Pith. sign in

REVIEW 3 major objections 5 minor 78 references

Across 16 DNN workloads and three IEEE 754 formats, the lowest fraction bits can be flipped without crossing a 1% task-quality bound, and this threshold can be turned into a cheaper unequal-error-protection memory design.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 12:13 UTC pith:YNS3ZPBP

load-bearing objection Solid empirical bit-sensitivity study, but the BF16 UEP design uses X=5 while the paper's own floor is X=4, undercutting the headline hardware savings. the 3 major comments →

arxiv 2607.19623 v1 pith:YNS3ZPBP submitted 2026-07-21 cs.AR cs.LG

From Bit-Position Sensitivity to Unequal Error Protection for DNN Inference Memory

Muhammad Husnain Mubarik , Karthik Mohan Kumar , Pedro Antonio Pena , Keshavan Varadarajan , Kunal Tyagi This is my paper
classification cs.AR cs.LG
keywords per-bit-position fault sensitivityunequal error protectionXsafe thresholdfloating-point reliabilityDNN inference memoryselective ECCdual-voltage SRAMsoft errors
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper's central claim is that the IEEE 754 floating-point formats used in DNN inference have a built-in safety margin: the lowest few mantissa bits can be corrupted by memory faults with less than 1% degradation in task quality, while bits near the exponent field cause catastrophic collapse. By systematically flipping each bit position across 16 models in FP16, BF16, and FP32, the paper derives a data-type-specific threshold, Xsafe (FP16=6, BF16=4, FP32=15), below which ECC coverage is unnecessary. It then builds an unequal-error-protection codec and a dual-voltage SRAM architecture that leaves those bits unprotected, cutting ECC logic area by 27.8% and BF16 read energy by about 17%, without retraining. A sympathetic reader should care because it turns a reliability cost into a function of the data format itself, and offers a concrete hardware design parameter.

Core claim

Flipping any of the least-significant fraction bits up to a data-type-specific threshold, Xsafe, degrades task metrics by under 1% in deterministic single-bit stress tests; sensitivity then ramps through the upper fraction bits and spikes at the exponent-mantissa boundary, where a single-bit flip is catastrophic. The paper argues this transition is a property of the IEEE 754 data type, not of any particular model architecture: attention-free CNNs show the same pattern as transformers, and statistical tests place significance only at exponent and sign positions. On that basis it claims conservative floors FP16 Xsafe=6, BF16 Xsafe=4, FP32 Xsafe=15 are stable across the 16 evaluated workloads,

What carries the argument

The central object is Xsafe, the number of least-significant fraction bits below which flips are benign; it is read off the sharp 'cliff' in per-bit-position sensitivity curves where quality holds flat then collapses near the exponent boundary. The carrying mechanism is a three-tier unequal-error-protection codec: sign/exponent bits get full SECDED (odd-parity codes), upper fraction bits get single-error correction (even-parity codes), and the Xsafe low bits are stored with no code, physically separated into a low-voltage SRAM partition selected by a per-cacheline 3-bit tag.

Load-bearing premise

The argument rests on assuming that measuring sensitivity at one injection site per model—pre-softmax QK^T scores for attention models and FFN/MLP outputs otherwise—reveals the most fault-sensitive tensor; if some other tensor (embeddings, normalization statistics, residual stream, or an untested layer) is more sensitive at low fraction bits, the Xsafe floor would not be a true worst-case bound.

