Emerging as a promising contender in this quest is the forward-forward network (FFN) framework, which diverges radically from the back-propagation approach. Rather than performing end-to-end training that propagates errors backward through the network’s layers, FFNs train each layer independently by optimizing a local goodness function. This paradigm promises faster convergence and more transparent training dynamics, potentially mimicking biological learning mechanisms more closely than conventional algorithms. However, the straightforward application of FFN to convolutional neural networks (CNNs)—the backbone of image-related AI tasks—has been fraught with challenges. CNNs rely on intricate spatial feature extraction, and splitting their training into independent layers risks losing critical information embedded within the input images, adversely impacting accuracy.

Addressing this formidable challenge, a team of researchers from Seoul National University of Science and Technology, led by Ph.D. candidate Gilha Lee and Associate Professor Hyun Kim from the Department of Electrical and Information Engineering, have crafted an innovative solution dubbed the Visual Forward-Forward Network (VFF-Net). Their pioneering study, published in the esteemed journal Neural Networks in October 2025, delineates how VFF-Net transcends the inherent limitations of FFNs when applied to convolutional architectures. By preserving spatial features and optimizing efficiency, their approach represents a significant leap towards more brain-like, resource-efficient AI training methods.

Central to the VFF-Net’s success is the introduction of three novel methodologies: Label-Wise Noise Labeling (LWNL), Cosine Similarity-based Contrastive Loss (CSCL), and Layer Grouping (LG). LWNL innovatively expands the training process by incorporating three distinct types of data: clean original images, positively labeled images correctly corresponding to their targets, and negatively labeled images deliberately assigned incorrect labels. This strategy counters the tendency of traditional FFNs to lose pixel-level information by reinforcing the network’s capacity to differentiate between genuine and corrupted inputs. Essentially, LWNL acts as a sophisticated noise-augmented training mechanism designed to safeguard the integrity of spatial details critical for accurate classification.

Complementing LWNL, the Cosine Similarity-based Contrastive Loss (CSCL) function refines the goodness criterion by leveraging the directional similarity of feature maps rather than mere magnitude comparisons. Unlike conventional approaches that often utilize Euclidean distances or simplistic statistical measures, CSCL evaluates how closely feature representations align in their vector orientations within high-dimensional space. This subtle yet powerful adjustment enables VFF-Net to retain intricate spatial relationships and nuanced patterns indispensable for discerning subtle differences across image categories. Consequently, CSCL imparts an enhanced ability to capture the essence of visual features, elevating classification performance.

To address the issue of training convolutional layers individually—a process that might otherwise degrade model performance due to inconsistent optimization—VFF-Net introduces Layer Grouping (LG). This methodology clusters layers exhibiting similar output characteristics into cohesive groups, training them collectively rather than in isolation. To further bolster this collective learning, auxiliary layers are incorporated, allowing the network to propagate meaningful gradients within grouped units. By doing so, LG mitigates the pitfalls of layer-wise greediness inherent in FFNs and nurtures a more holistic convergence resembling end-to-end training. The strategic grouping procedure significantly enhances stability and responsiveness, ensuring that the network’s layered structure contributes constructively to overall task performance.

The empirical results validating VFF-Net’s innovations are compelling. When applied to a CNN architecture with four convolutional layers, VFF-Net achieved notable reductions in test error rates on benchmark image datasets. Specifically, it decreased errors by 8.31% on CIFAR-10, a widely utilized dataset comprising 10 classes of natural images, and by 3.80% on the more challenging CIFAR-100 dataset containing 100 distinct categories. Additionally, when applied to a fully connected architecture, VFF-Net attained an impressively low test error of 1.70% on the MNIST dataset, a foundational corpus for handwritten digit recognition. These quantitative gains underscore VFF-Net’s capability to bridge the performance gap between FFNs and conventional back-propagation-trained CNNs while harnessing the benefits of layer-wise training.

Beyond the quantitative metrics, VFF-Net’s broader implications are profound. By circumventing the expensive and computationally intensive back-propagation paradigm, VFF-Net charts a path toward AI models that are inherently lighter and more biologically plausible. Such models have the transformative potential to decentralize AI processing, enabling the deployment of powerful neural architectures directly on edge devices like smartphones, medical instruments, and household electronics. This local computation paradigm reduces reliance on massive cloud-based data centers, helping curb the growing energy consumption footprint of AI and promoting sustainability in technological development.

