Spiffy: Efficient Implementation of CoLaNET for Raspberry Pi

Andrey Derzhavin; Denis Larionov

REVIEW 4 major objections 5 minor 14 references

Spiffy: Efficient Implementation of CoLaNET for Raspberry Pi

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that a single Rust-implemented CoLaNET network reaches 91.08% accuracy on MNIST at sub-millisecond timesteps on a Raspberry Pi 4, beating single-network CoLaNET by about two points.

desk verdict Useful open-source Rust port of CoLaNET for edge devices, but the accuracy gain over the baseline is not established because the hyperparameter search may have peeked at the test set, and 'hardware implementation' is an overstatement. read the letter →

arxiv 2506.18306 v1 pith:GTTEONED submitted 2025-06-23 cs.NE cs.AI

Andrey Derzhavin , Denis Larionov This is my paper · ORCID

classification cs.NEcs.AI

keywords spikingneuralnetworksCoLaNETlocallearningon-deviceMNISTRaspberryPiRustimplementationsynapticresourcefunction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Spiffy is a from-scratch Rust implementation of the CoLaNET spiking neural network, written to run on ordinary hardware rather than specialized neuromorphic chips or proprietary simulators. The paper's central claim is that one Spiffy network, trained with local rules on the MNIST dataset, reaches 91.08% mean accuracy over ten runs (rounded to 92% in the abstract), beating the 89% of an optimized single-network CoLaNET and approaching the 95% of a 15-network CoLaNET ensemble with 15× fewer neurons. On a Raspberry Pi 4 the same code trains on-device at 0.9 ms per timestep and runs inference at 0.45 ms per timestep, with 0.1 ms during silent timesteps. A sympathetic reader would care because this is a concrete demonstration that edge hardware, not just neuromorphic accelerators, can host spike-based local learning end-to-end.

What carries the argument

The load-bearing object is the synaptic resource $W$ and the function that turns it into a synaptic weight $w$. CoLaNET uses a saturating rational function of $W$; Spiffy instead uses $w = \min(w_{\max}, \max(w_{\min}, W))$, which is cheaper to evaluate and, in the reported experiments, slightly more accurate. On top of that map, the plasticity rule at time step 19 does three things: if the column for the correct label produced no spikes, all neurons in that column receive a dopamine reward (the group update); if some neurons fired correctly, one random correct neuron is potentiated; and every neuron that fired for the wrong label is depressed by the same dopamine quantum. Stochasticity comes from random weight initialization, while the winner-take-all and dopamine modules are implemented programmatically rather than as spiking neurons, which is what makes the whole network small and fast.

What would settle it

Run the same AdamW search on MNIST with the test set held out, select the best configuration on a validation split, then measure that exact configuration once on the untouched test set; if the accuracy lands within noise of the 89.84% (classical resource) or 89.87% (zero initialization) variants, the claimed benefit of the Spiffy modifications is not a genuine improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single Spiffy network—10 columns, 15 plastic neurons per column, plus a shared label module—outperforms a single optimized CoLaNET network on MNIST, reaching 91.08% ± 0.26% over ten independent runs. The improvement is attributed to three changes made together: replacing CoLaNET's saturating synaptic-resource function with the linear map $w = \min(w_{\max}, \max(w_{\min}, W))$, initializing synaptic resources randomly instead of at zero, and adding a group-update dopamine rule that rewards every neuron in the correct column when no neuron in that column fired. The paper's ablations show the linear function alone reaches 89.84% and random initialization with the linear function reaches 89.87%, so the advertised gain appears to come from the combination; the paper also observes that the group update makes learned receptive fields within a column similar to each other and flags this as a drawback.

Load-bearing premise

The 91.08% result is reported for the best configuration found by an AdamW hyperparameter search, and the paper does not describe a separate validation set; if the MNIST test set was used to choose that configuration, the figure is an optimistic estimate and the claimed two-point edge over the baselines is not established.

