{"id":"93d295f3-dede-461d-afcb-4ae1d8224bbc","arxiv_id":"2508.02313","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A t-SNE-based sampling algorithm with differential evolution and near-memory hardware is claimed to speed up edge DNN training and reduce memory energy.","lead":"This paper proposes DE-SNE, a way to pick a small representative set of training images using t-SNE with an evolutionary search, plus a near-memory chip layout that runs the selection next to DRAM. The authors claim better accuracy than existing coreset methods and much lower DDR energy for edge DNN training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Perplexity is monotonic in sigma, so the claimed need for DE-SNE is not established; Table II's accuracy gap is unexplained.","rationale":"The paper's two contributions are (1) DE-SNE, a perplexity-search variant of t-SNE claimed to fix binary-search failures, and (2) a near-memory accelerator for it. The correctness of contribution (1) rests entirely on the claim that perplexity is non-monotonic in sigma, stated in Section IV-B and Fig. 6 without proof. This claim is false for the standard perplexity definition in Eq. (3): entropy is monotonically increasing in sigma, so binary search is well-posed. That makes the DE-SNE algorithm's motivation unsound. The ablation in Table II then lacks a mechanistic explanation; it may reflect a broken t-SNE baseline rather than a genuine limitation of binary search. The paper also reports very large perplexity-error reductions (Section X, Fig. 15) without giving the implementation details, convergence criteria, or code needed to check whether the comparison is apples-to-apples. The hardware claims are less directly undermined, but the overall system's value proposition depends on the sampling algorithm being genuinely better, and that dependency is not met. Since the central algorithmic premise is mathematically incorrect and the empirical support is not independently verifiable, the appropriate verdict is REJECT.","tokens_in":19218,"tokens_out":2953,"duration_ms":40165,"concrete_test":"Recompute P_i(sigma) from Eq. (3) on at least 1000 random images from CIFAR10 and ImageNet-1K over the sigma bounds [1e-20, 1000] used in Algorithm 1, and test monotonicity. If P_i is monotonic, re-run the Table II comparison using a properly converged binary search (same epsilon=1e-10 and same sigma bounds as DE). If binary search attains the target perplexity but the DE-SNE accuracy advantage persists, then the advantage is not explained by perplexity-search failure and the central algorithmic claim is unsupported.","verdict_should_be":"REJECT","load_bearing_attack":"Section IV-B and Fig. 6 assert that perplexity sometimes overshoots the target as sigma changes, making the search non-monotonic and binary search unsuitable. This is contradicted by the standard t-SNE mathematics. With Eq. (3), set beta = 1/(2 sigma^2) and p_{j|i} = exp(-beta d_ij)/Z. The entropy H_i satisfies dH_i/dbeta = -beta * Var_p(d) < 0 for distinct distances, so dH_i/dsigma > 0; perplexity P_i = 2^{H_i} is strictly increasing in sigma. Binary search therefore converges, and the stated motivation for replacing it with differential evolution (Algorithm 1) collapses. Without this premise, the DE-SNE contribution is not motivated: Table II may simply compare against an under-converged or incorrectly bounded t-SNE implementation. The paper provides no error bars or code to rule that out, and Section X's claim of 3-16 orders of magnitude perplexity-error reduction over binary search is implausible if both methods converge to the same tolerance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NMS, a system for edge DNN training that combines DE-SNE, a t-SNE variant in which differential evolution replaces binary search for perplexity optimization, with a near-memory computing architecture. The authors claim that DE-SNE is a DNN-free data sampling method that improves generalization over DNN-based coresets, and that placing the sampling circuit near DRAM reduces DDR energy consumption. They report accuracy gains over DQ, DQAS, and NeSSA on six datasets, plus large DDR energy reductions and improved hardware efficiency when integrated with the THETA training accelerator.","tokens_in":19455,"tokens_out":5764,"duration_ms":68490,"significance":"If the claims held, the paper would address two real problems in edge training: generalization bias from DNN-based coresets and data-movement energy. The hardware portion is concrete, with a 28nm synthesis flow, CACTI-based energy estimates, and cycle-accurate simulation, which are strengths relative to purely algorithmic submissions. However, the core algorithmic motivation is mathematically incorrect: perplexity is monotonic in the Gaussian width, so the stated reason for replacing binary search is false. The accuracy comparisons lack error bars and code, and the hardware comparison mixes sampling-accelerator metrics with training-accelerator metrics. The central claims are therefore not supported in their current form.","major_comments":[{"comment":"The premise that perplexity is non-monotonic in sigma is mathematically false under Eqs. (1)-(3). Writing beta = 1/(2 sigma^2) and