What would settle it

Run the deterministic single-bit-position sweep (pword = 1, every word's bit toggled) on every tensor of a text-conditioned diffusion model—embeddings, normalization statistics, residual stream, and attention outputs—in BF16, and check whether any fraction bit at or below bit 3 degrades PSNR by more than the allowed bound (first bit before a single-step PSNR drop > 5 dB); if such a bit exists, the BF16 floor of 4 is not worst-case.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Bit-protection thresholds can be fixed in hardware at design time because they vary by at most one bit across datasets, injection sites, and sample counts.
  • A single 136-bit physical word and one codec serve FP16, BF16, and FP32; only the per-cacheline tag changes routing.
  • Multi-cell upsets that straddle the Xsafe boundary contract the safe region by at most one bit per additional flipped bit, so the floor remains usable under MCU-2 and MCU-3 faults.
  • Leaving the low partition unprotected at reduced voltage yields ~17% gross BF16 read-energy savings with a ~4% macro-area overhead recovered by the energy savings, and the most aggressive dual-voltage operating point keeps MTTF around 900 years.
  • Workload-aware tiers widen the unprotected region (up to X=10 for FP16) for vision, NLU, and resilient LLM workloads, boosting ECC savings to 37.5–62.5% without retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the same bit-sensitivity logic applies to weights once protection cost is amortized over a load-once lifetime; since it notes weight LSBs are tolerant, a weight-side UEP is a natural extension.
  • The per-layer sensitivity spreads, which exceed an order of magnitude across LLM layers, suggest that per-tensor protection tiers could reclaim more energy than the per-model tiers reported here.
  • Because Xsafe is defined against task-specific quality bounds, stricter safety contracts would tighten the floors monotonically; applications with hard quality requirements should treat the quoted floors as ceilings, not defaults.
  • The codec's reduced double-error detection means a safety-critical deployment should pay the extra check bit to restore full SECDED; the paper enables this as an option, making the throughput-versus-safety tradeoff a policy decision.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper characterizes per-bit-position fault sensitivity in FP16/BF16/FP32 values across 16 DNN workloads, reports a sharp transition between benign low-order fraction bits and catastrophic exponent-bit faults, and uses the observed threshold Xsafe (FP16=6, BF16=4, FP32=15) to motivate a three-tier Unequal Error Protection (UEP) codec and a dual-partition SRAM. The claimed contributions are: a data-type-level Xsafe floor stable across workloads; workload-aware protection tiers; an RTL-synthesized UEP codec with per-cacheline format tags; validation under random, persistent, and multi-bit correlated upsets; and area/energy estimates showing a 27.8% codec-area reduction and roughly 17% gross BF16 read-energy savings.

Significance. If the empirical floors and their generality hold, this is a practically relevant result for ML inference memories: a fixed hardware partition, a small per-cacheline tag, and no retraining could make selective ECC deployable across FP16/BF16/FP32 accelerators. The paper has genuine strengths: 16 models across 6 modalities, deterministic sweeps complemented by stochastic, persistent, and MCU-2/3 campaigns, statistical tests (McNemar, Wilcoxon), RTL verification with 146k+ vectors, and an honest discussion of limitations in §7. However, the load-bearing generalization from a few injection sites to all tensors, and the internal mismatch between the certified Xsafe floor and the implemented RTL bypass width, must be resolved before the hardware claims are credible.

