Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Hardware-Aware Fine-Tuning of Spiking Q-Networks on the SpiNNaker2 Neuromorphic Platform

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A quantized 8-bit spiking Q-network deployed on SpiNNaker2 matches a GPU-trained policy on two control tasks while using up to 32 times less energy per episode.

desk verdict The 24-32x energy claim is probably an artifact of excluding the enforced real-time pacing delay; the deployment work itself is real, but the headline comparison needs fixing. read the letter →

arxiv 2507.23562 v1 pith:QR5U6OTJ submitted 2025-07-31 cs.LG cs.AR

classification cs.LGcs.AR
keywords spikingneuralnetworksreinforcementlearningQ-learningquantizationneuromorphichardwareSpiNNaker2energyefficiencyclosed-loopcontrol
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

The paper tries to establish that a Q-learning agent expressed as a spiking neural network can be moved from floating-point simulation to 8-bit integer weights on the SpiNNaker2 neuromorphic chip without losing policy quality: CartPole still scores 200 and Acrobot -67, matching or slightly beating the software simulator baseline. The reason this matters is practical: the chip draws about 0.33 W for these episodes, against 7.3 to 7.4 W for a GTX 1650, giving 24x and 32x reductions in energy per episode while inference latency stays on par. The authors attribute the lossless transfer to hardware-aware fine-tuning: layer-wise weight scaling before quantization and a two-stage threshold search on the chip. If the energy accounting holds, low-power neuromorphic hardware becomes a credible platform for real-time closed-loop deep reinforcement learning.

What carries the argument

The argument is carried by three devices: post-training quantization with a uniform layer-wise scaling factor that expands the dynamic range of small weights, with a scaling factor of 3 for CartPole and 32 for Acrobot; a two-stage on-chip threshold grid search that rebalances neuron excitability after quantization; and a signed two-neuron rate-coding input scheme plus a final membrane-potential readout from output neurons with a very high threshold, so the network's decision is the action with the largest integrated voltage. A tick-based scheduling loop, using 1 ms ticks and 10 or 20 ticks per control step, paces closed-loop interaction with the environment.

What would settle it

Measure the board's total energy and wall-clock time over complete closed-loop episodes with the pacing delay explicitly included; if per-episode energy or latency comes out materially higher than the reported 0.006 J and 0.010 J or the reported 0.020 s and 0.030 s, the 24x and 32x reductions do not hold in deployed use.

Watch

Extended reading notes

Core claim

The central claim is that a quantized 8-bit spiking deep Q-network, trained in full precision and then scaled and rounded, can be deployed on SpiNNaker2 and solve the CartPole and Acrobot control tasks with no degradation in reward. On the chip, CartPole-v0 reaches the maximum reward of 200 and Acrobot-v1 reaches -67, equal to or slightly better than the simulator-trained baseline. The deployment achieves these rewards while measuring 0.006 J and 0.010 J per episode, compared with 0.145 J and 0.321 J on a GTX 1650, corresponding to 24x and 32x reductions in energy. The authors' evidence includes membrane-potential traces and spike rasters showing task-aligned dynamics, plus a threshold-sensitivity sweep.

Load-bearing premise

The headline energy advantage assumes that SpiNNaker2's reported power draw, about 0.33 W, and episode durations of 0.020 s for CartPole and 0.030 s for Acrobot capture the full closed-loop episode, including the artificial pacing delay used to match the environment's control rate.

Editorial extensions

If this is right

  • SpiNNaker2 can run a complete closed-loop RL policy at the environment's control rate, 100 Hz for CartPole and 50 Hz for Acrobot, without dropping reward.
  • The same quantization-and-threshold-tuning recipe should apply to other small fully connected spiking policies, with the caveat that the scaling factor and thresholds must be re-tuned per task.
  • Energy per episode is low enough, in the millijoule range, that battery-powered edge agents could sustain many episodes of control between charges.
  • Latency parity with a GPU means the neuromorphic chip is not trading speed for efficiency on these workloads.

