{"id":"3e57b861-0df7-4ab4-93f7-a6d16a8a6468","arxiv_id":"2607.17371","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ADEPT selects CNN blocks for fine-tuning on PIM accelerators using a sensitivity-to-EDP ratio plus a selective channel-wise low-rank adapter, cutting training EDP by up to 8.1x with roughly unchanged accuracy.","lead":"This paper presents ADEPT, a method for fine-tuning CNNs on processing-in-memory (PIM) accelerators by training only the most useful blocks, chosen by a score that weighs how much a block needs to change against the energy and delay cost of training it. It reports up to 8.1x lower energy-delay product than full-parameter fine-tuning, at comparable accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"8.1x EDP claim is benchmarked against AutoRGN, not full-parameter fine-tuning; FpT accuracy and EDP are never reported.","rationale":"The reader's verdict is CONDITIONAL, and the central concern I identify — the absence of a direct full-parameter fine-tuning baseline — reinforces that conditionality. The reader did mention the lack of a full-parameter EDP row in their rationale, but their stated weakest_assumption focuses on the stability of the one-time sensitivity calibration. I considered that assumption and agree it is a real risk: if the SER ranking becomes stale, ADEPT's block selection could be misdirected, and the claimed accuracy/EDP trade-off could degrade. However, the paper's results demonstrate competitive accuracy, and the nested sampling ensures all blocks receive non-zero update probability, partially mitigating stale rankings. In contrast, the missing FpT baseline is a direct evidentiary gap for the headline number itself. The 8.1x figure is computed against AutoRGN, not FpT, and no FpT accuracy is reported anywhere. Since the central claim is explicitly 'compared to full-parameter fine-tuning', the evidence does not yet establish that claim. This is not an internal inconsistency in the method, but it is a load-bearing absence of a control condition. A simple experiment — running FpT in the same simulator and evaluation setup — would settle the issue. My recommendation is to keep the paper CONDITIONAL pending this comparison, which is consistent with the reader's verdict, hence UNCHANGED.","tokens_in":22096,"tokens_out":7462,"duration_ms":79203,"concrete_test":"Run full-parameter fine-tuning (FpT) with the same hyperparameters (batch size, epochs, learning rates, 16-bit fixed-point precision, NeuroSim configuration) for ResNet-18, MbNet-V2, SnetV2-x1, and VGG-16 on the same datasets (CIFAR10, CIFAR10-c, Entity-30, ImageNet) and PIM platforms (PUMA, CIMAT, HuNT). Measure FpT accuracy (Table VII-style) and FpT fine-tuning EDP with the same simulator. Then compute ADEPT's EDP ratio relative to FpT (not AutoRGN) and the accuracy delta. If the ratio falls below 8.1x or the accuracy delta exceeds a predefined tolerance (e.g., >1%), the abstract's claim must be revised to specify the comparison baseline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and Key Contribution 3 state that ADEPT achieves up to 8.1x lower EDP compared to full-parameter fine-tuning while maintaining comparable accuracy. However, the experimental section (V.E) compares ADEPT only with AutoRGN, L2F, ADEPT(S), and a uniform-DCLoRA baseline; Fig. 10 and Fig. 11 report EDP normalized to ADEPT for these methods, with AutoRGN serving as the proxy for full-parameter training. AutoRGN is described in Section II.B as updating all layers 'similar to full-parameter training (FpT)', but it is not identical: it computes per-layer learning rates from gradient/parameter norms, which can affect both accuracy and training cost. Table VII, which lists post-fine-tuning accuracy, omits FpT entirely. Consequently, the central claim's two components — the 8.1x EDP reduction and the 'comparable accuracy' guarantee — are not directly supported by any measurement of FpT. If AutoRGN is more energy/latency expensive than plain FpT due to its learning-rate machinery, the true EDP reduction versus FpT would be smaller; if FpT achieves higher accuracy than AutoRGN, the accuracy gap between ADEPT and FpT could exceed the paper's implied 'comparable' range. This is a load-bearing gap because the headline result is precisely a comparison to FpT.