major comments (3)
  1. [§3.3, §4.1.1, Table 5, Table 6] Xsafe is defined as the number of fraction LSBs that can safely remain unprotected. Under that definition, for FP16 (10 fraction bits), BF16 (7 fraction bits), and FP32 (23 fraction bits), Xsafe cannot exceed the fraction width. Table 5 lists FP16 values 11, 12, 14 (ViT-L/16, BERT-base, Swin-B), BF16 values 8 (BERT, ViT, Swin), and FP32 values 24, 25, 27, all of which exceed the corresponding fraction widths. Table 6 deepens the contradiction: for ViT-L/16 it reports Xsafe=11 but labels b10 as the first bit above Xsafe, and b11 (which should be below the threshold) already shows a 3.94pp top-1 drop, crossing the 1pp bound. Since Table 5 feeds the workload tiers in Table 8, the tier boundaries inherit this inconsistency. The authors should correct the per-model Xsafe values, clarify the bit indexing used, or revise the definition so that Table 5 and Table 6 are mutually consistent.
  2. [§5.1.5, Table 11, Table 14, §6.4] The hardware implementation and validation do not use the certified BF16 floor Xsafe=4. Table 11 and Table 14 size BF16 Partition B as 40 data bits (5×8 words), i.e., X=5 unprotected fraction LSBs, and §6.4 confirms the RTL operating point is k=11 for BF16 with '5 unprotected LSBs, bits 0–4.' The text calls this a conservative boundary distinct from Xsafe, but the paper's headline energy and area claims (17% gross BF16 read-energy reduction, 27.8% codec-area reduction) and the end-to-end task validation in Tables 20–21 are all computed at this wider X=5 point. At the actual floor X=4, Partition B would be 32 bits/word, f_nc would drop from 40/136 to 32/136, and the energy saving would shrink accordingly. The paper needs to either validate and justify X=5 as its own design point, or re-derive the hardware claims and validation at Xsafe=4.
  3. [§3.3, Table 4, §4.1.1] The claim that Xsafe is a data-type-level property, not a model artifact, is load-bearing for the hardware design, because a single floor is applied to all inference-produced values. However, per §3.3, the primary injection site is the pre-Softmax QK^T score tensor for attention models and FFN/MLP outputs otherwise, with additional loci tested only 'for selected models when a within-model sensitivity ranking is needed.' Table 4 reports at most a one-bit shift across injection sites, but this axis is exercised on a subset of models only. The floor-setting models (e.g., SD3.5-Medium, PixArt-α) are not shown with per-tensor Xsafe sweeps covering residual streams, normalization statistics, embeddings, or KV-cache tensors. The authors should report per-site sweeps for at least the floor-setting models, or temper the universality conclusion.
minor comments (5)
  1. [Table 8] The tier table is hard to read and appears internally inconsistent. The 'Models in tier' entries (4, 7, 16) seem to mix non-cumulative and cumulative counts, and Tier 1 at X=9 cannot include 'all below' if models with Xsafe=6 are in the same list. Please recast the table so coverage counts and membership are unambiguous.
  2. [§4.1.1, Table 6] The prose says Llama-3 8B perplexity 'remains stable through bit 11' while Table 6 reports 7.01 at b11 against a baseline of 6.64, which is already above the 1% bound used to define Xsafe. Please align the narrative with the table values.
  3. [§3.2, §4.1.3] The C4/HellaSwag external-validity check for Llama-3 8B is mentioned in §3.2 but no result is reported. Either add the data or remove the claim.
  4. [§5.3.2, Eq. (3)–(4)] The read-energy saving is reported as 'gross' and 'net positive after level shifters,' but the dynamic/leakage split and the read vs. write energy mix are not stated. Clarifying the energy model would make the 17% figure easier to reproduce.
  5. [§6.4] The RTL operating points k=11/12/24 are described as 'conservative boundaries distinct from the Xsafe floors,' but the relationship between k and Xsafe is never formalized. A short derivation or table mapping k to the number of unprotected LSBs per format would remove ambiguity.

Circularity Check

0 steps flagged

No significant circularity: Xsafe thresholds are empirically calibrated and cross-validated under independent fault patterns; the BF16 X=5 vs Xsafe=4 mismatch is an internal consistency issue, not circularity.

full rationale

The paper's central claims are empirical rather than deductive. Xsafe is defined in §3.3 as the number of fraction LSBs that can remain unprotected while staying within a modality-specific quality bound, and then measured in §4.1.1 via deterministic single-bit-position sweeps. The later validations (§4.1.2–§4.1.5, §6.4) use different fault families—persistent stuck-at faults, stochastic BER sweeps, MCU-2/3, and tiered-BER end-to-end runs—so the safety claim is cross-validated, not obtained by restating the calibration. The hardware equations (§5.1.5, §5.3) are arithmetic consequences of the chosen code parameters, and the ECC area/energy numbers follow from the codeword construction rather than from the sensitivity data alone. No author-overlap citations are invoked as load-bearing uniqueness arguments, and no ansatz is imported via self-citation. The one notable issue is an internal inconsistency: the BF16 RTL operating point k=11 exposes 5 unprotected LSBs (§6.4) while the certified floor is Xsafe=4 (Table 5), and the headline BF16 energy saving is computed at f_nc=40/136 (§5.3.2). This is a correctness/consistency concern about whether the evaluated configuration matches the certified floor, not a circularity: the k=11 experiment is an independent (if mismatched) measurement, not a reduction of the conclusion to its premise.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central safety claim is an empirical calibration governed by quality bounds; the hardware gains follow from simulation models and the representativeness of the injection site. The Xsafe floors are fitted values, not derived constants, and the codec/partition are proposed designs with no external falsifiable handle outside the paper.