Reading between the lines

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

  • If the closed-loop pacing delay is included in the reported episode duration, the headline energy figures describe active inference rather than the full deployed episode; a system-level measurement that includes the delay and host overhead would be a stronger test of the 32x claim.
  • The same hardware-aware fine-tuning loop, quantize, scale, and sweep thresholds, is directly testable on larger discrete-action tasks or on continuous control with population-coded outputs.
  • The near-lossless membrane-potential readout suggests that the output layer could support on-chip learning rules without changing the encoding, connecting to in-the-loop training.
  • A fairer comparison would include the GPU's full-system power rail rather than its board draw; even then the chip's sub-watt operation is the qualitative difference.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper presents a pipeline for training spiking Q-networks on CartPole-v0 and Acrobot-v1, fine-tuning them for 8-bit quantized deployment on the SpiNNaker2 neuromorphic chip, and benchmarking them against an Nvidia GTX 1650 GPU. The authors report that the deployed quantized policies preserve episode rewards, that SpiNNaker2 consumes about 0.335 W and achieves 24x to 32x lower energy per episode than the GPU, and that inference latency is on par with or better than the GPU. The main evidence is Table III, supported by quantization-scale ablations, a two-stage threshold sensitivity analysis, and spike and voltage traces from on-chip execution. The paper also provides a public code repository.

Significance. If the reported efficiency numbers survive the measurement corrections discussed below, this would be a valuable contribution: it would be one of the few closed-loop deep-RL demonstrations on SpiNNaker2, with a practical quantization and threshold-tuning recipe, a direct GPU baseline comparison, and public code. The deployment pipeline itself, including the two-stage threshold search and the per-layer weight scaling analysis, is credibly executed, and the on-chip spike and voltage traces support that real spiking dynamics are being exercised. However, the headline up-to-32x energy reduction claim is currently not substantiated because the energy accounting omits the real-time pacing delay defined in Section IV-C, and the power measurement methodology is absent from the manuscript.

major comments (3)
  1. [IV-C, Table III] The energy and latency comparison in Table III appears to be computed from Tinf, the active inference time, rather than from the wall-clock control interval Tctrl that the closed-loop system enforces. For CartPole, Tctrl = Tsim x Ttick = 10 ms per step, so a 200-step episode requires at least 2.0 s of wall-clock time; Table III reports Duration = 0.020 s and Energy = 0.006 J for that episode, which equals 0.335 W x 0.020 s. The paper does not state that the chip is powered down during Tdelay = max(0, Tctrl - Tinf), and the GTX 1650 duration of 0.0197 s shows that the GPU baseline was not subject to the same real-time pacing. If SpiNNaker2 remains powered during Tdelay, the closed-loop per-episode energy for CartPole would be on the order of 0.335 W x 2.0 s = 0.67 J, exceeding the reported GPU value of 0.145 J and reversing the claimed 24x advantage. The paper must either include Tdelay in both platforms' durations and energy figures, or clearly separate active-compute metrics from wall-clock closed-loop metrics; as written, the headline energy and latency claims are not substantiated.
  2. [IV-D, Table III] The power values in Table III (0.335 W on SpiNNaker2 and 7.39 W on GTX 1650) are not supported by any description of the measurement procedure. The manuscript does not state whether power was measured on the chip itself or on the whole test board, what instrument and sampling method were used, whether the GPU reading includes memory, CPU, or data-transfer overhead, or whether idle power was subtracted. Because the central claim is an energy-per-episode comparison, these omissions make the efficiency numbers unverifiable. The authors should specify the measurement setup and report the measurement conditions in sufficient detail for the comparison to be reproduced.
  3. [IV-A, IV-B, V] The claim of lossless deployment in Section V is weakened because the two tuning parameters, the weight scaling factor lambda in Eq. (5) and the hidden-layer thresholds, were selected using the same evaluation rewards that are then reported as the deployment result. Section IV-A states that lambda = 3 and lambda = 32 were chosen because they gave the highest average rewards, and Section IV-B describes selecting thresholds by sweeping the average reward over 5 episodes. Table III then reports the resulting rewards of 200 and -67 without any held-out test set, error bars, or number of evaluation episodes. The deployment may still be valid as a tuned engineering result, but the paper should provide an out-of-sample evaluation and report statistics over multiple episodes and training seeds to support the lossless-deployment claim.
minor comments (5)
  1. [I] The Introduction contains a typo: 'ariti-ficial intelligence' should be 'artificial intelligence'.
  2. [III-C] The phrase 'two-neruon encoding scheme' contains a typo; it should be 'two-neuron encoding scheme'.
  3. [IV-C] The sentence 'A consistent finding across both environments is that across both tasks, we observe no degradation' is redundant; it should be shortened to avoid repeating 'across both environments' and 'across both tasks'.
  4. [V] The code URL in the Conclusion contains spaces and may not resolve correctly; please verify that it is a working link and format it properly.
  5. [Table III, Fig. 4] The Rewards row in Table III and the sensitivity curves in Fig. 4 are based on average rewards over only 5 episodes, but no standard deviation or number of episodes is reported; please add error bars and state the number of evaluation episodes for each reported reward.