Moreover, VFF-Net’s design echoes cognitive principles observed in neurological learning processes. By emphasizing local learning rules, noise handling, similarity-based contrastive criteria, and modular training units, it recaps aspects of synaptic plasticity and cortical area specialization found in biological brains. This convergence of neuroscience insights and computational innovation paves the way for more naturalistic, interpretable, and trustworthy AI systems. With interpretability being a critical concern in deploying AI in domains like healthcare and autonomous systems, VFF-Net’s brain-inspired paradigm could foster greater adoption and regulatory acceptance.

The conceptual significance of forward-forward algorithms extends well beyond the immediate context of CNN training. VFF-Net exemplifies how reimagining fundamental learning mechanisms can unlock new avenues for efficiency and capability. While back-propagation has demonstrated remarkable success over the past decades, its scaling limitations and biological implausibility leave room for alternative frameworks. Forward-forward networks and their derivatives such as VFF-Net could serve as pivotal milestones in the search for energy-efficient, interpretable, and robust AI.

As research progresses, the integration of VFF-Net with emerging hardware accelerators and neuromorphic chips could catalyze further acceleration and optimization. Its compatibility with computational simulations and modeling environments also invites broad experimental validation and extension across diverse modalities like speech recognition, natural language processing, and robotics. Thus, VFF-Net does not merely refine convolutional training—it ignites a paradigm shift in how neural computation can be conceptualized, engineered, and applied.

In summary, the Visual Forward-Forward Network offered by the Seoul National University of Science and Technology team marks a decisive evolution in AI training methodologies. By harmonizing label-wise noise interventions, cosine similarity-guided contrastive objectives, and strategic grouping of layers, VFF-Net transcends previous limitations of the forward-forward framework. This advancement promises faster, cheaper, and more biologically inspired AI, empowering powerful neural networks on devices far beyond traditional computing centers. The breakthrough holds promising ramifications for sustainability, accessibility, and trustworthiness in the future of artificial intelligence.

Subject of Research: Not applicable

Article Title: VFF-Net: Evolving forward–forward algorithms into convolutional neural networks for enhanced computational insights

News Publication Date: October 1, 2025

Web References: https://doi.org/10.1016/j.neunet.2025.107697

References: DOI: 10.1016/j.neunet.2025.107697

Image Credits: Hyung Kim from Seoul National University and Technology

Keywords: Artificial intelligence, Machine learning, Deep learning, Neural networks, Computer science, Algorithms, Computer architecture, Pattern recognition

The challenge of selecting the optimal neural network architecture for ophthalmic ultrasound imaging is profound. Traditionally, the lack of systematic evaluation methods has impeded the development of specialized detection models that cater to the unique complexities of ocular structures. The research team tackled this by proposing a modular ablation analysis framework based on orthogonal experimental design, a statistical technique that allows comprehensive evaluation of interactions between modular components within multi-version YOLO architectures. This methodical approach enables systematic dissection of network elements, offering unprecedented insights into their individual and combined impacts on performance.

To ground their analysis in clinical reality, the researchers curated an extensive dataset comprising 1,121 ocular ultrasound images. These images provided a diverse range of anatomical presentations, capturing the intricate details necessary for robust model training and evaluation. By decoupling YOLO versions 10 through 12 into three fundamental modules—backbone, neck, and head—they established a flexible experimental structure. The backbone module facilitates feature extraction, the neck module functions as a feature aggregator and enhancer, and the head module is responsible for prediction and localization. This modularization permitted precise isolation and manipulation of architectural variables to refine detection efficiency.

The investigative process unfolded across three key experimental stages. Initially, single-module benchmarking through controlled variable experiments allowed the researchers to assess the base impact of each module in isolation. This foundational step revealed nuanced performance dynamics, highlighting how each architectural component contributes uniquely to detection accuracy and computational speed. Following this, orthogonal combination experiments—implemented using an L9(3^4) array design—enabled the team to systematically explore inter-module interactions. These experiments were augmented by range analysis and interaction heatmap visualizations, tools that elucidate the intricate dependencies and synergies between modules.

Such rigorous experimentation culminated in the final phase: optimal architecture selection. Employing Pareto front analysis, a multi-objective optimization technique, the researchers identified network combinations that offered the best trade-offs between accuracy and speed. This approach embraces the practical constraints of real-world deployment, where computational resources and latency are just as critical as detection precision. Among the configurations tested, a hybrid model combining YOLOv11’s backbone and neck with YOLOv10’s head (Bv11–Nv11–Hv10) emerged as the top performer, achieving an impressive mean average precision (mAP) of 64.0% at 26 frames per second (FPS).