free parameters (5)
  • Xsafe floors (FP16/BF16/FP32) = 6 / 4 / 15
    Empirically selected bit positions where deterministic single-bit stress crosses each modality's quality bound; minima across 16 models. Not predicted from theory; shifts with quality bound (Table 4).
  • FP16 workload tiers X = 6 / 9 / 10
    Chosen from per-model Xsafe values (Table 5) to create tier groups; conservative family minima.
  • Modality quality bounds = PPL≤1.01x; acc≥-1pp; WER≤+1pp; first bit before PSNR drop >5dB
    These thresholds define Xsafe; a stricter bound tightens the floors monotonically, so Xsafe is calibration-dependent.
  • Low-voltage partition operating point Vlow = 1.2 V (sky130, Vnom=1.8 V)
    Selected from Figure 11 so that net read-energy savings remain positive after level-shifter overhead.
  • RTL validation boundaries k = BF16 k=11 (5 unprotected LSBs); FP16 k=12 (4); FP32 k=24 (8)
    Operating points used for end-to-end BER validation; note BF16's k=11 exposes one more LSB than its Xsafe floor of 4, an internal inconsistency.
axioms (5)
  • domain assumption The measured injection loci (QK^T scores for attention; FFN/MLP outputs otherwise) are worst-case representatives for all tensors in each model.
    Xsafe floors are model-level parameters derived from these loci; Table 4 tests injection-site variation on only a subset of models.
  • domain assumption Deterministic single-bit stress at p_word=1 plus random/MCU sweeps adequately captures SRAM soft-error behavior relevant to the design.
    The paper acknowledges in §7 that synthetic/bounded-neighborhood injection only partially captures field-correlated fault processes.
  • standard math The odd/even parity-class capacity split of Hsiao SECDED supports Tier 1/Tier 2 code assignment as encoded in Eqs. (1)-(2).
    Standard coding-theory counting of odd/even parity columns; used to set r=7 or r=8.
  • domain assumption SRAM bit-error rate at Vlow follows the Stutz mapping, and CACTI 7.0 plus published 7nm scaling equations give valid area/energy projections.
    Hardware numbers (Table 12, Table 14, Fig. 11) rely on these simulation models; no silicon validation.
  • domain assumption The terrestrial-neutron SER model (1000 FIT/Mbit) and lambda_SEU=1e-12/bit/hr are valid for the FIT/MTTF calculations.
    Table 19 relies on analytic SER models and assumed flux from references.
invented entities (2)
  • Three-tier UEP codec (Tier 1 SECDED / Tier 2 SEC / Tier 3 bypass) no independent evidence
    purpose: To provide graded error protection by bit significance, cutting ECC area vs uniform SECDED.
    A proposed hardware design; evaluated only through RTL synthesis and directed/random vector tests, not silicon or an external falsifiable prediction.
  • Dual-partition SRAM with per-cacheline format tag no independent evidence
    purpose: To store bypassed LSBs in a low-voltage partition while tags select protection mode per data type/workload.
    Architectural proposal; area/energy estimates from CACTI and analytical voltage-scaling equations, not measured hardware.

pith-pipeline@v1.3.0-alltime-deepseek · 27800 in / 26047 out tokens · 252379 ms · 2026-08-01T12:13:13.965443+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of From Bit-Position Sensitivity to Unequal Error Protection for DNN Inference Memory." pith.science (2026). https://pith.science/paper/YNS3ZPBP

@misc{pith2026260719623,
  author       = {Pith},
  title        = {Pith review of: From Bit-Position Sensitivity to Unequal Error Protection for DNN Inference Memory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YNS3ZPBP}},
  note         = {Machine review of arXiv:2607.19623}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We characterize per-bit-position fault sensitivity in ML inference across 16 workloads -- spanning transformer-based models and attention-free CNNs -- and across three floating-point formats. Our central empirical finding is a sharp bit-sensitivity transition: flipping any of the least-significant fraction bits up to a data-type-specific threshold, Xsafe, degrades task metrics by less than 1% under deterministic single-bit stress tests. Sensitivity rises through the upper fraction bits and spikes at the exponent-mantissa boundary, where a single-bit flip causes catastrophic collapse. Because low-order bits are largely inconsequential while high-order and exponent bits are critical, uniform SECDED protection -- which guards every bit equally at 12.5% storage overhead -- is unnecessarily conservative. We derive per-data-type Xsafe floors (FP16: 6, BF16: 4, FP32: 15) and workload-aware tiers that widen the unprotected region for resilient model classes, raising ECC savings to 37.5-62.5% without retraining. Text-conditioned diffusion models dictate the conservative floor; vision encoders, NLU models, and resilient LLMs tolerate wider bypass regions. These floors and tiers drive an Unequal Error Protection (UEP) codec with per-cacheline data-type tags and a dual-partition SRAM architecture for ML accelerators. Validation across 870+ fault-injection runs confirms selective protection holds under contiguous 2- and 3-bit upsets. The codec reduces ECC area by 27.8% relative to uniform SECDED; dual-voltage operation of the non-critical partition lowers gross BF16 read energy by about 17%, with a roughly 4% dual-partition macro-area overhead.