Circularity Check

1 steps flagged · score 4.0 of 10

Lossless-deployment claim is partially circular: λ and hidden-layer thresholds were tuned on the same reward metric later cited as evidence; the headline energy comparison is direct measurement.

  1. fitted input called prediction [Sections IV-A/IV-B (λ and threshold selection) -> IV-C (lossless claim); Table III]
    "These values were selected empirically after extensive experiments measuring task performance across a range of scaling factors, where the highest average rewards were obtained at λ = 3 for CartPole-v0 and λ = 32 for Acrobot-v1. ... The threshold yielding the highest average reward was selected. ... A consistent finding across both environments is that across both tasks, we observe no degradation in policy performance when deploying the trained DSQN models on SpiNNaker2."

    The reward values reported as evidence of lossless deployment (Table III: 200 and -67) are the same objective used to select the free parameters. λ was chosen as the setting giving 'the highest average rewards,' and hidden-layer thresholds were chosen as the settings giving 'the highest average reward' over 5 evaluation episodes. The subsequent 'no degradation' conclusion is therefore the selected configuration's score on the fitting metric, not an independent validation: the evidence coincides with the tuning objective, so it cannot support the lossless-deployment claim on its own. The energy and latency comparison in Table III is a direct measurement and is not circular.

full rationale

The paper's headline efficiency result (up to 32x energy reduction) is a direct measurement of power and duration on SpiNNaker2 versus a GTX 1650 GPU, so that central claim is not fit-derived or definitionally forced. However, the 'lossless deployment' claim is weakened by a partial circularity: the quantization scaling factor λ and the hidden-layer LIF thresholds were selected to maximize the same reward metric that is then reported as evidence that deployment caused 'no degradation in policy performance.' The tuning objective and the claimed validation metric coincide, making the losslessness evidence a restatement of the selection rule rather than an independent prediction. This is a moderate, localized circularity; the energy claim retains independent content. The separate concern that Table III's energy/duration may exclude the enforced real-time pacing delay Tdelay is a measurement-validity issue rather than circularity, so it is not counted in the circularity score.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The pipeline adds two empirically fitted deployment parameters: a quantization scaling factor and per-layer hidden thresholds. No new physical entities, particles, forces, or conserved quantities are introduced. The remaining assumptions are standard RL math, rate-coding fidelity, and SpiNNaker2 hardware behavior.

