Controlling the latency in temporal coded spiking models for neuromorphic datasets
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Controlling the latency in temporal coded spiking models for neuromorphic datasets".
Jane: Here's the gist of the paper. Exploiting temporal information in datasets with high temporal dependencies (such as spoken keyword recognition) by spiking models is still an open problem The single most informative sentence about the paper's subject and finding is that the Delay Neural Network (DeNN) formulated in real time allows for a theoretical control of the latency, through a simple bounding of its synaptic parameters Spiking neural networks (SNNs) are promising alternatives to conventional artificial neural networks due to their energy efficiency through sparse, event-driven, and temporally coded computation Latency in spike-based architectures is intrinsically tied to the dynamics of the underlying neuron model, which limits classical models like Leaky integrate-and-fire (LIF) neurons because the relationship between network parameters and spike timing is difficult to tract While some work has addressed this by encouraging early spikes or deriving analytical expressions for spiking time, these approaches remain fundamentally empirical, lacking a formal, input-independent guarantee on worst-case latency The authors study Delay Neural Networks (DeNN) models whose spike-timing dynamics are constructed so that latency can be expressed and controlled analytically Model Formulation and Preprocessing The paper introduces the Delay Neural Network (DeNN) as a spiking neuron model whose spike time is analytically derivable, and where applicable invertible, as a function of its input and parameters The preprocessing involves extending the event2time algorithm by computing the average Inter-Spike Intervals (ISIs) to serve as a proxy for spiking time This process transforms spiking activity into a single spike time for some time window, allowing for the use of a TTFS-like algorithm on each image Is The postsynaptic neuron's spiking time is computed using equation (3), which incorporates the integrated synaptic activity and synaptic delays Analytical Latency Control The latency of a given neuron is formally defined as the time elapsed between its spiking time and the last input integrated, specifically for a layer l, it is defined as Lj
Is: := etlj
Is: − tgl−1q
Is: Theorem 3.2 provides bounds on this latency Lj = tej − teq between the spiking time of a post synaptic neuron j and the last integrated input spiking time teq, showing that Lminj ≤ Lj < nE exp hzq − min i zil−1i − e−c + ReLU olj + Mlj + min(s, ν) min hαh Lemma 3.3 simplifies this by bounding the synaptic parameters when assuming a symmetric distribution of the α neuronal parameter, showing that bounding the synaptic parameters is enough to bound the required bounds for d sij Experimental Validation and Trade-offs The model was empirically validated on a subset of the Google Speech Commands dataset, showing that promising accuracy can be achieved while respecting an analytically guaranteed latency The study shows that models with one and two hidden layers behave as expected, demonstrating higher accuracies with higher latencies The model with three hidden layers reached the best accuracy (92.56%) but had a network latency of 1663.4711 ms for R = 0.1 Furthermore, the global latency can be reduced by using a classical approach based on classification confidence thresholds The learning process allocated the long-term memory to some of the neurons only, as evidenced by analyzing rank-order encoding and importance metrics Long-Term Memory Allocation The long-term memory mechanism is implemented using a matrix α where the parameter α1 weights the memory of one timestep in the past, and so on until αν (note that the indices 1, …, ν are superscripts, not exponents) The shifting constant τ must be updated with equation (9), which incorporates terms related to synaptic delays and long-term memory parameters The importance of each term of the memory is computed using equation (14), which shows that for all models, there are few neurons which are important, and that they are more important for a larger number of timesteps This indicates that with fewer neurons, activity is distributed over more neurons The work establishes the first theoretical understanding of the accuracy-latency tradeoff on a dataset with strong temporal dependencies in SNNs trained in the temporal dimension with synaptic delays The global architecture of the model impacts latency, suggesting that adapting bounds to each layer gives the best results While deeper DeNN models might make studying the tradeoff analytically harder, the paper shows high accuracy with relatively low latency on a temporal dataset This result could be decreased to lower latency with further understanding of the model The first, to the best of our knowledge, that analyses this tradeoff on a dataset with strong temporal dependencies such as the Google Speech Commands, in the framework of SNNs trained in the temporal dimension with synaptic delays References Akopyan, F., J.
Tom: Sawada, A.
Jane: Cassidy, R.
Tom: Alvarez-Icaza, J.
