architectural modifications or additional data (Zhu et al., 8 Aug 2025).
1. Problem setting and motivation
FMCE-Net++ is motivated by a limitation of standard end-to-end deep learning: optimization is usually guided only by a global task loss, such as classification cross-entropy, while intermediate modules and feature maps are not explicitly evaluated or supervised. In this framing, the absence of module-level assessment produces both an interpretability problem and a training problem. Internal feature maps are central to representation learning, yet vanilla training does not directly impose a principled notion of whether those representations are converging in a meaningful, stable, and semantically useful manner.
The framework addresses this gap by treating module-level feature map convergence as a usable supervisory signal rather than merely an analytical artifact. Its central claim is that if the convergence state of an intermediate feature map can be quantified, that quantity can be reused during training to encourage the backbone to produce representations that resemble those from later, better-converged stages of optimization. The intended consequence is not a change in classifier architecture, but a change in what the architecture is asked to optimize.
This positioning is important for interpreting the method correctly. FMCE-Net++ is not presented as a new backbone family, a data-augmentation procedure, or an auxiliary pretraining corpus. It is presented as a training framework that attaches an auxiliary convergence-evaluation mechanism to an otherwise standard classifier and feeds that signal back into learning (Zhu et al., 8 Aug 2025).
2. Antecedent concepts: FMCE, FMCS, and FMCE-Net
FMCE-Net++ is an extension of the earlier FMCE/FMCS framework introduced in prior work cited as Zhang et al. (2024). In that earlier formulation, the goal was independent evaluation of functional modules rather than training-time intervention. The original setting was motivated by autonomous driving perception models, which were described as being composed of multiple functional modules such as feature extraction backbones, perspective transformation modules, feature fusion modules, and downstream task heads. The authors argued that end-to-end black-box optimization obscures whether a given internal module has actually matured, whether its output features are stable and informative, and whether downstream modules are compensating for weaknesses upstream (Zhang et al., 2024).
The operational definition of feature map convergence in this line of work is based on training checkpoints. A backbone is sampled at several checkpoints along its convergence trajectory, . Feature maps extracted at those checkpoints are labeled by checkpoint index , and that index acts as an FMCS class. A lightweight network, FMCE-Net, is then trained to predict from a feature map which convergence stage it came from. In this construction, FMCS is a proxy for how far along the training trajectory a feature map lies.
The earlier work therefore established a quantitative metric, the Feature