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces ADEPT, a hardware-aware framework for fine-tuning CNNs on PIM accelerators. The central idea is a Sensitivity-EDP Ratio (SER) that ranks each CNN block by the ratio of its gradient-based sensitivity to its per-block Energy-Delay Product (EDP) cost. A stochastic nested training scheme then samples blocks for each mini-batch with probability proportional to their SER rank, and a dynamic channel-wise LoRA (DCLoRA) is applied selectively to channel-dense blocks to reduce gradient-memory traffic. The paper claims up to 8.1x lower EDP compared to full-parameter fine-tuning while maintaining comparable accuracy, and also presents a continual-learning extension that refreshes the sensitivity component via on-chip EMA updates. Experiments are carried out across four CNN models, four datasets, and three PIM platforms (PUMA, CIMAT, HuNT) using NeuroSim-based modeling.","tokens_in":22418,"tokens_out":4850,"duration_ms":50220,"significance":"If the central claims are supported, ADEPT is a useful contribution to efficient on-device fine-tuning for PIM architectures. The SER metric is a principled way to make layer selection architecture-aware, and the DCLoRA construction addresses a real mismatch between standard LoRA and CNN weight tensors. The paper is transparent about the hardware models and provides extensive evaluations across models, datasets, and accelerators. The continual-learning extension with EMA-based sensitivity refresh is a thoughtful, low-overhead addition. However, the headline comparison to full-parameter training is not directly measured, and the reported accuracy numbers are inflated by test-set-based rank selection. These issues are load-bearing for the stated contributions, so the empirical support is currently incomplete.","major_comments":[{"comment":"The headline claim in the abstract and Key Contribution 3 is that ADEPT achieves up to 8.1x lower EDP than full-parameter fine-tuning (FpT), but no FpT baseline is measured. The EDP comparisons in Fig. 10 are normalized to ADEPT and show AutoRGN, L2F, ADEPT(S), and a uniform DCLoRA baseline; AutoRGN is described as updating all layers 'similar to' FpT, but it also computes per-layer learning rates from gradient/parameter norms, which adds overhead absent from plain FpT. Table VII likewise omits FpT accuracy. The 8.1x figure is therefore only a comparison to AutoRGN, not to FpT. Please add actual FpT measurements (accuracy and EDP on the same platforms) or revise the claim to state the comparison is against AutoRGN.","section":"§V.E, Fig. 10"},{"comment":"The final LoRA rank is selected using test accuracy: 'After fine-tuning, we employ the rank with highest test accuracy in R for inference.' This is a form of test-set leakage. The reported ADEPT accuracies in Table VII are optimistic and do not reflect a deployment scenario where the test set is unavailable. The 'comparable accuracy' claim is therefore not supported for a realistic protocol. The rank should be chosen on a validation split, or the paper should report accuracy for a fixed rank (e.g., the largest rank) and show that the test-selected rank does not materially change the conclusions.","section":"§IV.C"},{"comment":"All accuracy and EDP results are point estimates with no repeated-seed variance, confidence intervals, or significance tests. In Table VII, ADEPT and AutoRGN differ by only 0.1-0.5 percentage points in several configurations; without a variance estimate, the claimed 'comparable accuracy' could be noise. Likewise, the EDP ratios in Figs. 9-11 come from single runs of the calibration pass, and the sensitivity estimates in Eq. (3) are computed from a small number of mini-batches, so the resulting SER ranking could vary across initializations. At minimum, report three to five seeds for a representative subset (e.g., ResNet-18 on CIFAR10, CIFAR10-c, and Entity-30) and show the ranges for Table VII and the EDP figures.","section":"§V.A, Table VII, Figs. 9-11"},{"comment":"The SER priority map is computed once during a short calibration pass and never updated during static fine-tuning. The continual-learning experiments in Table IX demonstrate that the sensitivity profile drifts across corruption stages, which raises the concern that