Paper summary: Tom: So, this paper is about tackling a problem in spiking models where they handle time-dependent data well, like spoken keyword recognition. They're looking at how to control latency in these systems.
Jane: That sounds really interesting because we always talk about how fast things need to process real-time information, and the idea here is that controlling the latency analytically is still an open problem.
Lu: Exactly. The core thesis of this paper, "Controlling the latency in temporal coded spiking models for neuromorphic datasets," is that they introduce a Delay Neural Network model that lets you theoretically control the latency through bounding its synaptic parameters.
Meng: I wonder what that actually means for deployment at scale. If you can bound the latency analytically, does it mean we can predict exactly how long a response will take without running exhaustive tests?
Lalam: Lalam has looked at this literature and is confident in this direction. It seems like the paper proposes a Delay Neural Network, or DeNN, which lets you express and control spike timing analytically.
Tom: So they're not just tweaking parameters empirically; they are building a model where the spike time dynamics are constructed so that latency can be expressed and controlled analytically. That sounds like it could really simplify the design process for real-time systems.
Jane: But the question is, does that control hold up when you move from simple static data, like MNIST, to complex temporal datasets where things actually change over time?
Lu: That’s where their work gets tricky. The paper shows that preprocessing involves extending the event2time algorithm by computing average Inter-Spike Intervals as a proxy for spike time.
Meng: Proxy measurements are always worrying when you're building production models; they introduce an estimation error somewhere in the pipeline. How much does that error affect the final latency guarantee?
Lalam: The paper then goes on to define latency formally, calling it Lj Is, as the time between a neuron's spiking time and the last input integrated spike time for a specific layer, and they provide bounds on this.
Paper summary: Tom: Bounds are good because they give you a guarantee, but I always want to know what those bounds look like in practice. What’s the actual mathematical expression for that latency bound?
Jane: The paper provides Theorem three point two which shows that the latency Lj is bounded by an expression involving exponential functions and ReLU functions, which gives an upper limit on the spike time difference between layers, and Lemma three point three simplifies this by bounding synaptic parameters under a symmetric distribution of neuron parameters <ref:2610.19732#pg1>.
Lu: That lemma is key because it shows that simply bounding those synaptic parameters is enough to bound the required bounds for the spike times, which means you can control things at design time instead of after training.
Meng: If we have to bound synaptic parameters, that implies we need some kind of prior knowledge about how those parameters should be distributed for a given task. That sounds like a big assumption on our side.
Lalam: The paper does address the memory aspect, showing they implement long-term memory using a matrix alpha where alpha one weights the memory of one timestep in the past, and there are terms related to synaptic delays and these long-term memory parameters need updating with equation nine.
Tom: So they’re managing temporal dependencies by explicitly modeling that history through these weighted steps, which is important because spoken commands aren't just single snapshots.
Jane: And when you look at the experimental validation on the Google Speech Commands dataset, they found promising accuracy can be achieved while respecting an analytically guaranteed latency. That’s a concrete result to anchor the theory to.
Lu: They specifically showed that models with one and two hidden layers behave as expected, demonstrating higher accuracies with higher latencies, but the model with three hidden layers reached the best accuracy at ninety-two point five six percent but had a network latency of one thousand six hundred sixty-three point four seven one one milliseconds for an input ratio of zero point one.
Meng: Sixteen hundred thirty-four milliseconds is pretty slow for real-time voice command processing; we need something much tighter than that to be viable in a consumer device.
Paper summary: Lalam: They also noted that the global latency can be reduced by using a classical approach based on classification confidence thresholds, which suggests a hybrid approach might work well.
Tom: It sounds like the paper is saying that while deep models are accurate, they come with high latency, but there are ways to prune or simplify the architecture to get better speed without losing too much accuracy.
Jane: And for someone who only listens to the show, it means that for systems needing fast responses on time-sensitive data, like speech recognition, there is a formal way now to design the network so you know its maximum delay before you even start training.
Lu: This work establishes what they call the first theoretical understanding of the accuracy latency tradeoff on a dataset with strong temporal dependencies in spiking neural networks trained in the temporal dimension with synaptic delays.
Meng: So, for deployment, it means we can use these analytical bounds to set hard performance requirements upfront instead of discovering latency issues during post-training optimization.