Editorial extensions

If this is right

A single Spiffy network, with 15× fewer neurons than the published CoLaNET ensemble, is claimed to beat single-network CoLaNET by about two percentage points on MNIST.
On-device training with local spiking rules is fast enough for interactive edge use: 0.9 ms per training timestep and 0.45 ms per inference timestep on a Raspberry Pi 4.
The linear synaptic-resource function is presented as both computationally cheaper and at least as accurate as the classical saturating function in the tested configuration.
Manual implementation in Rust removes the dependency on a proprietary simulator, so the architecture can be ported to any platform with a Rust toolchain.
The reported timings and accuracy can serve as a software benchmark against which future specialized neuromorphic implementations of CoLaNET can be compared.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An implication the authors leave implicit is that the group update's tendency to make within-column receptive fields similar might be counteracted by rewarding a random subset of the silent correct column, preserving the learning signal while adding diversity; this is a testable variant, not something the paper considers.
The ablation numbers suggest the three modifications interact rather than stack independently: neither the linear map nor random initialization alone reaches 91%, so a reader should treat the joint configuration, not any single change, as the contribution.
Because the paper reports only the Raspberry Pi timings, a natural extension would be to measure the same Rust code on other ARM and RISC-V boards to separate algorithmic efficiency from the Pi's memory and clock characteristics.
The comparison is against the original CoLaNET only; an independent benchmark against other local-learning SNNs on MNIST would locate Spiffy's accuracy relative to the broader edge-SNN landscape.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This one is a mixed bag. Spiffy is a Rust implementation of CoLaNET, an SNN architecture, with three tweaks: linear resource function, random weight init, and group update. What's genuinely useful is the open-source code and the honest engineering narrative. The authors walk through their reverse-engineering of the proprietary ArNI-X simulator, the initial 76% accuracy, and the iterative changes that got them to 91%. The latency figures on a Raspberry Pi 4 (0.9 ms/step training, 0.45 ms/step inference) are plausible and worth having, and the heatmaps in Fig. 3 are a nice touch even though they show the class prototypes are not cleanly separated.

The soft spots are real. The abstract says 92% but Table 1 says 91.08%; that's sloppy. The bigger issue is that the hyperparameters were tuned with AdamW and there is no mention of a held-out validation split. If the 'best configuration' was selected on the MNIST test set, the 91.08% is an optimistic post-selection estimate, not an unbiased generalization number. The 89% CoLaNET baseline is taken from prior work and not re-evaluated under the same protocol, so the 2-point margin could well shrink or vanish. Also, they call this the 'first hardware implementation' while it is software; Rust on a Raspberry Pi is not a hardware implementation. The latency measurement method is thin, with no detail on how timestamps were taken or what exactly was timed.

None of this sinks the core engineering value. The paper is an implementation study with a real artifact, and the code is available. The central accuracy claim needs either a clean nested evaluation or an explicit statement that the test set was used only for the final reported runs. And the 'first hardware' phrase should be corrected.

Who is this for? People building SNN prototypes on edge devices, or anyone wanting a portable reference implementation of CoLaNET. It's not a fundamental advance. I would send it to peer review because it is an honest, reproducible artifact that could be improved with a solid evaluation section. A serious referee would ask for the validation protocol and a corrected abstract, but the paper is worth that effort.

Referee Report

4 major / 5 minor

Summary. The paper presents Spiffy, a Rust implementation of the CoLaNET spiking neural network architecture, optimized for commodity hardware. Spiffy uses a single network instance with 150 plastic neurons, a linear synaptic resource function, random weight initialization, and a group-update plasticity mechanism. The authors report 91.08% average accuracy on MNIST over 10 runs (rounded to 92% in the abstract), exceeding the 89% accuracy of the previously reported optimized single-network CoLaNET configuration. They also report latencies of 0.9 ms per training timestep, 0.45 ms per inference timestep, and 0.1 ms per silent timestep on a Raspberry Pi 4. The code is released as open source.