free parameters (2)
  • Weight scaling factor lambda (Eq. 5) = lambda = 3 for CartPole-v0, lambda = 32 for Acrobot-v1
    Chosen empirically to maximize task reward after 8-bit quantization; it controls quantization resolution and clipping risk, and it underlies the lossless-deployment claim.
  • Hidden-layer firing thresholds after quantization = Not reported in the paper
    A two-stage grid search over both hidden layers, selecting values by 5-episode average reward; used to rebalance neuron excitability after quantization.
assumptions (5)
  • standard math Q-learning and the Bellman equation provide a valid training objective for the spiking network.
    Background in Section II-A, Eqs. 1-3; standard RL framework assumed without proof.
  • domain assumption Rate coding with Poisson spike trains over T timesteps faithfully represents continuous observation values for these tasks.
    Section III-C; follows prior work [17] and [29], with no ablation on encoding fidelity.
  • domain assumption SpiNNaker2 runs on 1 ms ticks and uses 8-bit signed synapse weights with 32-bit floating-point neuron state.
    Section III-B and IV-C; hardware behavior cited from the authors' own prior work [28] and py-spinnaker2 [30].
  • ad hoc to paper A uniform per-layer scaling factor and manually swept thresholds are sufficient to recover full-precision policy performance after 8-bit quantization.
    Section IV-A and IV-B; this is the core hardware-aware fine-tuning assumption, validated only on two tasks.
  • domain assumption Output-layer membrane potentials after the final timestep can be compared across neurons to select the argmax action.
    Section III-C; decoding scheme from [17] and [29], empirically supported by the traces in Figure 3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hardware-Aware Fine-Tuning of Spiking Q-Networks on the SpiNNaker2 Neuromorphic Platform." pith.science (2026). https://pith.science/paper/QR5U6OTJ

@misc{pith2026250723562,
  author       = {Pith},
  title        = {Pith review of: Hardware-Aware Fine-Tuning of Spiking Q-Networks on the SpiNNaker2 Neuromorphic Platform},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QR5U6OTJ}},
  note         = {Machine review of arXiv:2507.23562}
}
read the original abstract

Spiking Neural Networks (SNNs) promise orders-of-magnitude lower power consumption and low-latency inference on neuromorphic hardware for a wide range of robotic tasks. In this work, we present an energy-efficient implementation of a reinforcement learning (RL) algorithm using quantized SNNs to solve two classical control tasks. The network is trained using the Q-learning algorithm, then fine-tuned and quantized to low-bit (8-bit) precision for embedded deployment on the SpiNNaker2 neuromorphic chip. To evaluate the comparative advantage of SpiNNaker2 over conventional computing platforms, we analyze inference latency, dynamic power consumption, and energy cost per inference for our SNN models, comparing performance against a GTX 1650 GPU baseline. Our results demonstrate SpiNNaker2's strong potential for scalable, low-energy neuromorphic computing, achieving up to 32x reduction in energy consumption. Inference latency remains on par with GPU-based execution, with improvements observed in certain task settings, reinforcing SpiNNaker2's viability for real-time neuromorphic control and making the neuromorphic approach a compelling direction for efficient deep Q-learning.

Figures

Figures reproduced from arXiv: 2507.23562 by the authors.