p_{j|i} = exp(-beta d_{ij})/Z_i, the entropy H_i satisfies dH_i/dbeta = -beta * Var_p(d_{ij}) < 0 whenever the distances are not all equal. Since dbeta/dsigma < 0, H_i and the perplexity P_i = 2^{H_i} are strictly increasing in sigma. Binary search on sigma therefore converges to any target perplexity in exact arithmetic. The stated motivation for replacing binary search with differential evolution collapses, and the accuracy gap in Table II cannot be attributed to perplexity search error without additional evidence.","section":"Section IV-B and Fig. 6"},{"comment":"The ablation does not control for search convergence. Algorithm 1 specifies epsilon = 1e-10 and max iter = 10000; if binary search were run to the same tolerance on a monotonic continuous function, the resulting sigma values would agree to that tolerance and the downstream embeddings should be nearly identical. The reported 4-6 percentage point accuracy improvements are therefore unexplained. The claim in Section X that DE reduces perplexity error by 3, 16, 14, 10, and 12 orders of magnitude relative to binary search and other methods is implausible for a one-dimensional monotonic search and is not supported by any convergence diagnostics or code.","section":"Table II and Section X"},{"comment":"All accuracy numbers are reported as single values with no standard deviations, number of seeds, or code/data release. Since the claimed improvements over DQ, DQAS, and NeSSA are often only 1-3 percentage points, run-to-run variance could account for the differences. Additionally, the target perplexity of 15 and the DE hyperparameters are described as determined experimentally on the same datasets, so part of the reported gain may encode these choices. This is load-bearing because the central claim is superior accuracy.","section":"Tables I-II and Fig. 12"},{"comment":"The hardware comparison mixes a sampling accelerator with training accelerators. The 'Our Work' row reports 66.87 TOPS and 176.12 TOPS/W, but the system under discussion integrates the DE-SNE sampling circuit with the THETA training accelerator. No end-to-end power, latency, or energy-per-trained-epoch measurement is provided for the integrated system, so the statement that the system improves efficiency by 1.2x over THETA lacks support. The comparison to DAC'19, DAC'20, SparTANN, GANPU, and other training accelerators is not apples-to-apples unless all rows execute the same workload.","section":"Table III and Section XII"}],"minor_comments":[{"comment":"There is a typo: 'THATA' should be 'THETA' in the hardware setup paragraph, and 'AMD EPYCT M' should be 'AMD EPYC'.","section":"Section VII-A"},{"comment":"The sentence 'the DE algorithm achieves error reductions of 3, 16, 14, 10, and 12 orders of magnitude on six commonly used datasets' lists five numbers but six datasets and does not identify which comparison corresponds to which dataset.","section":"Section X"},{"comment":"The claim that the bisection search needs up to 5000 iterations is atypical; standard t-SNE implementations use on the order of 50 bisection steps. Please justify this search budget, since it is central to the runtime-bottleneck argument.","section":"Fig. 5"},{"comment":"The CPU and GPGPU energy comparison lacks a stated methodology: no workload details, power models, or measurement sources are given for the four-orders-of-magnitude and three-orders-of-magnitude energy reduction claims.","section":"Fig. 16 and Section XI"},{"comment":"Standard differential evolution samples three distinct indices a, b, and c that differ from i; the current pseudocode only excludes i and could choose a = b, making the mutation degenerate. Please clarify the sampling rule.","section":"Algorithm 1, line 8"}],"recommendation":"reject","confidential_remarks":"The manuscript's central algorithmic premise contradicts standard t-SNE mathematics, and the empirical and hardware evidence does not currently rehabilitate it. I recommend rejection. If the authors can supply code, convergence diagnostics, and an end-to-end hardware evaluation under a corrected motivation, a fresh submission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the hardware work is real: a t-SNE-based sampling accelerator placed near DRAM, with cycle-accurate simulation, RTL, and 28nm synthesis. Second, the paper's main algorithmic justification is false: they claim t-SNE's perplexity search is non-monotonic in sigma, so binary search fails and differential evolution (DE) is needed. In the standard t-SNE setup, perplexity is strictly increasing in sigma, so binary search converges. The entire DE-SNE motivation collapses.\n\nWhat's new and good: NMS is a genuine new combination—DNN-free coreset selection via t-SNE, implemented in near-memory logic, with a plausible energy model. The energy savings (74x vs NeSSA) are the kind of result that could matter for edge training. The accuracy tables show consistent gains over DQ/DQAS/NeSSA across six datasets, and avoiding DNN bias in coreset selection is a legitimate goal. The hardware details are concrete.