AdapSNE: Adaptive Sampling & Spiking Neural Nets

• AdapSNE is a multidimensional concept encompassing both DNN-free dataset sampling for edge training and adaptive spiking neural network techniques for sparse, asynchronous processing.
• It combines t-SNE–like dimensionality reduction with Fireworks Algorithm search and entropy-guided perplexity control to optimize exemplar selection and improve training efficiency.
• AdapSNE and its variants (including d-SNE) address distinct domain challenges ranging from edge learning and LLM fine-tuning to neuromorphic computational models and supervised domain adaptation.

Searching arXiv for papers and related usages of “AdapSNE” to ground the article. AdapSNE is a term with multiple uses in the arXiv literature. In its 2025 usage, it denotes “Adaptive Fireworks-Optimized and Entropy-Guided Dataset Sampling for Edge DNN Training,” a DNN-free dataset sampling framework for edge DNN and LLM training that combines t-SNE–like dimensionality reduction, Fireworks Algorithm search, entropy-guided target-perplexity adaptation, grid sampling, and a custom accelerator (Zhao et al., 19 Aug 2025). In earlier spiking-network literature, closely related labels such as “AdapSNE,” “ASNN,” and “AdSNN” refer to adapting or adaptive spiking neural networks built from adaptive spiking neurons that replace ANN units and implement adaptive spike-time coding (Zambrano et al., 2016, Zambrano et al., 2017). A further related SNE-derived line is “d-SNE,” a supervised domain-adaptation method based on stochastic neighborhood embedding and a modified Hausdorff distance (Xu et al., 2019).

1. Terminological scope

The supplied literature uses the label “AdapSNE” for distinct constructs rather than for a single standardized method. The 2025 paper uses it as the name of an edge-oriented dataset sampling framework. The 2016 and 2017 spiking papers use closely related naming for adapting or adaptive spiking neural networks. The 2019 domain-adaptation paper is explicitly titled $d$-SNE rather than AdapSNE, but it belongs to the same broader SNE-derived naming family and is relevant when the term is interpreted expansively.

Usage in the supplied literature	Core object	Principal mechanism
AdapSNE	DNN-free dataset sampling for edge DNN/LLM training	FWA, entropy-guided perplexity tuning, grid sampling, custom accelerator
ASNN / AdSNN / AdapSNE	Adapting or adaptive spiking neural networks	Adaptive spike-time coding, adaptive thresholds, ANN-to-SNN conversion
$d$-SNE	Supervised domain adaptation	SNE-style neighborhood probabilities and modified Hausdorff distance

This terminological overlap matters because the three lines target different problem settings. The 2025 framework addresses exemplar selection under edge constraints; the 2016–2017 work addresses sparse asynchronous neural computation with adaptive spiking neurons; the 2019 method addresses few-shot and semi-supervised domain adaptation.

2. AdapSNE as DNN-free dataset sampling for edge training

In “AdapSNE: Adaptive Fireworks-Optimized and Entropy-Guided Dataset Sampling for Edge DNN Training” (Zhao et al., 19 Aug 2025), AdapSNE is defined as a DNN-free dataset sampling framework designed to make training deep neural networks—including large models and LLMs—feasible directly on edge devices. The motivation is explicit: edge devices need local training or fine-tuning for domain adaptation, privacy preservation, and heterogeneous data, but conventional full-dataset training is prohibitive in compute, memory, and energy. The paper states that training Gemini required on the order of $10^3$ MWh, far beyond typical edge batteries of approximately $10^{-1}$ MWh.

AdapSNE is positioned as an extension of Near-Memory Sampling (NMS), the prior DNN-free state of the art. NMS first reduces dimensionality, then performs exemplar sampling in the reduced space. Its attraction is that it avoids the architectural bias of DNN-based selection methods, which use learned features tied to a particular backbone. The 2025 paper identifies two limitations in NMS. First, the search method is mismatched to the non-monotonic property of the perplexity error function, which leads to outliers in the reduced representation. Second, the target perplexity is selected empirically, introducing arbitrariness and causing uneven sampling. These two effects are described as producing representative bias of exemplars and degrading training accuracy.

The framework therefore targets representative exemplar selection without relying on a DNN to guide sampling. In the paper’s formulation, the objective is to compress large training datasets into small representative exemplar sets that preserve training accuracy and can be computed efficiently on edge hardware. This is the central sense of AdapSNE in the 2025 usage.