Figure 1
Figure 1. Pipeline Overview: Closed-loop SNN-based RL with SpiNNaker2. Observations from the OpenAI Gym environ￾ments are processed by a quantized spiking neural network deployed on the SpiNNaker2 neuromorphic platform. The system runs in closed-loop, with selected actions sent back to the environment. The SNN architecture includes two fully connected hidden layers; additional details are provided in Table II. transitions (s,… view at source ↗
Figure 2
Figure 2. Effect of Uniform Full-Layer Quantization Scaling in CartPole-v0 and Acrobot-v1. (a) In CartPole-v0, applying a uniform scaling factor across all layers (×3) increases the dynamic range of early-layer weights without saturating the network, preserving temporal sparsity and enabling maximum reward performance. (b) For Acrobot-v1, a more challenging control task, full-layer scaling (×32) is essential to expand hidden-… view at source ↗
Figure 3
Figure 3. Voltage dynamics in CartPole-v0 and Acrobot-v1 recorded for one episode on SpiNNaker2. Horizontal axis: the number of time steps of the episode multiplied by the simulation time of each trial, listed in Table II. The dots indicate which action was selected at each time step, based on the neuron with the higher membrane potential. In the example runs shown here, the CartPole-v0 episode lasted 200 time steps (reward =… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Average rewards over 5 episodes for various hidden [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Spike activity for CartPole-v0 and Acrobot-v1 across input and hidden layers. Each column corresponds to a task (CartPole-v0 on the left, Acrobot-v1 on the right), and each row displays spike activity for a layer (input, hidden 1, hidden 2). The first row shows spike c…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 21 canonical work pages

  1. [17]

    Porting deep spik- ing q-networks to neuromorphic chip loihi,

    M. Akl, Y . Sandamirskaya, F. Walter, and A. Knoll, “Porting deep spik- ing q-networks to neuromorphic chip loihi,” in International Conference on Neuromorphic Systems 2021 , 2021, pp. 1–7

  2. [1]

    Networks of spiking neurons: the third generation of neural network models,

    W. Maass, “Networks of spiking neurons: the third generation of neural network models,” Neural networks, vol. 10, no. 9, pp. 1659–1671, 1997

  3. [2]

    Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip,

    M. Yao, O. Richter, G. Zhao, N. Qiao, Y . Xing, D. Wang, T. Hu, W. Fang, T. Demirci, M. De Marchi et al. , “Spike-based dynamic computing with asynchronous sensing-computing neuromorphic chip,” Nature Communications, vol. 15, no. 1, p. 4464, 2024

  4. [3]

    Spiking neural networks and their applications: A review,

    K. Yamazaki, V .-K. V o-Ho, D. Bulsara, and N. Le, “Spiking neural networks and their applications: A review,”Brain sciences, vol. 12, no. 7, p. 863, 2022

  5. [4]

    Language modeling on a spinnaker2 neuromorphic chip,

    K. K. Nazeer, M. Sch ¨one, R. Mukherji, B. V ogginger, C. Mayr, D. Kappel, and A. Subramoney, “Language modeling on a spinnaker2 neuromorphic chip,” in 2024 IEEE 6th International Conference on AI Circuits and Systems (AICAS) . IEEE, 2024, pp. 492–496

  6. [5]

    Snn- bert: Training-efficient spiking neural networks for energy-efficient bert,

    Q. Su, S. Mei, X. Xing, M. Yao, J. Zhang, B. Xu, and G. Li, “Snn- bert: Training-efficient spiking neural networks for energy-efficient bert,” Neural Networks, vol. 180, p. 106630, 2024

  7. [6]

    Spinnaker2: A large-scale neuromorphic system for event-based and asynchronous machine learning,

    H. A. Gonzalez, J. Huang, F. Kelber, K. K. Nazeer, T. Langer, C. Liu, M. Lohrmann, A. Rostami, M. Sch¨one, B. V oggingeret al., “Spinnaker2: A large-scale neuromorphic system for event-based and asynchronous machine learning,” arXiv preprint arXiv:2401.04491 , 2024

  8. [7]

    Low-power neuromorphic hardware for signal processing appli- cations: A review of architectural and system-level design approaches,

    B. Rajendran, A. Sebastian, M. Schmuker, N. Srinivasa, and E. Elefthe- riou, “Low-power neuromorphic hardware for signal processing appli- cations: A review of architectural and system-level design approaches,” IEEE Signal Processing Magazine , vol. 36, no. 6, pp. 97–110, 2019

Show all 31 references
  1. [8]