\n\nThe soft spots, in order of severity. (1) The false premise is load-bearing, not cosmetic. Table II claims DE-SNE beats t-SNE by 3-8 points; if both should find the same sigma, the gap suggests a broken binary search implementation or a different objective. No code, no perplexity errors, no sigma values are provided to rule that out. The '3-16 orders of magnitude' error reduction over binary search is implausible if both converge to the same tolerance. (2) No error bars anywhere, and the reproduced DQ/DQAS numbers are not described well enough to check. (3) The 60% keeping ratio for the THETA comparison is simply asserted; the 'less than 1% degradation' is not shown. (4) 'First DNN-free' overstates novelty: DNN-free coresets (k-center, herding) predate this work; the claim is only defensible for near-memory implementation.\n\nVerdict: as written, no. The central contribution is not established, and the paper overclaims. But the hardware direction is worth a serious revision. If the authors drop the false perplexity claim and present DE-SNE as an empirical alternative to binary search, with real perplexity measurements and error bars, this could become a decent systems paper. I would not send this to peer review in its current state; I'd desk-reject and invite resubmission.","headline":"Hardware story is plausible but the central algorithmic claim—that t-SNE perplexity search needs differential evolution—is mathematically false; paper overreaches.","tokens_in":19950,"tokens_out":3838,"would_cite":false,"duration_ms":44902,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"No-DNN sample picker beats coresets and cuts edge training energy 74x","keywords":["edge DNN training","coreset selection","t-SNE","differential evolution","near-memory computing","dataset compression","generalization","DRAM energy"],"falsifier":"Take random samples from CIFAR-10, CIFAR-100, and ImageNet-1K, and for each image sweep $\\sigma_i$ over a fine grid from $10^{-20}$ to $10^3$, plotting the resulting perplexity $P_i = 2^{H_i}$. If every curve is monotonically increasing, the paper's stated reason for replacing binary search is false and the reported DE-SNE gains would need another explanation; if any curve has a local maximum or multiple crossings at the target perplexity, the paper's mechanism is directly confirmed.","tokens_in":19044,"feed_emoji":"🧠","tokens_out":9328,"duration_ms":102343,"temperature":0.7,"pith_summary":"Training a deep network on an edge device is impractical if every image in a large dataset must cross a long PCB bus from DRAM to the compute chip, and existing dataset-compression methods add a second cost: they use a DNN to judge which samples matter, so the chosen subset inherits that network's bias and transfers poorly to other architectures. This paper proposes a sampling pipeline that avoids DNNs altogether: an improved t-SNE embedding, DE-SNE, maps images into a low-dimensional manifold, grids that space, and picks representative samples that cover the whole feature distribution. The same routine is implemented in logic placed next to DRAM cells, so only the selected images travel to the training accelerator. The claim is that this removes the generalization gap of DNN-based coresets while cutting DRAM energy by up to 74x relative to the closest prior system, and the paper reports accuracy gains on ImageNet-1K of 11.9% over DQ, 9.7% over DQAS, and 4.7% over NeSSA.","feed_headline":"No-DNN sample picker beats coresets and cuts edge energy 74x","feed_subtitle":"An evolutionary t-SNE selects generalizable training subsets inside DRAM, keeping only useful images on the wire.","key_machinery":"The load-bearing machinery is DE-SNE, defined as t-SNE with differential evolution substituted for the binary search that fixes each point's Gaussian width $\\sigma_i$. t-SNE builds conditional probabilities $p_{j|i}$ from $\\sigma_i$ and targets a fixed perplexity; DE-SNE evolves a population of candidate $\\sigma$ values through mutation, crossover, and selection on perplexity error, which the paper argues is needed because perplexity is non-monotonic in $\\sigma$. The embedding output is gridded and sampled in parallel, converting the manifold coordinates back into a covered subset of original images. Around this sits the near-memory sampling circuit, in which processing-element arrays compute the distance matrix, probability matrices, and gradients inside DRAM, with tree elements handling comparisons, so that data movement over the board is limited to the final selected images.","core_discovery":"The central discovery, stated on the paper's own terms, is that representative samples for DNN training can be chosen from the raw data distribution alone, with no forward passes through the target network. t-SNE is used to collapse high-dimensional images into a low-dimensional manifold while preserving local distances, and the usual binary search for the Gaussian width $\\sigma$ is replaced by differential evolution, because the authors find perplexity to be a non-monotonic function of $\\sigma$ and therefore unsuited to bisection. The embedding is then gridded and sampled in parallel so the selected subset covers the feature space. Since no DNN is involved, the same subset transfers across ResNet, ShuffleNet, MobileNet, and ViT. Implementing the DE-SNE arithmetic in 3D-stacked near-memory