REVIEW 4 major objections 5 minor 31 references
FFGAF-SNN: The Forward-Forward Based Gradient Approximation Free Training Framework for Spiking Neural Networks
T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A gradient-free, forward-only training rule brings convolutional spiking networks to 99.58% accuracy on MNIST and 75.64% on CIFAR-10.
desk verdict Plausible FF-SNN recipe with a load-bearing gap: the channel-allocation features are unspecified, so the headline margins may not survive a clean re-run. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the black-box spiking layer: integrate-and-fire neurons are frozen and never receive weight updates, so their non-differentiable activation function never needs a gradient. Around this sits a channel-wise competitive loss, taken from prior Forward-Forward work, which partitions the channel dimension by class and measures each class's 'goodness' as the mean squared spatiotemporal activation; a locally normalized softmax-style cross-entropy loss trains each layer block independently. Two supporting components are an improved ReLU activation (a clipped, level-shifted function borrowed from ANN-SNN conversion) that reduces encoding error at the spike layer, and a regularization layer that rescales outputs by a trainable threshold to keep them inside the spiking neuron's expressible range. The allocation rule works by computing class-mean feature vectors, cosine similarities between classes, standardized similarity sums, and a proportion-based channel split with a uniformity hyperparameter fixed at 2.
What would settle it
Re-run the CIFAR-10 experiment with the channel-allocation scores recomputed from class-mean features of only the training split, produced by a randomly initialized, untrained convolutional network, and compare against both the reported 75.64% and the uniform-allocation baseline; if accuracy falls to the uniform-allocation level (about 75.28% for the comparable model), the allocation advantage came from information about the test classes rather than from the complexity measure itself.
Extended reading notes
Core claim
The central discovery is that the non-differentiability of spiking neurons does not have to be worked around at all: if the spiking layer is inserted as a frozen black box that encodes rectified, normalized activations into spike trains, a Forward-Forward training signal defined on the squared, spatiotemporal mean of per-class channel subsets can drive the trainable convolutional layers to competitive accuracy. The paper further asserts that distributing channels according to normalized inter-class similarity scores, rather than uniformly, improves accuracy under fixed parameter budgets because harder classes receive more representational resources. The method is claimed to be the first Forward-Forward training of convolutional spiking networks, and to reduce parameter counts to about 2.3% of a baseline MLP while lowering memory-access energy.
Load-bearing premise
The channel-allocation advantage rests entirely on the class-mean feature vectors in Eq. (1): the paper does not say which feature extractor produced them or whether the statistics were taken from the training split, the validation split, or the full dataset including test images, so if those vectors encode test knowledge, the reported accuracy gains over uniform allocation would not transfer to unseen data.
Editorial extensions
If this is right
- Convolutional spiking networks can be trained end-to-end with purely forward passes, with no surrogate gradient, no backward pass, and no stored intermediate activations for backpropagation.
- Assigning more channels to confusable classes improves accuracy when parameters are scarce; the gap over uniform allocation closes as the model's capacity grows, so the benefit appears exactly in the edge-deployment regime the paper targets.
- The same frozen-black-box trick applies to any non-differentiable neuron model, so extending to richer spiking dynamics does not require new training mathematics.
- At 433,600 parameters and 0.081 mJ total estimated energy per inference, the trained network is small enough for on-device learning if the Forward-Forward updates are implemented in hardware.
- The reported accuracies beat every Forward-Forward-trained spiking network in the comparison table while using fewer epochs on MNIST and Fashion-MNIST than most baselines.
Reading between the lines
- If the class-mean features used for allocation (Eq. 1) are computed from the training split only with a feature extractor that does not already encode the target classes, the method becomes a fully data-driven architecture allocation rule; the paper leaves this provenance unstated, and the claimed allocation gains should be re-measured under that constraint.
- Because allocation is computed once before training, an obvious extension the paper does not explore is to update the allocation during training as class-mean features drift, which might further raise the CIFAR-10 result where inter-class difficulty differences are small.
- The black-box spiking layer suggests a modular co-design: any neuromorphic core that maps rectified inputs to spikes could be dropped into the architecture without altering the learning rule, connecting the method to heterogeneous hardware accelerators.
- The energy comparison depends on theoretical estimates from a specific analytical model; the interesting test is whether these ratios hold on actual neuromorphic hardware, where memory-access energy and spike sparsity interact differently.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes FFGAF-SNN, a Forward-Forward-based training framework for convolutional spiking neural networks. Spiking layers are frozen and treated as non-trainable black-box encoders, while trainable convolutional blocks are optimized by a cross-entropy-style goodness loss. A class-complexity analysis of dataset features (Eqs. 1-4) is used to allocate channels per class, and an improved ReLU/quantization activation (Eq. 13) plus output regularization is introduced to reduce spiking-layer conversion error. Experiments report test accuracies of 99.58% on MNIST, 92.13% on Fashion-MNIST, and 75.64% on CIFAR-10, claiming to surpass all existing FF-based SNN approaches and several non-backpropagation baselines, with lower parameter counts and estimated energy consumption.