Significance. If the claimed accuracy and latency numbers are reliable, the paper provides a useful engineering reference: a complete, portable, open-source implementation of CoLaNET with local learning that runs on a Raspberry Pi, plus evidence that a linear resource function and group updates improve over the original mechanics. The paper's strengths are its concrete algorithmic description (Algorithm 1), reproducibility-oriented open-source release, and reporting of means and standard deviations over 10 runs. The significance is nevertheless constrained by the lack of a disclosed hyperparameter-selection protocol, the absence of a re-evaluated CoLaNET baseline under the same conditions, and incomplete latency measurement details. The central quantitative claims are therefore not yet fully established.

major comments (4)

[Section 3, Table 1] The reported 91.08% accuracy is labeled as the 'best Spiffy configuration' obtained by optimizing hyperparameters with AdamW, but the manuscript does not disclose how the test set was used during this optimization. If the MNIST test set was used for model selection, the reported average is an optimistic post-selection estimate, and the 10 random-initialization runs do not correct for this bias because they are repetitions of the already-selected configuration. Please specify the data split used for AdamW optimization, the objective function, the number of trials, and a statement of whether any test-set information influenced the final configuration. Absent this, the comparison with the 89% single-network CoLaNET result is not conclusive.
[Section 3 (latency paragraph)] The latency figures (0.9 ms per training cycle, 0.45 ms per inference cycle, 0.1 ms per silent cycle) are presented without any measurement protocol. The manuscript does not state how many repeated measurements were taken, whether warm-up was excluded, which Raspberry Pi model and clock speed were used, what compiler flags or optimization levels were applied, and whether the reported times include I/O or only computation. Without these details the efficiency claims are not reproducible and cannot be compared with other implementations.
[Section 2.1 and Section 4] The claimed advantage over a single-network CoLaNET rests on comparing Spiffy's 91.08% to the 89% figure reported in [9]. The 89% value is not re-evaluated with the same protocol, same data split, same hyperparameter search, or same evaluation software, so the difference could be due to experimental procedure rather than architectural improvements. A re-evaluation of the original CoLaNET configuration under the same evaluation protocol, or at least a careful statement about comparability, is needed before claiming Spiffy 'outperforms' CoLaNET.
[Section 3, Tables 1-3] The paper compares three configurations (linear vs. classic resource function; random vs. zero initialization) and concludes that the linear/random combination is best, but it provides no statistical test or confidence interval for the differences among the three average accuracies. The per-class standard deviations are large (up to 5.0), and the averages are within a narrow range, so the claimed benefits of random initialization and the linear function should be supported by a proper statistical comparison or by reporting the full distribution of per-run accuracies.

minor comments (5)

[Abstract] The abstract states '92% accuracy', but Table 1 reports 91.08% with standard deviation 0.26. The abstract should quote 91.08% (or '91.1%') and preferably include the standard deviation.
[General] The paper repeatedly calls Spiffy a 'hardware implementation' (e.g., Introduction and Section 4), but Spiffy is a software implementation running on general-purpose hardware. Suggest replacing 'hardware implementation' with 'software implementation on commodity hardware' to avoid misleading terminology.
[Section 2.2] The text refers to 'Listing 1' but the algorithm is presented as 'Algorithm 1'. Please align the reference.
[Section 3] The description of the AdamW optimization gives only the CPU models on which it was run. Please provide the hyperparameter ranges, the number of optimization iterations, and the chosen hyperparameter values for the best configuration, so that the experiments can be reproduced.
[References] Reference [10] for ArNI-X is given only as a year, with no version or public repository identifier; since the paper says the authors used the ArNI-X source code, a version or commit reference would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the work is an empirical implementation study; the only self-citation is a non-load-bearing introductory survey.

full rationale

Spiffy is an implementation study, not a derivation. Its central claims are that the Rust implementation runs on a Raspberry Pi with the reported latencies and that a particular configuration reaches 91.08% accuracy on MNIST, exceeding the cited 89% single-network CoLaNET baseline. Neither claim is circular: the accuracy is an empirical measurement of a trained SNN, and the baseline comes from Kiselev's external CoLaNET papers ([7, 8, 9]), not from parameters fitted in this paper. The only self-citation is reference [1], a general neuromorphic survey co-authored by one of the present authors, used for introductory motivation ('According to [1], modern artificial intelligence (AI) systems still fall far short...'); it is not load-bearing for any numerical or architectural result. The AdamW hyperparameter optimization described in Section 3 could produce an optimistic accuracy estimate if the MNIST test set was used for model selection, but that is a statistical validity concern (selection bias), not circularity: the reported 91.08% is the measured accuracy of a trained configuration, not a quantity that is definitionally fixed by the optimization target. No equation is both input and output, and no fitted parameter is renamed as a prediction. Therefore no circular step is present.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