Notably, the investigation also prioritized mobile optimization, recognizing the growing need for portable diagnostic tools in diverse clinical settings. The variant tailored for mobile implementation (Bv10–Nv10–Hv11) balanced compactness and accuracy, maintaining a competitive mAP of 63.5% while drastically reducing parameter count to just 8.6 MB. This underscores the study’s potential to facilitate deployment on resource-constrained devices without sacrificing diagnostic quality, a crucial advancement for point-of-care ophthalmic assessments in underserved regions.

Beyond detection, the research integrated an automated biometric analysis pipeline by applying a segmented sound velocity matching algorithm. This innovation allowed precise measurement of critical ocular biometric parameters, including anterior chamber depth, lens thickness, and axial length, directly from the ultrasound images. These parameters are vital inputs for diagnosis, surgical planning, and monitoring of ocular diseases like glaucoma and cataracts. By automating these measurements, the framework promises to significantly enhance workflow efficiency while reducing operator-dependent variability inherent in manual assessment.

Empirical validation of the automated measurements revealed strong concordance with manually obtained references. The mean absolute error across assessed parameters remained impressively low, at or below 0.133 millimeters, while the intraclass correlation coefficient (ICC) values exceeded 0.839, indicating high reliability and consistency. This level of agreement establishes confidence that the optimized YOLO architectures can serve as dependable tools in clinical practice, ensuring precision without compromising throughput or introducing bias.

From a technical standpoint, the modular ablation framework validated the feasibility of cross-version module combinations within the YOLO family. This innovative strategy breaks away from monolithic network designs, showcasing how modular engineering can capitalize on the strengths of different algorithm versions while mitigating their individual weaknesses. The backbone modules were found to bolster both accuracy and computational efficiency, whereas the neck and head modules presented a balance between speed and precision that varied depending on their configuration. The neck showed the greatest influence on detection accuracy, while the head exerted dominant control over computational load.

The implications of this research extend far beyond ophthalmic imaging. It provides a robust, quantitative foundation for network architecture design applicable to other medical imaging domains where similar trade-offs exist. The modular ablation and orthogonal design methodology represents a scalable framework to accelerate the iterative improvement of detection models, expediting the pathway from algorithmic innovation to bedside deployment. Such systematic approaches are essential as deep learning models become increasingly integral to diagnostic processes.

Clinicians and engineers alike are poised to benefit from this work. For ophthalmologists, the enhanced performance and efficiency in ocular ultrasound image analysis translate to more timely and accurate diagnoses, potentially improving patient outcomes through early detection and intervention. For medical device developers, the demonstrated adaptability and lightweight models open avenues for integrating advanced AI algorithms into handheld and portable ultrasound devices, democratizing access to high-quality ophthalmic imaging.

As the medical community continues to integrate artificial intelligence into routine practice, studies like this underscore the importance of methodological rigor and practical relevance in developing AI tools. The balance struck in this research among accuracy, speed, and deployability exemplifies a thoughtful approach to model optimization, ensuring that technological advancements translate into tangible clinical benefits. The study’s findings herald a new era of AI-assisted ocular biometry, characterized by precision, reproducibility, and accessibility across diverse healthcare environments.

Future directions inspired by this work may include expanding the dataset to incorporate pathological variations, facilitating the development of detection models sensitive to a wider array of ophthalmic conditions. Moreover, real-time integration with clinical workflows and validation within multi-center trials could pave the way for regulatory approval and widespread clinical adoption. The synergy of modular architecture design and orthogonal experimental methodologies is poised to drive continual improvements across medical imaging AI applications, with ophthalmology serving as a pioneer field.

In conclusion, the network architecture optimization for ophthalmic ultrasound image detection presented in this study represents a significant leap forward in medical imaging AI. By harnessing modular ablation, orthogonal design, and comprehensive multi-version YOLO evaluations, the research delivers a nuanced, data-driven strategy for advancing automated ocular diagnostics. Its potential to enhance both clinical accuracy and operational efficiency while accommodating device constraints marks a transformative milestone in the journey toward AI-powered precision medicine in ophthalmology.

Subject of Research: Network architecture optimization for ophthalmic ultrasound image detection using modular ablation of multi-version YOLO.

Article Title: Network architecture optimization for ophthalmic ultrasound image detection based on modular ablation of multi-version YOLO.

Article References:
Li, Z., Wang, X., Yu, X. et al. Network architecture optimization for ophthalmic ultrasound image detection based on modular ablation of multi-version YOLO. BioMed Eng OnLine 24, 121 (2025). https://doi.org/10.1186/s12938-025-01459-5

Image Credits: AI Generated

DOI: https://doi.org/10.1186/s12938-025-01459-5