Figures

Figures reproduced from arXiv: 2607.19623 by Karthik Mohan Kumar, Keshavan Varadarajan, Kunal Tyagi, Muhammad Husnain Mubarik, Pedro Antonio Pena.

Figure 1
Figure 1. Figure 1: Per-model SSIM under deterministic single-bit-position stress tests [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Paired McNemar tests for single-bit flips at matrix-multiply outputs [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Attention-path Xsafe for five decoder LLMs plotted against their heads/KV ratio — the number of query heads sharing each key-value head pair — on a log scale, for FP16, BF16, and FP32. A higher heads/KV ratio (as in multi￾query attention) compresses key-value memory across more query heads and could in principle confer redundancy that masks bit errors. Spearman correlation is non-significant across all dat… view at source ↗
Figure 4
Figure 4. Figure 4: Permanent stuck-at-0 (solid lines) and stuck-at-1 (dashed lines) faults [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Aggregate SD3.5-Medium PSNR under random bit-flips as a function [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Left: BF16 and FP16 Xsafe under MCU-2 and MCU-3 compared to single-bit faults, showing a contraction of at most one bit per additional flipped bit. Right: cross-architecture BF16 MLP sensitivity-transition positions under MCU-2 and MCU-3, confirming that the contraction pattern is consistent across all evaluated models [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Matched-variance comparison across four noise families (Gauss [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 10
Figure 10. Figure 10: Mean HumanEval pass@1 drop (baseline minus injected), averaged [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 9
Figure 9. Figure 9: Tagged selective-ECC access path for one 128-bit granule: the per [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Net SRAM read-energy savings vs.𝑉low for BF16, FP16, and FP32. The adopted operating point 𝑉low=1.2 V (dashed line) yields positive net savings for all three formats after level-shifter overhead [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Design-space tradeoffs for the 𝑋-boundary sweep. Left: codec area across BF16, FP16, and FP32 as the protection boundary 𝑋 is varied. Right: relative BCH𝑡 = 2 and Hamming SECDED area for the representative protected payload widths used in this section. 6.2 MCU-Aware Design Selection BCH 𝑡=2 costs 3.9–4.2× Hamming SECDED [20, 24] area across tested codeword widths (64, 128, 256 bits). Once bit-plane inter￾… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

78 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, et al. 2024. Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone. arXiv:2404.14219 https://arxiv.org/abs/2404.14219

    Pith/arXiv arXiv 2024

  2. [2]

    Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebron, and Sumit Sanghai. 2023. GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 4895–4901. doi:10.18653/v1/2023.emnlp-main.298 12 From Bit-Position Sen...

    doi:10.18653/v1/2023.emnlp-main.298 2023

  3. [3]

    Irina Alam and Puneet Gupta. 2022. COMET: On-Die and In-Controller Collabora- tive Memory ECC Technique for Safer and Stronger Correction of DRAM Errors. InDSN. IEEE, Baltimore, MD, USA, 124–136. doi:10.1109/DSN53405.2022.00024

    arXiv 2022

  4. [4]

    Ebtesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, et al . 2023. The Falcon Series of Open Language Models. arXiv:2311.16867 https://arxiv.org/abs/ 2311.16867

    Pith/arXiv arXiv 2023

  5. [5]

    1999.Design and Analysis of Fast Low-Power SRAMs

    Bharadwaj S Amrutur. 1999.Design and Analysis of Fast Low-Power SRAMs. Stanford University

