🤖 machine learning

Forward-Only Convolutional Neural Networks with Learnable Channel-Class Assignment

This paper introduces a novel Forward-Only Convolutional Neural Network framework that employs learnable channel-class assignments and loss-aware layer weighting to achieve state-of-the-art performance on benchmark datasets, significantly narrowing the gap with backpropagation-based methods.

Original authors: Mohammadnavid Ghader, Saeed Reza Kheradpisheh, Bahar Farahani, Mahmood Fazlali

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammadnavid Ghader, Saeed Reza Kheradpisheh, Bahar Farahani, Mahmood Fazlali

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a team of 100 specialized workers (the "channels" in a computer's brain) to sort a massive pile of mixed-up mail into different bins (the "classes," like "Cat," "Dog," or "Car").

For a long time, the standard way to teach these workers was Backpropagation. Think of this like a strict manager who watches the final result, realizes a mistake was made, and then walks all the way back through the entire factory line, telling every single worker exactly how they messed up. While effective, this process is slow, requires a lot of memory, and doesn't really match how biological brains learn.

A newer method called Forward-Forward (FF) tries to fix this. Instead of a manager walking backward, every worker just looks at the mail in front of them, makes a guess, and checks if they feel "good" about it. If the mail looks right for their bin, they feel good; if it looks wrong, they feel bad. They learn just by looking forward.

However, the original Forward-Forward method had a big flaw: Rigid Assignments.
Imagine the manager told the workers, "You, Worker #1 through #10, are only allowed to look at 'Cat' mail. You, Worker #11 through #20, are only for 'Dog' mail." This is called Static Channel Assignment.

The Problem: What if a picture of a cat has a lot of "dog-like" fur texture? Worker #1 (the cat specialist) might miss it because they are strictly forbidden from looking at those features. The workers are stuck in their boxes, unable to adapt.

The Paper's Solution: A Flexible Team

This paper introduces a new way to organize these workers called Learnable Channel–Class Assignment. Here is how it works, using simple analogies:

1. The "Flexible Badge" System (Learnable Assignment)

Instead of giving workers permanent, rigid badges that say "Cat Only," the team leader gives them smart, adjustable badges.

At the start, Worker #1 might think, "I'm 80% sure I'm a Cat specialist, but maybe 20% Dog."
As they process more mail, they can learn and adjust these percentages. If they realize they are actually really good at spotting "fluffy textures" which appear in both cats and dogs, their badge updates to say, "I'm 50% Cat, 50% Dog."
The Result: The workers aren't stuck in rigid boxes. They can dynamically specialize based on what the data actually looks like, making the whole team much more efficient.

2. The "Team Harmony" Rules (Regularization)

To make sure the workers don't all just decide to be "Cat Specialists" and ignore the rest, the paper adds two rules (mathematical nudges):

The "Sparsity" Rule (Entropy): This encourages workers to focus. It says, "Don't try to be good at everything. Pick a few specific things to be really good at." This prevents confusion.
The "Diversity" Rule (Orthogonality): This ensures the team doesn't all do the same job. It says, "If Worker #1 is great at spotting 'ears,' Worker #2 should be great at spotting 'tails,' not ears." This forces the team to cover different angles of the problem.

3. The "Smart Voting" System (Layer Contribution)

In a deep neural network, there are many layers of workers (like a factory with many assembly lines). Usually, only the final line's decision counts.

The Old Way: Ignore the early workers; only listen to the last guy.
The New Way: The paper proposes a Loss-Aware Layer Contribution Strategy. Imagine a voting system where every layer gets to cast a vote for the final answer.
How it works: The system checks the "score" (loss) of each layer. If Layer 3 is doing a great job at identifying the object, its vote counts for a lot. If Layer 5 is confused and making mistakes, its vote counts for very little.
The Result: The final decision is a weighted average of all the layers, giving more weight to the ones that are actually performing well.

What Did They Find?

The authors tested this new "Flexible Team" approach on standard image datasets (like recognizing cats, dogs, cars, and tiny images of objects).

Beating the Competition: Their method consistently outperformed previous Forward-Forward methods. It was better at sorting the mail than the old "rigid badge" systems.
Doing More with Less: Because the workers were so flexible, the team could be smaller. They achieved better results using 50% fewer workers (channels) than the previous best methods.
Closing the Gap: While the old "Backward Manager" (Backpropagation) is still the absolute best, this new "Forward-Only" method got much closer to that top performance than ever before, proving that you don't need a backward manager to get great results.

In Summary

This paper teaches computers to learn by looking forward only, but with a twist: instead of forcing parts of the brain to be rigid specialists, it lets them adaptively specialize based on the data. It also uses a smart voting system to decide which parts of the brain are doing the best job at any given moment. The result is a faster, more efficient, and highly accurate way for computers to learn from images without needing the complex "backward" training process.