    Loihi: A neuromorphic manycore processor with on-chip learning,

    M. Davies, N. Srinivasa, T.-H. Lin, G. Chinya, Y . Cao, S. H. Choday, G. Dimou, P. Joshi, N. Imam, S. Jain et al. , “Loihi: A neuromorphic manycore processor with on-chip learning,” Ieee Micro, vol. 38, no. 1, pp. 82–99, 2018

  2. [9]

    Brainqn: Enhancing the robustness of deep rein- forcement learning with spiking neural networks,

    S. Feng, J. Cao, Z. Ou, G. Chen, Y . Zhong, Z. Wang, J. Yan, J. Chen, B. Wang, C. Zou et al., “Brainqn: Enhancing the robustness of deep rein- forcement learning with spiking neural networks,” Advanced Intelligent Systems, vol. 6, no. 9, p. 2400075, 2024

  3. [10]

    Temporal dendritic heterogeneity incorporated with spiking neural networks for learning multi-timescale dynamics,

    H. Zheng, Z. Zheng, R. Hu, B. Xiao, Y . Wu, F. Yu, X. Liu, G. Li, and L. Deng, “Temporal dendritic heterogeneity incorporated with spiking neural networks for learning multi-timescale dynamics,” Nature Communications, vol. 15, no. 1, p. 277, 2024

  4. [11]

    Stdp-based spiking deep convolutional neural networks for object recognition,

    S. R. Kheradpisheh, M. Ganjtabesh, S. J. Thorpe, and T. Masquelier, “Stdp-based spiking deep convolutional neural networks for object recognition,” Neural Networks, vol. 99, pp. 56–67, 2018

  5. [12]

    Con- version of continuous-valued deep networks to efficient event-driven networks for image classification,

    B. Rueckauer, I.-A. Lungu, Y . Hu, M. Pfeiffer, and S.-C. Liu, “Con- version of continuous-valued deep networks to efficient event-driven networks for image classification,” Frontiers in neuroscience , vol. 11, p. 682, 2017

  6. [13]

    Exploring spiking neural networks for deep reinforcement learning in robotic tasks,

    L. Zanatta, F. Barchi, S. Manoni, S. Tolu, A. Bartolini, and A. Ac- quaviva, “Exploring spiking neural networks for deep reinforcement learning in robotic tasks,” Scientific Reports , vol. 14, no. 1, p. 30648, 2024

  7. [14]

    Directly-trained spiking neural networks for deep re- inforcement learning: Energy efficient implementation of event-based obstacle avoidance on a neuromorphic accelerator,

    L. Zanatta, A. Di Mauro, F. Barchi, A. Bartolini, L. Benini, and A. Acquaviva, “Directly-trained spiking neural networks for deep re- inforcement learning: Energy efficient implementation of event-based obstacle avoidance on a neuromorphic accelerator,” Neurocomputing, vol. 56...

  8. [15]

    Neurorobotic reinforcement learning for domains with parametrical uncertainty,

    C. Amaya and A. V on Arnim, “Neurorobotic reinforcement learning for domains with parametrical uncertainty,” Frontiers in Neurorobotics, vol. 17, p. 1239581, 2023

  9. [16]

    Deep reinforcement learning with population-coded spiking neural network for continuous control,

    G. Tang, N. Kumar, R. Yoo, and K. Michmizos, “Deep reinforcement learning with population-coded spiking neural network for continuous control,” in Conference on Robot Learning . PMLR, 2021, pp. 2016– 2029

  10. [18]

    Training spiking neural networks using lessons from deep learning,

    J. K. Eshraghian, M. Ward, E. Neftci, X. Wang, G. Lenz, G. Dwivedi, M. Bennamoun, D. S. Jeong, and W. D. Lu, “Training spiking neural networks using lessons from deep learning,” Proceedings of the IEEE , vol. 111, no. 9, pp. 1016–1054, 2023

  11. [19]