    1999

  6. [6]

    2019.Armv8.5-A Memory Tagging Extension White Paper

    Arm. 2019.Armv8.5-A Memory Tagging Extension White Paper. Arm. https://developer.arm.com/-/media/Arm%20Developer%20Community/PDF/ Arm_Memory_Tagging_Extension_Whitepaper.pdf

    2019

  7. [7]

    Kahng, Naveen Muralimanohar, Ali Shafiee, and Vaishnav Srinivas

    Rajeev Balasubramonian, Andrew B. Kahng, Naveen Muralimanohar, Ali Shafiee, and Vaishnav Srinivas. 2017. CACTI 7: New Tools for Interconnect Exploration in Innovative Off-Chip Memories.ACM Trans. Archit. Code Optim.14, 2 (2017), 14:1–14:25. doi:10.1145/3085572

    doi:10.1145/3085572 2017

  8. [8]

    Robert C. Baumann. 2005. Radiation-Induced Soft Errors in Advanced Semicon- ductor Technologies.IEEE Transactions on Device and Materials Reliability5, 3 (2005), 305–316. doi:10.1109/TDMR.2005.853449

    arXiv 2005

  9. [9]

    Izan Catalán, José Flich, and Carles Hernández. 2025. Exploiting Neural Networks Bit-Level Redundancy to Mitigate the Impact of Faults at Inference.The Journal of Supercomputing81 (2025), 183. doi:10.1007/s11227-024-06693-7

    work page doi:10.1007/s11227-024-06693-7 2025

  10. [10]

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhong- dao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. 2023. PixArt-𝛼: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthe- sis. arXiv:2310.00426 https://arxiv.org/abs/2310.00426

    Pith/arXiv arXiv 2023

  11. [11]

    C. J. Clopper and E. S. Pearson. 1934. The Use of Confidence or Fiducial Limits Illustrated in the Case of the Binomial.Biometrika26, 4 (1934), 404–413. https: //barestatistics.nl/uploads/1/1/7/9/11797954/clopper__pearson_1934.pdf

    1934

  12. [12]

    Tri Dao and Albert Gu. 2025. FlashAttention-4: Hardware-Friendly Fused At- tention with Warp-Specialization and Pingpong Scheduling.arXiv preprint arXiv:2603.05451(2025)

    arXiv 2025

  13. [13]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. InNAACL-HLT. https://arxiv.org/abs/1810.04805

    Pith/arXiv arXiv 2019

  14. [14]

    Jesse Dodge, Maarten Sap, Ana Marasović, William Agnew, Gabriel Ilharco, Dirk Groeneveld, Margaret Mitchell, and Matt Gardner. 2021. Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 1286–...

    2021

  15. [15]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. InICLR. https://arxiv.org/abs/2010.11929

    Pith/arXiv arXiv 2021

  16. [16]

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. 2024. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. InProceedings of the 41st International Conference on Machine Learning...

    2024

  17. [17]

    David Goldberg. 1991. What Every Computer Scientist Should Know About Floating-Point Arithmetic.Comput. Surveys23, 1 (1991), 5–48. doi:10.1145/ 103162.103163

    arXiv 1991

  18. [18]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, et al. 2024. The Llama 3 Herd of Models. arXiv:2407.21783 https://arxiv.org/abs/2407.21783

    Pith/arXiv arXiv 2024

  19. [19]

    Taeuk Ha, Gyuri Kim, Chanki Kim, and Sang-Hyo Kim. 2025. Chipkill-Level ECC Using 4-Bit Symbol Reed–Solomon Codes for DDR5 DRAM. InICTC. IEEE, Jeju, South Korea, 371–375. doi:10.1109/ICTC66702.2025.11387968

    arXiv 2025

  20. [20]

    Richard W. Hamming. 1950. Error Detecting and Error Correcting Codes.Bell System Technical Journal29, 2 (1950), 147–160. doi:10.1002/j.1538-7305.1950. tb00463.x

    arXiv 1950

  21. [21]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual Learning for Image Recognition. InCVPR. http://openaccess.thecvf.com/content_ cvpr_2016/html/He_Deep_Residual_Learning_CVPR_2016_paper.html

    2016

  22. [22]