Spiffy introduces no new theoretical entities. Its accuracy claim rests on the external CoLaNET architecture and on a set of hyperparameters tuned by AdamW, whose values are not disclosed.

free parameters (1)

Spiffy hyperparameter set = Not reported in the paper
The reported 91.08% accuracy depends on hyperparameters such as the dopamine quantum, membrane time constant, and weight bounds; these were optimized with AdamW (Section 3), and the values are not disclosed.

assumptions (4)

domain assumption CoLaNET architecture and its learning rules as described in refs [7,8,9] are a valid basis for MNIST classification
The paper builds directly on Kiselev's CoLaNET papers and ArNI-X; it does not independently justify the architecture's design.
standard math LIF neuron with delta-synapse current model (Eq. 1) accurately models the membrane potential dynamics
Standard model from Gerstner et al. [2]; assumed without validation.
domain assumption Deterministic spike-count encoding is equivalent to CoLaNET's probabilistic rate coding for static pixel inputs
Section 2.2 states this equivalence; it is plausible but not proven in the paper.
domain assumption MNIST labels are presented without noise and the reward/punishment (dopamine) signal is perfectly correlated with the correct class
Section 2.2 and Algorithm 1 rely on a supervised label signal at step 19; real-world labels may be noisy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Spiffy: Efficient Implementation of CoLaNET for Raspberry Pi." pith.science (2026). https://pith.science/paper/GTTEONED

@misc{pith2026250618306,
  author       = {Pith},
  title        = {Pith review of: Spiffy: Efficient Implementation of CoLaNET for Raspberry Pi},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GTTEONED}},
  note         = {Machine review of arXiv:2506.18306}
}

read the original abstract

This paper presents a lightweight software-based approach for running spiking neural networks (SNNs) without relying on specialized neuromorphic hardware or frameworks. Instead, we implement a specific SNN architecture (CoLaNET) in Rust and optimize it for common computing platforms. As a case study, we demonstrate our implementation, called Spiffy, on a Raspberry Pi using the MNIST dataset. Spiffy achieves 92% accuracy with low latency - just 0.9 ms per training step and 0.45 ms per inference step. The code is open-source.

Figures

Figures reproduced from arXiv: 2506.18306 by the authors.

**Figure 2.** Spiffy - inference During the training phase, on the final time step of the corresponding interval (step 19), the label is presented via the LAB component, and the DOP is activated ( [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Heatmaps of Spiffy network L-layer neuron receptive fields post-training on MNIST. Weights are zero [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages

[9]

A digital machine learning algorithm simulating spiking neural network CoLaNET.arXiv preprint arXiv:2503.17111, 2025

Mikhail Kiselev. A digital machine learning algorithm simulating spiking neural network CoLaNET.arXiv preprint arXiv:2503.17111, 2025

arXiv 2025
[1]

Neuromorphic artificial intelligence systems.Frontiers in Neuroscience, 16, 2022

Dmitry Ivanov, Aleksandr Chezhegov, Mikhail Kiselev, Andrey Grunin, and Denis Larionov. Neuromorphic artificial intelligence systems.Frontiers in Neuroscience, 16, 2022

work page 2022
[2]

Kistler, Richard Naud, and Liam Paninski

Wulfram Gerstner, Werner M. Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. InNeuronal Dynamics, 2014

work page 2014
[3]

Palmarium Academic Publishing, 2020

Mikhail Kiselev.Spike Neural Networks: Information Representation, Learning, Memory. Palmarium Academic Publishing, 2020

work page 2020
[4]

The hardware lottery.Communications of the ACM, 64:58 – 65, 2020

Sara Hooker. The hardware lottery.Communications of the ACM, 64:58 – 65, 2020

work page 2020
[5]