it may also drift within a single fine-tuning task. If the initial gradient-norm ratios are unrepresentative of later iterations, the stochastic freezing pattern is misdirected and the claimed accuracy/EDP trade-off weakens. Please provide an ablation in which the SER map is recomputed mid-training (or compare the static SER against an oracle that re-ranks periodically) to show that the one-time calibration is sufficient.","section":"§IV.A, Algorithm 1, Table IX"}],"minor_comments":[{"comment":"The parameter-reduction ratio expression is hard to parse: the parentheses and the 'r/(k)^2' term need clarification. State more explicitly that C_in = C_out = C and rewrite the ratio with unambiguous parentheses.","section":"Eq. (11)"},{"comment":"The y-axis is normalized independently per architecture, so the reader cannot compare absolute EDP between PUMA and TPU. Add an explicit statement in the caption that each architecture is normalized to its own Ip.conv EDP.","section":"Fig. 6"},{"comment":"The table is dense and the row/column structure is difficult to follow. Consider splitting into separate tables per model or adding boldface/color to distinguish metrics and baselines.","section":"Table VII"},{"comment":"The caption for Fig. 10 says 'Normalized EDP of ADEPT and baseline fine-tuning techniques' but the legend includes ADEPT(S) and a uniform DCLoRA baseline; make the normalization reference explicit to avoid confusion.","section":"§V.E"},{"comment":"The EMA update in Eq. (12) is written as S_b ← S_b + (1-λ)(x_b - S_b). This is algebraically correct, but the subsequent sentence says 'with S_b initialized during the one-time calibration step' while also using λ; the notation λ and k should be defined immediately before first use in the equation.","section":"§IV.E, Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the framework is interesting and well-motivated. The main gap is that the headline claim is not supported by the experiments as reported: the 8.1x EDP figure compares to AutoRGN, not full-parameter fine-tuning, and the test-set-based rank selection inflates the accuracy numbers. These issues are fixable with additional experiments and a revised claim. If the authors add a real FpT baseline and adopt a validation-based rank selection, the paper could be acceptable; otherwise, the central contribution should be reframed. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, ADEPT is a sensible hardware-aware fine-tuning framework: the SER metric (sensitivity over EDP) and the channel-wise LoRA variant are genuinely new combinations, and the paper shows that hardware cost changes which blocks you should fine-tune, which earlier methods miss. Second, the headline claim is not actually supported by the data. The abstract and conclusion say 8.1x lower EDP than full-parameter fine-tuning, but every comparison in Section V.E is against AutoRGN, which updates all layers but with a different learning-rate rule. There is no full-parameter (FpT) row in Table VII or in Figures 9-11. AutoRGN is a proxy, not the thing the claim is about. That's a load-bearing gap.\n\nThe real contribution is the combination of ideas. SER is a clean way to fold hardware cost into layer selection, and the architecture-dependent rankings (memory-centric PUMA vs compute-centric TPU) are a nice result. DCLoRA makes sense for CNNs, where filter sizes are too small for standard LoRA; applying it selectively based on EDP reduction is defensible. The continual-learning extension with the EMA refresh is thoughtful, and the overhead numbers are plausible. The paper is well-written and the sensitivity analysis (Fig. 7) justifies why static heuristics fail.\n\nSoft spots, in order of size. The missing FpT baseline is the main one; without it the central claim collapses to 'better than AutoRGN,' which is still useful but not the abstract's promise. Second, the final LoRA rank is chosen by test-set accuracy, which leaks test information into model selection. Use a validation split or report accuracy across ranks. Third, no error bars or repeated seeds, and no code or data; given the whole evaluation is simulated, that would help. The one-time sensitivity calibration is a fair concern but the continual-learning experiment shows the authors know it can go stale; it's not a fatal flaw.