Significance. If the reported results are reproducible, the work is a meaningful step: it appears to be the first demonstration of Forward-Forward-style training on convolutional spiking networks, and the frozen black-box treatment of spiking neurons is a clean way to avoid surrogate-gradient approximations. The comparison in Table 1 covers a broad set of FF, STDP, and direct-feedback-alignment baselines, and the parameter/energy analysis in Table 2 is useful. The central novelty, however, is the class-complexity channel allocation of Eq. (4), and its empirical support currently rests on an incompletely specified feature-extraction step. Because the claimed margins over the closest CNN baseline are small, the leakage concern is not a presentation detail but a threat to the headline comparison.
major comments (4)
- [Inter-class Similarity Analysis, Eqs. (1)-(4)] The feature extractor that produces the vectors f_i in Eq. (1), and the data split D used to compute the class means, are not specified. If D includes test images, or if the features come from a network trained on the target classes, then the channel allocation of Eq. (4) uses test information, and the reported accuracy gain over uniform allocation (e.g., 75.64% vs. 75.28% on CIFAR-10) would not transfer to unseen data. Please specify the extractor and the split, and rerun the comparison using statistics computed from the training split only.
- [Experiments and results, Table 1] All accuracies are single-run values with no standard deviations or error bars. Given that the headline comparison against CwComp* rests on differences below 0.5 percentage points on MNIST and CIFAR-10, it is impossible to judge whether the improvement is statistically meaningful. Report mean and standard deviation over at least five seeds for the proposed method and for the re-implemented baselines.
- [Loss Function, Eq. (6)] The loss is written as cross-entropy applied directly to goodness values G_n,j, but no softmax or normalization is defined, and G_n,j is a sum of squared activations rather than a probability. Because each class j is assigned disjoint channels, the positive-class term does not appear to compete with the other classes, so the objective may simply drive up the absolute goodness of the correct class without suppressing incorrect ones. Please clarify how this acts as a classification objective, or compare with a properly normalized softmax cross-entropy variant.
- [Tables 1 and 2] FF* and CwComp* are described as 'our implemented models', but no implementation details, hyperparameters, or code are provided, and the energy values in Table 2 are analytical estimates. This makes the comparison hard to verify and weakens the claim of surpassing all existing FF-based SNN approaches. Provide architecture details and training hyperparameters for the re-implemented baselines, and state the assumptions behind the energy model.
minor comments (5)
- [Inter-class Similarity Analysis, Eqs. (2) and (4)] The symbol S is used both for the class-similarity matrix S_{c,c'} and for the channel-allocation variable S_j, which is confusing; please use distinct notation.
- [Inter-class Similarity Analysis, Eq. (1)] The text says 'where D denotes the number of datasets', but D should denote the dataset itself; please correct this typo.
- [Table 1] The header 'Test Er. (%)' is inconsistent with the entries, which are accuracies rather than errors; also correct the typos 'Fashion-MIST' and 'Fashion-MINST'.
- [Ablation Study] The ablation section states that on CIFAR-10 the advantage of the proposed allocation is 'not fully manifested' because inter-class difficulty variations are slight; this tempers the abstract's claim that the mechanism is a key source of improvement and should be acknowledged in the conclusion.
- [Network, Eq. (13)] Equation (13) is called an improved ReLU function, but it is a clipped quantization function; the connection to ReLU and to conversion-error reduction should be stated more explicitly.
Circularity Check
No significant circularity: the reported accuracies are empirical outcomes of a forward-forward training pipeline, not quantities defined by the method's inputs.
full rationale
The claimed derivation chain is self-contained: spiking layers are frozen black-box encoders, the goodness score is computed by Eq. (5) from allocated channel subsets, the loss is the cross-entropy in Eq. (6), and the channel allocation in Eq. (4) is computed from dataset class-mean features via Eqs. (1)-(3). None of these equations is defined in terms of the final test accuracies, and no fitted parameter is renamed as a prediction. The hyperparameter phi=2 is fixed, not tuned to the reported results. The self-citations (Xu, Liu, and Yang 2023, 2024; Pei et al. 2023) appear only in the introductory motivation and are not load-bearing for the method or its comparison. External citations (Hinton 2022; Papachristodoulou et al. 2024; Bu et al. 2023; Zheng et al. 2021) supply the FF framing, channel partitioning, improved ReLU, and temporal normalization without being used to equate an output with an input. The main caveat, that Eq. (1) does not identify the feature extractor or data split used for the class-mean vectors, is a potential data-leakage/soundness concern rather than a circularity: the allocation is a precomputed design choice, not a post-hoc fit to the accuracies being reported. Therefore no circular step can be exhibited from the paper's equations, and the circularity score is 0.