DRAM logic means only the chosen images cross the PCB, yielding the reported average Top-1 improvements on ImageNet-1K of 11.9% over DQ, 9.7% over DQAS, and 4.7% over NeSSA, and DRAM energy reductions up to 74x at a 10% keeping ratio.","pith_inferences":["The paper leaves implicit that the same differential-evolution fix could stabilize $\\sigma$ search in other t-SNE applications, such as single-cell RNA visualization or word-embedding plots, where the binary-search bottleneck is identical.","The near-memory blueprint generalizes: any DNN-free selection rule cheap enough to run inside DRAM logic could make dataset pruning, active learning, and replay-buffer construction nearly energy-free for edge devices.","A natural but untested extension is to check whether DE-SNE subsets also transfer across optimizers, augmentation policies, and label shifts, where DNN-based coresets are known to be brittle.","If the non-monotonic perplexity claim is confirmed, it would suggest that many published t-SNE embeddings have silently used failed or imprecise $\\sigma$ searches, so DE-SNE could improve visualization fidelity independently of its use in coresets."],"forward_implications":["Because sampling no longer depends on a DNN, the same DE-SNE subset can be reused across heterogeneous architectures and can be computed once per dataset rather than once per model.","Moving the selector into DRAM removes most of the energy cost of reading the full dataset: reported DRAM energy savings range from about 32x to 74x depending on the keeping ratio.","The approach scales in its advantage: the largest accuracy gains over DQ, DQAS, and NeSSA appear on ImageNet-1K, the largest dataset tested.","DE-SNE's evolutionary search makes t-SNE's $\\sigma$ computation stable enough for fixed-point hardware, enabling manifold sampling on battery-powered devices.","At a 60% keeping ratio, training on the sampled set loses less than 1% accuracy compared with full data while the combined sampling-plus-training system improves energy efficiency by 1.2x over a sparse training accelerator."],"supporting_citations":[{"why":"Supplies the t-SNE algorithm whose manifold embedding and perplexity search DE-SNE modifies.","marker":"[19]"},{"why":"The DQ coreset baseline that motivates DNN-free sampling and is compared for accuracy and DRAM energy.","marker":"[13]"},{"why":"The NeSSA near-storage sampling system, the closest prior architecture and main DRAM-energy baseline.","marker":"[30]"},{"why":"Provides the differential evolution optimizer that replaces binary search in the perplexity step.","marker":"[37]"},{"why":"The binary-search method that DE-SNE replaces when solving for the Gaussian width $\\sigma$.","marker":"[50]"},{"why":"The DQAS adaptive-sampling coreset baseline used in the accuracy comparisons.","marker":"[46]"},{"why":"The distribution-matching condensation method cited for the across-architecture generalization failure of DNN-based sampling.","marker":"[12]"},{"why":"Provides the 0.5 pJ/bit UCIe near-memory energy figure used in the DRAM energy accounting.","marker":"[44]"},{"why":"Provides the 10 pJ/bit PCIe board-level DRAM access energy figure used in the comparison.","marker":"[45]"}],"fun_headline_variants":["DNN-free t-SNE sampling in DRAM reduces edge training energy 74x","Near-memory manifold sampling reduces DDR energy 74x for edge training","DE-SNE selects subsets in memory to reduce DNN bias and energy","In-memory t-SNE sampling for DNN-free edge training, 74x less DDR energy","Manifold-based sampling without DNNs reduces edge energy 74x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the claim that t-SNE's perplexity is a non-monotonic function of the Gaussian width $\\sigma$, so binary search can fail and an evolutionary search is needed; the paper asserts this without proof.","fun_headline_variants_meta":{"raw":{"variants":["DNN-free t-SNE sampling in DRAM reduces edge training energy 74x","Near-memory manifold sampling reduces DDR energy 74x for edge training","DE-SNE selects subsets in memory to reduce DNN bias and energy","In-memory t-SNE sampling for DNN-free edge training, 74x less DDR energy","Manifold-based sampling without DNNs reduces edge energy 74x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000912,"raw_usage":{"total_tokens":4000,"prompt_tokens":1109,"completion_tokens":2891,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":725,"completion_tokens_details":{"reasoning_tokens":2785}},"tokens_in":725,"tokens_out":2891,"duration_ms":22711,"temperature":1.0,"reasoning_tokens":2785,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:01:06.713768+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take random samples from CIFAR-10, CIFAR-100, and ImageNet-1K, and for each image sweep $\\sigma_i$ over a fine grid from $10^{-20}$ to $10^3$, plotting the resulting perplexity $P_i = 2^{H_i}$. If every curve is monotonically increasing, the paper's stated reason for replacing binary search is false and the reported DE-SNE gains would need another explanation; if any curve has a local maximum or multiple crossings at the target perplexity, the paper's mechanism is directly confirmed.","supporting_citations":[],"review_version":1}