3. Algorithmic structure: FWA search, entropy-guided perplexity control, and grid sampling

AdapSNE retains a t-SNE–like probabilistic embedding, but modifies both the search over local bandwidths and the selection of target perplexity. In the high-dimensional space, conditional similarities are defined by Gaussian kernels,

$$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$

with local entropy

$H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$

perplexity

$R_i(\sigma_i) = 2^{H(P_i)},$

and perplexity error

$f(\sigma_i) = \big| R_i(\sigma_i) - \Pi_t \big|.$

The low-dimensional similarities remain

$$q_{ij} = \frac{(1 + \|y_i - y_j\|^2)^{-1}} {\sum_{k \ne l} (1 + \|y_k - y_l\|^2)^{-1}},$$

and the embedding is optimized through

$$C = KL(P\|Q) = \sum_i \sum_j p_{ij} \log \frac{p_{ij}}{q_{ij}}.$$

The paper’s key analytical claim is that the perplexity function $d$0 is non-monotonic. It derives

$d$1

and shows that the terms inside $d$2 have mixed signs, so $d$3 can change sign depending on $d$4. This is used to justify replacing NMS’s differential-evolution search with the Fireworks Algorithm (FWA), which the paper characterizes as suitable for large, high-dimensional, non-monotonic objectives.

The FWA component optimizes

$d$5

Its loop is described in terms of explosion, mutation, and selection. Sparks are generated by

$d$6

mutated fireworks by

$d$7

with clipping to $d$8, and the next population is formed from the best candidates under the fitness $d$9. In the AdapSNE pipeline, this optimization is invoked each time the target perplexity $10^3$0 is updated.

The second major modification is entropy-guided optimization of $10^3$1. After embedding the data into a $10^3$2 grid, cell frequencies define

$10^3$3

The threshold is then set as

$10^3$4

with $10^3$5. If the current entropy is below threshold, $10^3$6 is updated, first by a step

$10^3$7

and then by a Newton-style rule using

$10^3$8

followed by

$10^3$9

Only after this closed loop produces a sufficiently uniform embedding does AdapSNE perform grid sampling to select exemplars.

A plausible implication is that the framework treats manifold uniformity as a control target rather than as an incidental property of dimensionality reduction. That interpretation follows from the paper’s explicit coupling of entropy measurement, perplexity update, and exemplar extraction.

4. Empirical behavior and accelerator design

The 2025 AdapSNE paper evaluates the method on small-scale image datasets—CIFAR-10, CINIC-10, and SVHN—large-scale image datasets—CIFAR-100, ImageNet-100, and ImageNet-1K—and an LLM setting that fine-tunes LLaMA-7B with LoRA on Alpaca and evaluates on InstructEval and MMLU (Zhao et al., 19 Aug 2025). The models include ResNet-18, ResNet-50, ShuffleNetV2, MobileNetV2, ViT, and LLaMA-7B. Keeping ratios for compressed datasets are 10%, 20%, and 30%.

Across the reported image benchmarks, AdapSNE consistently improves over DQ, DQAS, NeSSA, and NMS. On CIFAR-10, the average Top-1 accuracy gain is 3.9% versus DQ, 3.8% versus DQAS, 2.8% versus NeSSA, and 1.0% versus NMS, with maximum gains of 10.2%, 9.3%, 4.9%, and 1.9%, respectively. On ImageNet-1K, the average gains are 14.4% versus DQ, 12.2% versus DQAS, 7.3% versus NeSSA, and 2.5% versus NMS, with maximum gains of 25.9%, 25.5%, 15.6%, and 3.9%. In the LLM setting on MMLU, AdapSNE improves over DQAS by 3.4%, 4.1%, and 2.9% at KR $10^{-1}$0, $10^{-1}$1, and $10^{-1}$2, and over NMS by 2.2%, 2.8%, and 2.1% at the same keeping ratios. The paper also reports an ablation against vanilla t-SNE: CIFAR-10 rises from 86.7% to 91.6%, CIFAR-100 from 58.9% to 65.3%, and ImageNet-1K from 56.0% to 63.3%.

The hardware contribution is a custom accelerator with three modules: an FWA-optimized t-SNE module, an entropy calculation module, and a grid sampling module. The FWA accelerator comprises a RAND unit, SPK RAM, POP RAM, MUT RAM, NPOP RAM, arg min units, and an Evaluate module. The entropy accelerator uses four-way parallel processing, reconfigurable grid parameters, and time-multiplexed shared RAM for grid counts. According to the paper’s synthesis results in TSMC 28nm, FWA search and entropy add 2.7% of total power and 6.1% of total area over the t-SNE core. The stated purpose of this design is to make the iterative FWA and entropy computations inexpensive enough for edge deployment.

5. Earlier usage: adapting and adaptive spiking neural networks

In the 2016 and 2017 spiking-network papers, AdapSNE refers to a different line of work centered on adapting or adaptive spiking neural networks rather than on dataset sampling (Zambrano et al., 2016, Zambrano et al., 2017). The 2016 paper describes an Adapting Spiking Neural Network (ASNN) in which every ReLU neuron of a standard ANN is replaced by an adaptive spiking neuron implementing adaptive Asynchronous Pulsed Sigma-Delta modulation. The 2017 paper develops Adaptive SNNs (AdSNNs) built from Adaptive Spiking Neurons (ASNs), formalizes an effective transfer function, and shows direct ANN-to-SNN conversion for deep benchmarks.