Technical Summary: Forward-Only Convolutional Neural Networks with Learnable Channel–Class Assignment

Problem Statement
The Forward-Forward (FF) algorithm offers a biologically plausible alternative to backpropagation by replacing global gradient-based credit assignment with local, forward-only objectives. While recent adaptations have extended FF to Convolutional Neural Networks (CNNs), existing formulations (e.g., CwComp, DeeperForward) rely on static channel–class partitions. In these methods, convolutional channels are permanently assigned to specific classes via fixed grouping mechanisms. The authors argue that this rigidity limits the network's ability to dynamically adapt channel specialization to evolving feature distributions during training, leading to inefficient feature utilization, reduced representational flexibility, and an inability to fully leverage deeper architectures. Furthermore, standard FF inference often relies solely on the final layer or simple aggregation, potentially ignoring valuable intermediate representations.

Methodology
The proposed method introduces a Learnable Channel–Class Assignment mechanism and a Loss-Aware Layer Contribution Strategy integrated into residual CNNs.

Learnable Channel–Class Assignment (CAW-Conv):
- Instead of statically assigning channels to classes, the authors introduce a learnable parameter matrix $U \in \mathbb{R}^{C_{out} \times J}$ (where $C_{out}$ is the number of output channels and $J$ is the number of classes).
- A softmax function is applied to the rows of $U$ to generate a weight matrix $M$ , representing the contribution rate of each channel to each class. This allows channels to dynamically specialize based on data structure.
- The "goodness" score for each class is computed as a pooled representation of the feature maps ( $A_{pooled}$ ) multiplied by the assignment matrix $M$ .
- Regularization: To prevent degenerate solutions and promote effective learning, two regularization terms are added to the loss function:
  - Entropy Regularization: Minimizes the entropy of the class-channel weights to encourage sparsity, ensuring channels contribute significantly to only a few classes.
  - Orthogonality Regularization: Enforces orthogonality among the normalized class-channel weight vectors to promote diversity in feature extraction across different classes.
Loss-Aware Layer Contribution Strategy:
- Unlike traditional networks that make decisions only at the final layer, the proposed framework aggregates predictions from all layers.
- The contribution of each layer to the final decision is weighted adaptively based on its validation loss. Layers with lower validation losses (indicating better predictive reliability) are assigned higher weights via a softmax transformation.
- The final prediction is a weighted sum of the goodness scores from all layers.
Training and Inference:
- The network is trained using a forward-only, layer-wise contrastive process. Each layer is optimized independently using a local cross-entropy loss combined with the regularization terms.
- Instance normalization is applied between layers to decouple feature scale from representational direction.
- During inference, the weighted aggregation of layer outputs is used to determine the final class prediction.

Key Contributions

Dynamic Specialization: The paper proposes replacing static channel grouping with a learnable assignment mechanism, enabling adaptive, data-driven specialization of convolutional channels.
Enhanced Representational Capacity: By utilizing entropy and orthogonality regularization, the method promotes sparse and diverse channel utilization, improving feature discriminability.
Adaptive Inference: The introduction of a loss-aware layer contribution strategy allows the model to leverage intermediate representations more effectively than methods relying solely on final-layer outputs.
Efficiency: The proposed method achieves superior performance with significantly fewer channels compared to static grouping approaches.

Experimental Results
The method was evaluated on MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet using three architectures: TinyCNN-4, VGG-14, and ResNet-17.

Performance on CIFAR-10: The proposed method achieved 89.37% accuracy with a 17-layer ResNet, outperforming the previous state-of-the-art FF-based method (DeeperForward) by over 3 percentage points. It also surpassed DeeperForward by more than 5% with a shallow 4-layer architecture.
Performance on Challenging Datasets: On CIFAR-100, the proposed method reached 63.52% (vs. 53.09% for DeeperForward). On Tiny-ImageNet, it achieved 49.87% (vs. 41.36% for DeeperForward).
Channel Efficiency: The proposed method consistently outperformed DeeperForward even when the channel size was reduced by 30–50% (e.g., achieving 86.52% on CIFAR-10 with ResNet-17 at 0.6x channel scale, compared to 86.22% for DeeperForward at full scale).
Comparison to Backpropagation: While still trailing behind full backpropagation (BP) in absolute accuracy (e.g., 89.37% vs. 95.52% for a 152-layer BP model on CIFAR-10), the proposed FF-based method substantially narrows the performance gap compared to previous FF attempts, particularly on MNIST and Fashion-MNIST where it surpassed some BP-based baselines.

Significance and Claims
The paper claims that introducing learnable channel specialization and layer contribution weighting significantly enhances the representational capacity of forward-only learning in deep CNNs. The authors assert that their approach establishes new state-of-the-art performance among FF-based models, demonstrating that forward-only learning can be scaled effectively to deeper architectures and complex tasks without the constraints of static channel assignments. The results suggest that the proposed framework is a viable, biologically inspired alternative that can compete with gradient-based methods while offering greater scalability and efficiency in channel usage. The authors modestly note that while the learnable matrix improves performance, future work could focus on auxiliary mechanisms to further guide its learning and better regulate global loss influence.