Lalam: The implication is that adapting those bounds to each layer gives the best results, suggesting that the global architecture of the model significantly impacts latency.
Tom: So, for future work, they suggest that deeper DeNN models might make studying this tradeoff analytically harder, but they still show high accuracy with relatively low latency on temporal datasets.
Jane: It’s a step forward because it moves us from empirical testing to having a formal understanding of the relationship between model structure and response time.
Lu: They also point out that this is the first analysis of this tradeoff on a dataset with strong temporal dependencies like the Google Speech Commands in the framework of spiking neural networks trained in the temporal dimension with synaptic delays.
Meng: I'm still curious about how we can get those latency numbers down even further; if we could reduce that sixteen hundred thirty-four millisecond latency significantly, that would make a huge difference for low-power edge devices.
Lalam: The paper suggests that further understanding of the model is needed to decrease this latency even more, which points toward future refinement of the DeNN architecture itself.
Conclusion: Tom: So, we're wrapping up this look at the paper "Controlling the latency in temporal coded spiking models for neuromorphic datasets." What's actually going on here in simple terms?
Jane: Well, basically, the authors are showing how you can formally control how slow a spiking neural network runs when it deals with time-dependent data like speech recognition. They introduce this Delay Neural Network model that lets them predict and limit the latency mathematically by just adjusting some parameters.
Lu: I think what they're really highlighting is that for these networks, latency isn't just an empirical mess; they can define a strict mathematical upper bound on it based on how you set up the synaptic delays.
Meng: That sounds promising for deployment because it moves us away from just hoping our models are fast enough to a situation where we know exactly what the worst-case delay is before we even launch.
Lalam: I'm quite sure of this direction, as they establish the first theoretical framework connecting accuracy and latency on these time-dependent datasets using synaptic delays in spiking neural networks.
Tom: So they’re saying that by building the network right this way, you get a guarantee on performance that you can control with just some math on the parameters.
Jane: Exactly. It's about taking something that used to be purely empirical—just testing and seeing how slow it was—and making it something where we have a formal understanding of the time delay involved.
Lu: The authors prove that bounding those synaptic parameters is enough to guarantee bounds on the spike times between layers, which means you control the system at design time instead of after training.
Meng: If we can bound these parameters, it suggests that we can design networks for specific latency targets upfront, which is a huge win for production systems where real-time response matters.
Lalam: And they even showed that this method works well when validated on complex data like the Google Speech Commands dataset, achieving good accuracy while keeping latency within those guaranteed limits.
Tom: It sounds like the implication is that we can start designing these temporal models with performance in mind, rather than treating latency as an afterthought during optimization.
Jane: That's right. This work establishes a foundational understanding of this trade-off for spiking networks dealing with long-term temporal dependencies, and it opens up a new way to approach system design in AI.
Alban Gattepaille, Alexandre Muzy
Universite Côte d’Azur - CNRS - I3S · International Laboratory on Learning Systems (ILLS)
cs.NE
Submitted: 2026-10-14
Updated: 2026-11-09
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 81/100
The gist: Exploiting temporal information in datasets with high temporal dependencies (such as spoken keyword recognition) by spiking models is still an open problem The single most informative sentence about…
Key concepts
- Delay Neural Network (DeNN)
- A specific spiking neuron model where the time it spikes is mathematically derivable from its inputs and internal parameters. This formulation allows researchers to analytically control and express the resulting spike timing dynamics, which is crucial for managing latency in temporal tasks.
- Analytical Latency Control
- The ability to formally define and bound the latency (the time between a neuron's spike and the last integrated input) using mathematical theorems. This control is achieved by bounding the synaptic parameters within the DeNN model, providing a guarantee on worst-case latency regardless of input.
- Temporal Coding
- A method where information is encoded not just in the amplitude of signals, but in when spikes occur over time. In this context, it means using spike timing dynamics to represent sequential data like speech commands efficiently within the spiking neural network architecture.