The neuron model is SRM-style. In the 2017 formulation, activation is

$10^{-1}$3

the refractory response is

$10^{-1}$4

the adaptive threshold is

$10^{-1}$5

and the effective membrane potential is

$10^{-1}$6

A spike is emitted when

$10^{-1}$7

The 2017 paper’s central bridge to ANN training is an analytically derived effective transfer function $10^{-1}$8, used as the activation of Artificial Analog Neurons during training and then matched by actual ASNs after conversion. The paper states that this transfer is rectified, behaves like a half-sigmoid above threshold, and can be normalized so that $10^{-1}$9, with spike height $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$0 serving as a scaling factor on trained weights.

The conversion pipeline is direct. An ANN is trained with the effective transfer function $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$1, Batch Normalization is used during training, BN parameters are absorbed into weights, and AAN units are replaced by ASNs with matching $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$2’s, $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$3, $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$4, and $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$5. In the 2017 experiments, typical biological-like constants are $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$6, $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$7, $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$8, and $$p_{j|i} = \frac{\exp\left(-\frac{\|x_i - x_j\|^2}{2\sigma_i^2}\right)} {\sum_{k \ne i} \exp\left(-\frac{\|x_i - x_k\|^2}{2\sigma_i^2}\right)},$$9. Pooling is merged into the next ASN layer, biases are added to postsynaptic activation $H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$0, and the last layer uses a slower membrane filter with $H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$1.

The empirical claims in the spiking literature concern sparse, asynchronous computation with ANN-level accuracy. The 2016 ASNN paper reports exact matching to the corresponding ANN on IRIS, SONAR, and MNIST, with average firing rates such as 36 Hz on IRIS, 59.7 Hz on SONAR, 14.6 Hz on MNIST FF-ASNN, and 8.6 Hz on MNIST C-ASNN, versus approximately 1000 Hz for the compared Poisson SNN on MNIST. The 2017 AdSNN paper extends this to MNIST, CIFAR-10, CIFAR-100, and ILSVRC-2012. It reports that baseline AdSNNs match ANN performance with firing rates between 24 Hz and 68 Hz and matching time around 300 ms, while state-of-the-art accuracies for most benchmarks are reached after about 200 ms. It also introduces an arousal mechanism that switches from low-precision to high-precision parameters on uncertain samples; the reported effect is roughly a halving of average firing rate with increased latency, and on MNIST specifically the firing rate falls from 67 Hz to 12 Hz while maintaining 99.56% accuracy.

These spiking papers therefore use AdapSNE in a neuromorphic and coding-theoretic sense. Their central objects are adaptive thresholds, spike-time coding, and conversion of analog ANN computation into sparse spiking dynamics, not manifold sampling.

6. Related SNE-based adaptation and distinctions from $H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$2-SNE

A related but separate SNE-derived method is “$H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$3-SNE: Domain Adaptation using Stochastic Neighborhood Embedding” (Xu et al., 2019). It addresses supervised few-shot and semi-supervised domain adaptation, not exemplar sampling and not spiking computation. The problem setting consists of a large labeled source domain $H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$4 and a small labeled target domain $H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$5, with identical label spaces but different data distributions. The method learns a shared latent feature space in which source and target samples of the same class are neighbors and samples of different classes are separated by a margin.

The probabilistic core is defined by

$H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$6

with the class-consistent probability

$H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$7

To avoid the scaling issues of the sum-of-exponentials likelihood, the paper replaces it with a modified Hausdorff-style loss,

$H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$8

combined with cross-entropy losses on source and target, and extended in the semi-supervised case with a Mean Teacher consistency term.

The reported results are specific to domain adaptation. On MNIST $H(P_i) = - \sum_{j \ne i} p_{j|i} \log_2 p_{j|i},$9 USPS with one target sample per class, $R_i(\sigma_i) = 2^{H(P_i)},$0-SNE achieves 73.01% versus 65.40% for CCSA and FADA; with seven target samples per class, it reaches 96.13%. On Office-31 with ResNet-101 and three target samples per class, the average accuracy across six transfers is 90.01. On VisDA-C, semi-supervised $R_i(\sigma_i) = 2^{H(P_i)},$1-SNE reaches 86.15%, exceeding 85.40% for SE and 77.10% for G2A.

The distinction among these SNE-derived usages is substantive. The 2025 AdapSNE is a closed-loop sampling system for edge training; the 2016–2017 AdapSNE/ASNN/AdSNN line is a spiking-network conversion framework for asynchronous neural computation; $R_i(\sigma_i) = 2^{H(P_i)},$2-SNE is a class-aware domain-adaptation loss built on stochastic neighborhood embedding. Any technical reading of the term therefore requires disambiguation by paper, year, and problem setting.