\n\nWho this is for: people working on on-device training, PIM accelerators, and PEFT for CNNs. It's a plausible systems contribution with a fixable evidential gap. I'd send it to peer review, but the referee should require an actual FpT comparison and a model-selection fix before the 8.1x number appears again.","headline":"Good hardware-aware fine-tuning idea, but the 8.1x EDP claim compares to AutoRGN, not full-parameter training.","tokens_in":22933,"tokens_out":3119,"would_cite":false,"duration_ms":32667,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ADEPT shows CNN fine-tuning on processing-in-memory accelerators can achieve up to 8.1x lower energy-delay product than full-parameter training with comparable accuracy, by ranking blocks by sensitivity-to-hardware-cost and adapting channel","keywords":["processing-in-memory","CNN fine-tuning","energy-delay product","low-rank adaptation","hardware-aware neural network training","layer selection","continual learning","edge accelerators"],"falsifier":"Run ADEPT on a PIM simulator but replace the pre-computed SER ranking with a ranking recomputed every epoch (or every few mini-batches) from the same sensitivity formula; if the fully-refreshed version beats ADEPT's static ranking by more than a small margin in accuracy at equal EDP, the one-time calibration assumption is falsified. Equivalently, compare the top-k SER block sets between the first epoch and the last epoch under a distribution shift like CIFAR10-c: if they overlap by less than half the blocks, the ranking was not stable.","tokens_in":21984,"feed_emoji":"⚡","tokens_out":7530,"duration_ms":65727,"temperature":0.7,"pith_summary":"This paper sets out to show that the energy cost of fine-tuning a pre-trained CNN on a processing-in-memory (PIM) accelerator can be cut by nearly an order of magnitude without losing accuracy, if the choice of which blocks to train is driven by the actual hardware cost of training each block rather than by gradient sensitivity alone. The framework, ADEPT, ranks each block by a Sensitivity-EDP Ratio (SER)—the block's relative gradient norm divided by its per-block energy-delay product on the target architecture—and then uses a nested stochastic schedule in which a block's update frequency is proportional to its rank. It pairs this with a channel-wise low-rank adaptation (DCLoRA) that shrinks gradient traffic selectively in channel-dense later layers, while leaving early layers alone because their cost is dominated by stored activations. On four CNN models and multiple datasets, the paper reports up to 8.1x lower energy-delay product than full-parameter fine-tuning, with accuracy within roughly a point of full retraining, and shows the same mechanism extends to continual learning via an on-chip sensitivity refresh.","feed_headline":"Fine-tuning CNNs in-memory: 8.1x lower energy-delay","feed_subtitle":"Hardware-aware layer selection plus low-rank adapters matches full retraining accuracy at a fraction of energy cost.","key_machinery":"Sensitivity-EDP Ratio (SER): a per-block score SER_b = (S_b)^n / (T_b)^m, where S_b = ||g_b||_2 / ||theta_b||_2 (relative gradient norm) and T_b is the block's aggregated energy-delay product from a PIM-specific latency/energy model (Eqs. 4-7, including off-chip DRAM traffic). Nested block-wise training: the ranked blocks form subsets V_k = {B_1,...,B_k}; each mini-batch samples one V_k, so block B_i trains with probability (z-i+1)/z, making update frequency exactly track SER rank. Dynamic Channel-wise LoRA (DCLoRA): low-rank factors A_l (k x k x C_in x r) and B_l (r x C_out x 1 x 1) are added along channels rather than spatial filters, with rank sampled per mini-batch from {8,...,64}; adapt","core_discovery":"This paper claims that hardware-aware fine-tuning can largely neutralize the off-chip memory bottleneck that dominates CNN fine-tuning on PIM accelerators. Its central object is the SER ranking: each block receives a score S_b^n / T_b^m, where sensitivity S_b is the ratio of gradient norm to parameter norm and training overhead T_b is the block's EDP computed from a memory-system model that includes on-chip communication and off-chip DRAM accesses. Blocks are sorted by SER and arranged into nested subsets; each mini-batch samples one