Terminology
Summary
Exploiting temporal information in datasets with high temporal dependencies (such as spoken keyword recognition) by spiking models is still an open problem The single most informative sentence about the paper's subject and finding is that the Delay Neural Network (DeNN) formulated in real time allows for a theoretical control of the latency, through a simple bounding of its synaptic parameters
Motivation and Problem Statement
Spiking neural networks (SNNs) are promising alternatives to conventional artificial neural networks due to their energy efficiency through sparse, event-driven, and temporally coded computation Latency in spike-based architectures is intrinsically tied to the dynamics of the underlying neuron model, which limits classical models like Leaky integrate-and-fire (LIF) neurons because the relationship between network parameters and spike timing is difficult to tract While some work has addressed this by encouraging early spikes or deriving analytical expressions for spiking time, these approaches remain fundamentally empirical, lacking a formal, input-independent guarantee on worst-case latency The authors study Delay Neural Networks (DeNN) models whose spike-timing dynamics are constructed so that latency can be expressed and controlled analytically
Model Formulation and Preprocessing
The paper introduces the Delay Neural Network (DeNN) as a spiking neuron model whose spike time is analytically derivable, and where applicable invertible, as a function of its input and parameters The preprocessing involves extending the event2time algorithm by computing the average Inter-Spike Intervals (ISIs) to serve as a proxy for spiking time This process transforms spiking activity into a single spike time for some time window, allowing for the use of a TTFS-like algorithm on each image Is The postsynaptic neuron's spiking time is computed using equation (3), which incorporates the integrated synaptic activity and synaptic delays
Analytical Latency Control
The latency of a given neuron is formally defined as the time elapsed between its spiking time and the last input integrated, specifically for a layer l, it is defined as Lj [Is]:= etlj[Is] − tgl−1q[Is] Theorem 3.2 provides bounds on this latency Lj = tej − teq between the spiking time of a post synaptic neuron j and the last integrated input spiking time teq, showing that Lminj ≤ Lj < nE exp hzq − min i zil−1i − e−c + ReLU olj + Mlj + min(s, ν) min hαh Lemma 3. p. 53 simplifies this by bounding the synaptic parameters when assuming a symmetric distribution of the α neuronal parameter, showing that bounding the synaptic parameters is enough to bound the required bounds for d sij
Experimental Validation and Trade-offs
The model was empirically validated on a subset of the Google Speech Commands dataset, showing that promising accuracy can be achieved while respecting an analytically guaranteed latency The study shows that models with one and two hidden layers behave as expected, demonstrating higher accuracies with higher latencies The model with three hidden layers reached the best accuracy (92.56%) but had a network latency of 1663. p. 64711 ms for R = 0. p. 61 Furthermore, the global latency can be reduced by using a classical approach based on classification confidence thresholds The learning process allocated the long-term memory to some of the neurons only, as evidenced by analyzing rank-order encoding and importance metrics
Long-Term Memory Allocation
The long-term memory mechanism is implemented using a matrix α where the parameter α1 weights the memory of one timestep in the past, and so on until αν (note that the indices 1, …, ν are superscripts, not exponents) The shifting constant τ must be updated with equation (9), which incorporates terms related to synaptic delays and long-term memory parameters The importance of each term of the memory is computed using equation (14), which shows that for all models, there are few neurons which are important, and that they are more important for a larger number of timesteps This indicates that with fewer neurons, activity is distributed over more neurons
Conclusion and Future Directions