    Deep reinforcement learning with spiking q-learning,

    D. Chen, P. Peng, T. Huang, and Y . Tian, “Deep reinforcement learning with spiking q-learning,” arXiv preprint arXiv:2201.09754 , 2022

  12. [20]

    Openai gym,

    G. Brockman, V . Cheung, L. Pettersson, J. Schneider, J. Schul- man, J. Tang, and W. Zaremba, “Openai gym,” arXiv preprint arXiv:1606.01540, 2016

  13. [21]

    Reinforcement co-learning of deep and spiking neural networks for energy-efficient mapless nav- igation with neuromorphic hardware,

    G. Tang, N. Kumar, and K. P. Michmizos, “Reinforcement co-learning of deep and spiking neural networks for energy-efficient mapless nav- igation with neuromorphic hardware,” in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2020, pp. 6090–6097

  14. [22]

    A self-driving robot using deep convolutional neural networks on neuromorphic hardware,

    T. Hwu, J. Isbell, N. Oros, and J. Krichmar, “A self-driving robot using deep convolutional neural networks on neuromorphic hardware,” in2017 International Joint Conference on Neural Networks (IJCNN) . IEEE, 2017, pp. 635–641

  15. [23]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al. , “Human-level control through deep reinforcement learning,” nature, vol. 518, no. 7540, pp. 529–533, 2015

  16. [24]

    Neuronlike adaptive elements that can solve difficult learning control problems,

    A. G. Barto, R. S. Sutton, and C. W. Anderson, “Neuronlike adaptive elements that can solve difficult learning control problems,” IEEE transactions on systems, man, and cybernetics, no. 5, pp. 834–846, 1983

  17. [25]

    Generalization in reinforcement learning: Successful examples using sparse coarse coding,

    R. S. Sutton, “Generalization in reinforcement learning: Successful examples using sparse coarse coding,” Advances in neural information processing systems, vol. 8, 1995

  18. [26]

    Learning repre- sentations by back-propagating errors,

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning repre- sentations by back-propagating errors,” nature, vol. 323, no. 6088, pp. 533–536, 1986

  19. [27]

    Navigating local minima in quantized spiking neural networks,

    J. K. Eshraghian, C. Lammie, M. R. Azghadi, and W. D. Lu, “Navigating local minima in quantized spiking neural networks,” in 2022 IEEE 4th International Conference on Artificial Intelligence Circuits and Systems (AICAS). IEEE, 2022, pp. 352–355

  20. [28]

    Efficient deployment of spiking neural networks on spinnaker2 for dvs gesture recognition using neuromorphic intermediate representation,

    S. Arfa, B. V ogginger, C. Liu, J. Partzsch, M. Schone, and C. Mayr, “Efficient deployment of spiking neural networks on spinnaker2 for dvs gesture recognition using neuromorphic intermediate representation,” arXiv preprint arXiv:2504.06748 , 2025

  21. [29]

    Toward robust and scal- able deep spiking reinforcement learning,

    M. Akl, D. Ergene, F. Walter, and A. Knoll, “Toward robust and scal- able deep spiking reinforcement learning,” Frontiers in Neurorobotics, vol. 16, p. 1075647, 2023

  22. [30]

    py-spinnaker2,

    B. V ogginger, F. Kelber, M. Jobst, G. B ´ena, S. Arfa, Y . Yan, P. Gerhards, M. Weih, M. Akl, H. A. Gonzalez, and C. Mayr, “py-spinnaker2,” Aug. 2024. [Online]. Available: https://zenodo.org/ doi/10.5281/zenodo.10202109

  23. [31]

    Pynn: a common interface for neuronal network simulators,

    A. P. Davison, D. Br ¨uderle, J. M. Eppler, J. Kremkow, E. Muller, D. Pecevski, L. Perrinet, and P. Yger, “Pynn: a common interface for neuronal network simulators,” Frontiers in neuroinformatics, vol. 2, p. 388, 2009

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.