subset, so more-sensitive, cheaper blocks are updated more often and frozen blocks generate no activation or gradient traffic. On the rank side, DCLoRA applies","pith_inferences":["Editorial extension: the SER idea is not tied to CNNs or LoRA; any block-structured PEFT method (adapter, pruning mask, or even prompt) could be gated by the same sensitivity-to-hardware-cost ratio, though the paper only demonstrates it for CNN blocks with channel-wise LoRA.","Editorial extension: because update probability is a pure function of SER rank, ADEPT's training schedule is effectively 'stochastic depth with hardware-informed survival rates'; this suggests a clean ablation—hold EDP fixed and shuffle the ranking—to isolate how much of the accuracy win comes from the hardware-aware ordering versus from stochastic depth itself.","Editorial extension: the calibration assumption could be stress-tested cheaply by recomputing S_b from later epochs and comparing the top-k block sets; if the ranking flips frequently on a drifting distribution, the static version of ADEPT will eventually underperform, which is exactly the scenario the EMA extension addresses.","Editorial extension: the reported 8.1x EDP gain is relative to full fine-tuning on the same platform; a deployment would also need to amortize one-time costs (loading the pretrained weights and the calibration pass) across many fine-tuning rounds before the gain materializes."],"forward_implications":["If the results transfer, on-device fine-tuning for federated and continual-learning deployments becomes materially cheaper, since those workloads fine-tune repeatedly rather than once.","Layer selection becomes architecture-dependent by construction: the same model will fine-tune different blocks on a memory-centric PIM than on a compute-centric accelerator, because off-chip weight fetches change each block's effective cost.","DCLoRA should not be applied uniformly: the paper's per-block EDP analysis shows little benefit in early layers where activations dominate, implying that hardware-aware gating is essential for LoRA-like methods on CNNs.","The on-chip controller that implements stochastic freeze/unfreeze and rank sampling adds under 0.5% tile area/power, so the EDP savings are achievable without a host CPU in the loop.","The continual-learning variant tracks a full-training baseline within 0.3% accuracy across sequential corruption stages at under 2% extra EDP, suggesting the hardware-aware ranking can be maintained under distribution drift."],"fun_headline_variants":["ADEPT: Adaptive CNN fine-tuning for PIM accelerators","Hardware-aware CNN fine-tuning: 8.1x lower EDP","ADEPT: Cut CNN fine-tuning energy 8.1x on PIM","Layer-aware tuning on PIM: 8.1x energy-delay drop"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything hinges on a one-time estimate: the framework measures each block's importance from a few initial mini-batches and a static hardware-cost model, and then freezes or trains blocks for the whole run according to that fixed ranking — if those early estimates mislead, the whole accuracy-versus-energy trade-off breaks.","fun_headline_variants_meta":{"raw":{"variants":["ADEPT: Adaptive CNN fine-tuning for PIM accelerators","Hardware-aware CNN fine-tuning: 8.1x lower EDP","ADEPT: Cut CNN fine-tuning energy 8.1x on PIM","Layer-aware tuning on PIM: 8.1x energy-delay drop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1230,"prompt_tokens":740,"completion_tokens":490,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":408}},"tokens_in":484,"tokens_out":490,"duration_ms":4809,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T18:09:28.896552+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ADEPT on a PIM simulator but replace the pre-computed SER ranking with a ranking recomputed every epoch (or every few mini-batches) from the same sensitivity formula; if the fully-refreshed version beats ADEPT's static ranking by more than a small margin in accuracy at equal EDP, the one-time calibration assumption is falsified. Equivalently, compare the top-k SER block sets between the first epoch and the last epoch under a distribution shift like CIFAR10-c: if they overlap by less than half the blocks, the ranking was not stable.","supporting_citations":[],"review_version":1}