The work establishes the first theoretical understanding of the accuracy-latency tradeoff on a dataset with strong temporal dependencies in SNNs trained in the temporal dimension with synaptic delays The global architecture of the model impacts latency, suggesting that adapting bounds to each layer gives the best results While deeper DeNN models might make studying the tradeoff analytically harder, the paper shows high accuracy with relatively low latency on a temporal dataset This result could be decreased to lower latency with further understanding of the model The first, to the best of our knowledge, that analyses this tradeoff on a dataset with strong temporal dependencies such as the Google Speech Commands, in the framework of SNNs trained in the temporal dimension with synaptic delays
References
Akopyan, F., J. Sawada, A. Cassidy, R. p. 9 Alvarez-Icaza, J. p. 9 Arthur, P. Merolla, N. p. 9 Imam, Y. Nakamura, P. p. 9 Datta, G.-J. Nam, B. Taba, M. Beakes, B.- Brezzo, J.-B Kuangge r., R., Manohar (2015). p. 9 Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip
Ayasi, B., A. M. Garc´ıa-Vico, C. p. 9 J. Carmona, and M. p. 9 Saleh (2024). p. 9 Advancing computational frontiers: Spiking neural networks in high-energy efficiency computing across diverse domains
Aydin, A., M. Gehrig, D. Gehrig, and D. p. 9 Scaramuzza (2024). p. 9 A hybrid ann-snn architecture for low-power and lowlatency visual perception
Bacho, F. and D. Chu (2023). p. 1 Exploring trade-offs in spiking neural networks
Bittar, A. and P. Garner (2022). p. 0 A surrogate gradient spiking baseline for speech command recognition
Bu, T., J. Ding, Z. Yu, and T. Huang (2022). p. 9 Optimized potential initialization for low-latency spiking neural networks
Bu, T., W. Fang, J. Ding, P. Dai, Z. Yu, and T. Huang (2023). p. 9 Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks
Che, K., W. Fang, P. Xue, Y. Huang, Z. Ma, and Y. Tian (2026). p. 9 Parallel training time-to-first-spike spiking neural networks
Chowdhury, S. S., N. Rathi, and K. p. 9 p. 9 Roy (2021). p. 9 One timestep is all you need: Training spiking neural networks with ultra low latency
Chowdhury, S. S., N. Rathi, and K. Roy (2022). p. 9 Towards ultra low latency spiking neural networks for vision and sequential tasks using temporal pruning
Chu, K. T. N., B. Amornpaisannon, Y. p. 9 Tavva, V. P. K Miriyala, J Wu, M Zhang, H Li, T E Carlson et al (2020). p. 9 You only spike once: Improving energy-efficient neuromorphic inference to ann-level accuracy
Comsa, I. M., K. Potempa, L Versari, T Fischbacher, A Gesmundo, and J Alakuijala (2020). p. 9 Temporal coding in spiking neural networks with alpha synaptic function
Datta, G. and P A Beerel (2022). p. 1 Can deep neural networks be converted to ultra low-latency spiking neural networks? p. 9
Datta, G., S Kundu, and P A Beerel (2021). p. 9 Training energy-efficient deep spiking neural networks with single-spike hybrid input encoding
Davies, M., N. Srinivasa, T. p. 9-H Lin, G. Chinya, Y Cao, S. H Choday, G Dimou, P Joshi, N Imam et al (2018). p. 0 Loihi: A neuromorphic manycore processor with on-chip learning
Garg, I., S S Chowdhury, and K Roy (2021). p. 9 Dct-snn: Using dct to distribute spatial information over time for lowlatency spiking neural networks
Gattepaille, A. and A Muzy (2025). p. 1 Delay neural networks for exploiting temporal information in event-based datasets, to appear in Nature Communications Engineering
Ge, B., L Du, Y Guo, J Song, X Yang, and H Liu (2026). p. 1 Training low-latency spiking neural networks through discrete knowledge distillation over output and multiple intermediate features
Goltz, J., L Kriener, A Baumbach, S Billaudelle et al. (2021). Fast and energy-efficient neuromorphic deep learning with first-spike times
Han, B., G. Srinivasan, and K Roy (2020). p. 9 Rmp-snn: Residual membrane potential neuron for enabling deeper highaccuracy and low-latency spiking neural network
Haowen, F., S Amar, Z Ziyi, and Q Qinru (2020, 07). p. 0 Exploiting neuron and synapse filter dynamics in spatial temporal learning of deep spiking neural network
Ho, N.-D. and I.-J Chang (2023). p. 1 Mice: An ann-to-snn conversion technique to enable high accuracy and low latency
Hwang, S., J Chang, M. p. 1-H Oh, K K Min, T Jang, K Park et al (2021). p. 0 Low-latency spiking neural networks using pre-charged membrane potential and delayed evaluation
Im, J., J Kim, H. p. 1-N Yoo, J W Baek et al. (2022). On-chip trainable spiking neural networks using time-to-first-spike encoding