Assumptions & free parameters
free parameters (4)
- phi (channel allocation uniformity) =
2
- thresh (regularized output scale)
- L and lambda_l (quantization levels and scaling)
- channel allocation S_j =
data-derived
assumptions (4)
- domain assumption Forward-Forward local layer-wise objectives produce useful features for classification.
- domain assumption The IF neuron model in Eq. (11) with reset describes the spiking black-box layer accurately enough.
- domain assumption The modified ReLU from Bu et al. reduces conversion error between analog activations and spike encodings.
- ad hoc to paper Cross-entropy applied to goodness values G_n,j (Eq. 6) is a valid training objective without an explicit softmax or normalization.
Cite this review
Pith. "Pith review of FFGAF-SNN: The Forward-Forward Based Gradient Approximation Free Training Framework for Spiking Neural Networks." pith.science (2026). https://pith.science/paper/NWBWX34U
@misc{pith2026250723643,
author = {Pith},
title = {Pith review of: FFGAF-SNN: The Forward-Forward Based Gradient Approximation Free Training Framework for Spiking Neural Networks},
year = {2026},
howpublished = {\url{https://pith.science/paper/NWBWX34U}},
note = {Machine review of arXiv:2507.23643}
}
read the original abstract
Spiking Neural Networks (SNNs) offer a biologically plausible framework for energy-efficient neuromorphic computing. However, it is a challenge to train SNNs due to their non-differentiability, efficiently. Existing gradient approximation approaches frequently sacrifice accuracy and face deployment limitations on edge devices due to the substantial computational requirements of backpropagation. To address these challenges, we propose a Forward-Forward (FF) based gradient approximation-free training framework for Spiking Neural Networks, which treats spiking activations as black-box modules, thereby eliminating the need for gradient approximation while significantly reducing computational complexity. Furthermore, we introduce a class-aware complexity adaptation mechanism that dynamically optimizes the loss function based on inter-class difficulty metrics, enabling efficient allocation of network resources across different categories. Experimental results demonstrate that our proposed training framework achieves test accuracies of 99.58%, 92.13%, and 75.64% on the MNIST, Fashion-MNIST, and CIFAR-10 datasets, respectively, surpassing all existing FF-based SNN approaches. Additionally, our proposed method exhibits significant advantages in terms of memory access and computational power consumption.
Figures
Figures from the paper (2 more)
Reference graph
Works this paper leans on
-
[1]
Bohte, S. M.; Kok, J. N.; and La Poutr \'e , J. A. 2000. SpikeProp: backpropagation for networks of spiking neurons. In ESANN, volume 48, 419--424. Bruges
work page 2000
-
[2]
Bu, T.; Fang, W.; Ding, J.; Dai, P.; Yu, Z.; and Huang, T. 2023. Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347
arXiv 2023
-
[3]
Diehl, P. U.; and Cook, M. 2015. Unsupervised learning of digit recognition using spike-timing-dependent plasticity. Frontiers in computational neuroscience, 9: 99
work page 2015
-
[4]
U.; Neil, D.; Binas, J.; Cook, M.; Liu, S.-C.; and Pfeiffer, M
Diehl, P. U.; Neil, D.; Binas, J.; Cook, M.; Liu, S.-C.; and Pfeiffer, M. 2015. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), 1--8. ieee
2015
-
[5]
Dooms, T.; Tsang, I. J.; and Oramas, J. 2023. The trifecta: Three simple techniques for training deeper forward-forward networks. arXiv preprint arXiv:2311.18130
arXiv 2023
-
[6]
u gel, K.; Coquelin, D.; Weiel, M.; Debus, C.; Streit, A.; and G \
Fl \"u gel, K.; Coquelin, D.; Weiel, M.; Debus, C.; Streit, A.; and G \"o tz, M. 2023. Feed-forward optimization with delayed feedback for neural networks. arXiv preprint arXiv:2304.13372
-
[7]
Frenkel, C.; Lefebvre, M.; and Bol, D. 2021. Learning without feedback: Fixed random learning signals allow for feedforward training of deep neural networks. Frontiers in neuroscience, 15: 629892
work page 2021
-
[8]
Backpropagation-free Spiking Neural Networks with the Forward-Forward Algorithm
Ghader, M.; Kheradpisheh, S. R.; Farahani, B.; and Fazlali, M. 2025. Backpropagation-free Spiking Neural Networks with the Forward-Forward Algorithm. arXiv preprint arXiv:2502.20411