Rumelhart, Geoffrey E

David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors.Nature, 323:533–536, 1986

work page 1986
[6]

Softhebb: Bayesian inference in unsupervised hebbian soft winner-take-all networks.Neuromorphic Computing and Engineering, 2, 2021

Timoleon Moraitis, Dmitry Toichkin, Adrien Journé, Yansong Chua, and Qinghai Guo. Softhebb: Bayesian inference in unsupervised hebbian soft winner-take-all networks.Neuromorphic Computing and Engineering, 2, 2021

work page 2021
[7]

Colanet – a spiking neural network with columnar layered architecture for classification.arXiv preprint arXiv:2409.01230, 2024

Mikhail Kiselev. Colanet – a spiking neural network with columnar layered architecture for classification.arXiv preprint arXiv:2409.01230, 2024

arXiv 2024

Show all 14 references

[8]

Classifying images with colanet spiking neural network - the mnist example.arXiv preprint arXiv:2409.07833, 2024

Mikhail Kiselev. Classifying images with colanet spiking neural network - the mnist example.arXiv preprint arXiv:2409.07833, 2024

2024 arXiv
[10]

Arni-x: A simple but powerful and flexible simulator for spiking neural networks, 2020

Mikhail Kiselev. Arni-x: A simple but powerful and flexible simulator for spiking neural networks, 2020

2020
[11]

Mikhail V . Kiselev. A synaptic plasticity rule providing a unified approach to supervised and unsupervised learning.2017 International Joint Conference on Neural Networks (IJCNN), 2017

2017
[12]

Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition.Proceedings of the IEEE, 86(11):2278–2324, 1998

1998
[13]

Grishanov

Alexander N. Grishanov. Neuromorphic processor AltAI for energy-efficient computing.Nanoindustry Russia, 96:531–538, 2019

2019
[14]

Kaspersky neuromorphic platform (knp), 2024

KNP. Kaspersky neuromorphic platform (knp), 2024. 7

2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [9]

A digital machine learning algorithm simulating spiking neural network CoLaNET.arXiv preprint arXiv:2503.17111, 2025

Mikhail Kiselev. A digital machine learning algorithm simulating spiking neural network CoLaNET.arXiv preprint arXiv:2503.17111, 2025

arXiv 2025

[2] [1]

Neuromorphic artificial intelligence systems.Frontiers in Neuroscience, 16, 2022

Dmitry Ivanov, Aleksandr Chezhegov, Mikhail Kiselev, Andrey Grunin, and Denis Larionov. Neuromorphic artificial intelligence systems.Frontiers in Neuroscience, 16, 2022

work page 2022

[3] [2]

Kistler, Richard Naud, and Liam Paninski

Wulfram Gerstner, Werner M. Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. InNeuronal Dynamics, 2014

work page 2014

[4] [3]

Palmarium Academic Publishing, 2020

Mikhail Kiselev.Spike Neural Networks: Information Representation, Learning, Memory. Palmarium Academic Publishing, 2020

work page 2020

[5] [4]

The hardware lottery.Communications of the ACM, 64:58 – 65, 2020

Sara Hooker. The hardware lottery.Communications of the ACM, 64:58 – 65, 2020

work page 2020

[6] [5]

Rumelhart, Geoffrey E

David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors.Nature, 323:533–536, 1986

work page 1986

[7] [6]

Softhebb: Bayesian inference in unsupervised hebbian soft winner-take-all networks.Neuromorphic Computing and Engineering, 2, 2021

Timoleon Moraitis, Dmitry Toichkin, Adrien Journé, Yansong Chua, and Qinghai Guo. Softhebb: Bayesian inference in unsupervised hebbian soft winner-take-all networks.Neuromorphic Computing and Engineering, 2, 2021

work page 2021

[8] [7]

Colanet – a spiking neural network with columnar layered architecture for classification.arXiv preprint arXiv:2409.01230, 2024

Mikhail Kiselev. Colanet – a spiking neural network with columnar layered architecture for classification.arXiv preprint arXiv:2409.01230, 2024

arXiv 2024