Jiang, Y., F Chen, Y Li, Y Liu et al. p. 1 (2026). Adaptive fission: Post-training encoding for low-latency spike neural networks
Kheradpisheh, S. R. and T Masquelier (2020). p. 0 Temporal backpropagation for spiking neural networks with one spike per neuron
Kheradpisheh, S. R., M Mirsadeghi, and T Masquelier (2022). p. 0 Bs4nn: Binarized spiking neural networks with temporal coding and learning
Kim, Y., A Kahana, R Yin, Y Li et al. (2024). Rethinking skip connections in spiking neural networks with time-to-first-spike coding
Kingma, D. P. (2014). Adam: A method for stochastic optimization
Krizhevsky, A., G Hinton, et al. (2009). Learning multiple layers of features from tiny images
Lew, D., K Lee, and J Park (2022). p. 0 A time-to-first-spike coding and conversion aware training for energy-efficient deep spiking neural network processor design
Li, C., Z Shang, L Shi, W Gao, and S Zhang (2023). p. 0 Ic-snn: Optimal ann2snn conversion at low latency
Loshchilov, I. and F Hutter (2016). p. 6 Sgdr: Stochastic gradient descent with warm restarts
Lu, Y., J Ding, and Z Yu (2026). p. 0 A latency coding framework for deep spiking neural networks with ultra-low latency
Maass, W. (1997). Networks of spiking neurons: The third generation of neural network models
Mirsadeghi, M., M Mirbagheri, and S R Kheradpisheh (2026). p. 0 A multiplication-free spike-time learning algorithm and its efficient fpga implementation for on-chip snn training
Mostafa, H. (2017). Supervised learning based on temporal coding in spiking neural networks
Nomura, O., Y Sakemi, T Hosomi, and T Morie (2022). p. 1 Robustness of spiking neural networks based on time-tofirst-spike encoding against adversarial attacks
Oh, S., D Kwon, G Yeom, W.-M Kang et al. (2022). Neuron circuits for low-power spiking neural networks using time-tofirst-spike encoding
Oh, S., S Lee, S Y Woo, D Kwon et al. (2021). Spiking neural networks with time-to-first-spike coding using tft-type synaptic device model
Orchard, G., G Cohen, A Jayawant, and N Thakor (2015). p. 0 Converting static image datasets to spiking neuromorphic datasets using saccades
Painkras, E., L A Plana, J Garside et al. (2013). Spinnaker: A 1-w 18-core system-on-chip for massively-parallel neural network simulation
Park, S., S Kim, B Na, and S Yoon (2020). p. 0 T2fsnn: Deep spiking neural network with time-tofirst-spike coding
Park, S. and S Yoon (2021). p. 0 Training energy-efficient deep spiking neural networks with time-to-first-spike coding
Pes, L., R Luiken, F Corradi, and C Frenkel (2024). p. 1 Active dendrites enable efficient continual learning in time-tofirst-spike neural networks
Qu, J., Z Gao, T Zhang, Y Lu et al. p. 0 (2024). Spiking neural network for ultralow-latency and highaccurate object detection
Rueckauer, B. and S.-C Liu (2018). p. 1 Conversion of analog to spiking neural networks using sparse temporal coding
Sakemi, Y., K Yamamoto, T Hosomi, and K Aihara (2023). p. 1 Sparse-firing regularization methods for spiking neural networks with time-to-first-spike coding
Singh, S., A Sarma, S Lu et al. (2021). Gesture-snn: Co-optimizing accuracy, latency and energy of snns for neuromorphic vision sensors
Stewart, K., T Shea, N Pacik-Nelson et al. (2023). Speech2spikes: Efficient audio encoding pipeline for real-time neuromorphic systems
Stockl, C. and W Maass (2020). p. 0 Recognizing images with at most one spike per neuron
Thorpe, S., A Delorme, and R Van Rullen (2001). p. 7 Spike-based strategies for rapid processing
Wang, B., J Cao, J Chen, S Feng, and Y Wang (2023). p. 1 A new ann-snn conversion method with high accuracy, low latency and good robustness
Wang, Z., Y Zhang, S Lian, X Cui et al. p. 0 (2025). Toward high-accuracy and low-latency spiking neural networks with two-stage optimization
Warden, P. (2018). Speech commands: A dataset for limited-vocabulary speech recognition
Wei, W., M Zhang, H Qu, A Belatreche et al. p. 0 (2023). Temporal-coded spiking neural networks with dynamic firing threshold: Learning with event-driven backpropagation
Xia, L., X Wang, R Wang, Y Yin et al. p. 1 (2026). Enhancing ann-snn conversion: Addressing low latency and negative thresholds with modified activation and bptt fine-tuning
Yamazaki, K., V Vo-Ho, D Bulsara, and N Le (2022). p. 0 Spiking neural networks and their applications: A review