work page Pith review arXiv 2025
Show all 31 references
-
[9]
Hao, Y.; Huang, X.; Dong, M.; and Xu, B. 2020. A biologically plausible supervised learning method for spiking neural networks using the symmetric STDP rule. Neural Networks, 121: 387--395
2020
-
[10]
Hinton, G. 2022. The forward-forward algorithm: Some preliminary investigations. arXiv preprint arXiv:2212.13345, 2(3): 5
2022 arXiv
-
[11]
Karkehabadi, A.; Homayoun, H.; and Sasan, A. 2024. FFCL: forward-forward net with cortical loops, training and inference on edge without Backpropogation. In proceedings of the Great Lakes symposium on VLSI 2024, 626--632
2024
-
[12]
R.; Ganjtabesh, M.; Thorpe, S
Kheradpisheh, S. R.; Ganjtabesh, M.; Thorpe, S. J.; and Masquelier, T. 2018. STDP-based spiking deep convolutional neural networks for object recognition. Neural Networks, 99: 56--67
2018
-
[13]
Krizhevsky, A.; and Hinton, G. 2009. Learning multiple layers of features from tiny images. Technical report, University of Toronto
2009
-
[14]
LeCun, Y.; Cortes, C.; and Burges, C. J. 1998. The MNIST database of handwritten digits
1998
-
[15]
Lee, H.-C.; and Song, J. 2023. Symba: Symmetric backpropagation-free contrastive learning with forward-forward algorithm for optimizing convergence. arXiv preprint arXiv:2303.08418
2023 arXiv
-
[16]
Lemaire, E.; Cordone, L.; Castagnetti, A.; Novac, P.-E.; Courtois, J.; and Miramond, B. 2022. An analytical estimation of spiking neural networks energy efficiency. In International conference on neural information processing, 574--587. Springer
2022
-
[17]
Lorberbom, G.; Gat, I.; Adi, Y.; Schwing, A.; and Hazan, T. 2024. Layer collaboration in the forward-forward algorithm. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 14141--14148
2024
-
[18]
Ororbia, A.; and Mali, A. 2023. The predictive forward-forward algorithm. arXiv preprint arXiv:2301.01452
2023 arXiv
-
[19]
Ororbia, A. G. 2024. Contrastive signal--dependent plasticity: Self-supervised learning in spiking neural circuits. Science Advances, 10(43): eadn6076
2024
-
[20]
Papachristodoulou, A.; Kyrkou, C.; Timotheou, S.; and Theocharides, T. 2024. Convolutional channel-wise competitive learning for the forward-forward algorithm. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 14536--14544
2024
-
[21]
Pei, Y.; Xu, C.; Wu, Z.; Liu, Y.; and Yang, Y. 2023. ALBSNN: ultra-low latency adaptive local binary spiking neural network with accuracy loss estimator. Frontiers in Neuroscience, 17: 1225871
2023
-
[22]
Taylor, L.; King, A.; and Harper, N. 2022. Robust and accelerated single-spike spiking neural network training with applicability to challenging temporal tasks. arXiv preprint arXiv:2205.15286
2022 arXiv
-
[23]
B.; Del Ser, J.; Mart \' nez-Seras, A.; and Garcia-Bringas, P
Terres-Escudero, E. B.; Del Ser, J.; Mart \' nez-Seras, A.; and Garcia-Bringas, P. 2024. On the Robustness of Fully-Spiking Neural Networks in Open-World Scenarios using Forward-Only Learning Algorithms. arXiv e-prints, arXiv--2407
2024
-
[24]
WANG, J.; WANG, J.-p.; SUN, W.-h.; and CHEN, S. 2020. A neuromorphic hardware design of a spiking convolutional neural network. Microelectronics & Computer, 37(12): 1--5
2020
-
[25]
Xiao, H.; Rasul, K.; and Vollgraf, R. 2017. Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747
2017 arXiv
-
[26]
Xu, C.; Liu, Y.; and Yang, Y. 2023. Ultra-low latency spiking neural networks with spatio-temporal compression and synaptic convolutional block. Neurocomputing, 550: 126485
2023
-
[27]
Xu, C.; Liu, Y.; and Yang, Y. 2024. STCSNN: High energy efficiency spike-train level spiking neural networks with spatio-temporal conversion. Neurocomputing, 607: 128364
2024
-
[28]
Zhang, Y.; Inoue, K.; Nakajima, M.; Hashimoto, T.; Kuniyoshi, Y.; and Nakajima, K. 2024. Training Spiking Neural Networks via Augmented Direct Feedback Alignment. arXiv preprint arXiv:2409.07776
2024 arXiv
-
[29]
Zheng, H.; Wu, Y.; Deng, L.; Hu, Y.; and Li, G. 2021. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 35, 11062--11070
2021
-
[30]
, " * write output.state after.block = add.period write newline
ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...
-
[31]
write newline
" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.