Yan, Z., K Tang, J Zhou, and W. p. 1-F Wong (2025). p. 1 Low latency conversion of artificial neural network models to rate-encoded spiking neural networks
Yang, Q., M Zhang, J Wu, K C Tan, and H Li (2023). p. 0 Lc-ttfs: Toward lossless network conversion for spiking neural networks with ttfs coding
Yang, R., Z Li, J Yao et al. (2024). Compact artificial neurons with time-to-first-spike coding for fast and energy-efficient federated neuromorphic computing
Yang, Y., Q Xie, Z Xuan, S Chen, and Y Kang (2025a). p. 1 A neuromorphic hardware architecture based on ttfs coding with temporal quantization for spiking neural networks
Yang, Y., Q Xie, Z Xuan, S Chen, and Y Kang (2025b). p. 1 A neuromorphic hardware architecture based on ttfs coding with temporal quantization for spiking neural networks
Zhang, L., S Zhou, T Zhi, Z Du, and Y Chen (2019). p. 1 Tdsnn: From deep neural networks to deep spike neural networks with temporal-coding
Zhang, M., J Wang, J Wu, A Belatreche et al. p. 1 (2021). Rectified linear postsynaptic potential function for backpropagation in deep spiking neural networks
Zhang, W. and P Li (2020). p. 0 Temporal spike sequence learning via backpropagation for deep spiking neural networks
Zhang, Y., Y Sun, H Li, Y Deng, and W Yao (2026). p. 0 Pseudo-spiking neurons: A noise-based training framework for heterogeneous-latency spiking neural networks
Zhang, Z. and Q Liu (2021). p. 1 Spike-event-driven deep spiking neural network with temporal encoding
Zhou, S., X Li, Y chen, S Chandrasekaran, and A Sanyal (2021). p. 0 Temporal-coded deep spiking neural network with easy training and robust performance
Zhu, Y., W Fang, X Xie, T Huang, and Z Yu (2023). p. 9
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Analytical latency control in Delay Neural Networks (DeNN) allows designers to
enforce a target latency budget by construction rather than through indirect training penalties.
This enables systems to meet hard real-time constraints or certification requirements by specifying or verifying a provable latency bound at design time. -
The model can be used to
predict the exact, or provably bounded, latency of any neuron or decision layer prior to deployment,
shifting fromempirical latency measurement to analytical latency guarantee.
This provides a formal input-independent guarantee on worst-case response time. -
By varying the parameter R after training, systems can navigate the tradeoff between accuracy and latency by applying bounds derived from Lemma 3.3, which allows for
bounding the synaptic parameters
based on aO(1) complexity
check. -
The system can dynamically select layer-specific bounds to optimize performance, as the paper shows that
adapting the bounds to each layer gives the best results,
mitigating issues where global architecture impacts latency beyond single-layer theorems. -
The model can allocate long-term memory resources optimally, as analysis shows that
the learning process allocated the long-term memory to some of the neurons only,
indicating a mechanism for sparse and efficient resource allocation during training on temporal data.
Sources
- Optimal ANN-SNN Conversion for High-accuracy and Ultra-low-latency Spiking Neural Networks
- One Timestep is All You Need: Training Spiking Neural Networks with Ultra Low Latency
- You Only Spike Once: Improving Energy-Efficient Neuromorphic Inference to ANN-Level Accuracy
- Adam: A Method for Stochastic Optimization
- SGDR: Stochastic Gradient Descent with Warm Restarts
- Latency Coding for Efficient and Low-Latency Deep Spiking Neural Networks
- A Multiplication-Free Spike-Time Learning Algorithm and its Efficient FPGA Implementation for On-Chip SNN Training
- Training Energy-Efficient Deep Spiking Neural Networks with Time-to-First-Spike Coding
- Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition
Related papers
- Evolutionary Ensemble of Agents
- EvoDR: Evolving Dispatching Rules via Large Language Model for Dynamic Flexible Assembly Flow Shop Scheduling
- RECAP: Local Hebbian Prototype Learning as a Self-Organizing Readout for Reservoir Dynamics
- Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets
- Design and Analysis of Temporal Representations in Neural Networks
- Evolutionary Computation: A Paradigm of Natural Generative AI and Creative Synthesis