    Sanghyun Hong, Pietro Frigo, Yigitcan Kaya, Cristiano Giuffrida, and Tudor Dumitras. 2019. Terminal Brain Damage: Exposing the Graceless Degradation in Deep Neural Networks Under Hardware Fault Attacks. InUSENIX Security Symposium. https://www.usenix.org/conference/usenixsecurity19/presentation/ hong

    2019

  23. [23]

    Le, and Hartwig Adam

    Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasude- van, Quoc V. Le, and Hartwig Adam. 2019. Searching for MobileNetV3. InICCV. http://openaccess.thecvf.com/content_ICCV_2019/html/Howard_ Searching_for_MobileNetV3_ICCV_2019_paper.html

    2019

  24. [24]

    M. Y. Hsiao. 1970. A Class of Optimal Minimum Odd-Weight-Column SEC- DED Codes.IBM Journal of Research and Development14, 4 (1970), 395–401. doi:10.1147/rd.144.0395

    work page doi:10.1147/rd.144.0395 1970

  25. [25]

    Ibe, Hitoshi Taniguchi, Yasuo Yahagi, Ken-ichi Shimbo, and Tadanobu Toba

    Eishi H. Ibe, Hitoshi Taniguchi, Yasuo Yahagi, Ken-ichi Shimbo, and Tadanobu Toba. 2010. Impact of Scaling on Neutron-Induced Soft Error in SRAMs From a 250 nm to a 22 nm Design Rule.IEEE Transactions on Electron Devices57, 7 (2010), 1527–1538. doi:10.1109/TED.2010.2047907

    arXiv 2010

  26. [26]

    2019.IEEE Standard for Floating-Point Arithmetic

    IEEE Computer Society. 2019.IEEE Standard for Floating-Point Arithmetic. Insti- tute of Electrical and Electronics Engineers, New York, NY, USA. doi:10.1109/ IEEESTD.2019.8766229 IEEE Std 754-2019 (Revision of IEEE 754-2008)

    arXiv 2019

  27. [27]

    Jiang, Alexandre Sablayrolles, Arthur Mensch, et al

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, et al. 2023. Mistral 7B. arXiv:2310.06825 https://arxiv.org/abs/2310.06825

    Pith/arXiv arXiv 2023

  28. [28]

    Vinay Joshi, Manuel Le Gallo, Simon Haefeli, Irem Boybat, S. R. Nandakumar, Christophe Piveteau, Martino Dazzi, Bipin Rajendran, Abu Sebastian, and Evan- gelos Eleftheriou. 2020. Accurate Deep Neural Network Inference Using Com- putational Phase-Change Memory.Nature Communications11 (2020), 2473. doi:10.1038/s41467-020-16108-9

    doi:10.1038/s41467-020-16108-9 2020

  29. [29]

    Sullivan, Seong-Lyong Gong, and Mattan Erez

    Jungrae Kim, Michael B. Sullivan, Seong-Lyong Gong, and Mattan Erez. 2015. Frugal ECC: Efficient and Versatile Memory Error Protection Through Fine- Grained Compression. InSC. ACM, Austin, TX, USA, 12:1–12:12. doi:10.1145/ 2807591.2807659

    arXiv 2015

  30. [30]

    Tae Hyun Kim, Hanwool Jeong, Juhyun Park, Hoonki Kim, Taejoong Song, and Seong-Ook Jung. 2020. An Embedded Level-Shifting Dual-Rail SRAM for High- Speed and Low-Power Cache.IEEE Access8 (2020), 187126–187139. doi:10.1109/ ACCESS.2020.3030099

    arXiv 2020

  31. [31]

    Giray Yağlıkçı, Roknoddin Azizi, Taha Shahroodi, Konstantinos Kanellopoulos, and Onur Mutlu

    Skanda Koppula, Lois Orosa, A. Giray Yağlıkçı, Roknoddin Azizi, Taha Shahroodi, Konstantinos Kanellopoulos, and Onur Mutlu. 2019. EDEN: Enabling Energy- Efficient, High-Performance Deep Neural Network Inference Using Approximate DRAM. InMICRO. ACM, Columbus, OH, USA, 218–231. doi:10.1145/3352460. 3358280

    doi:10.1145/3352460 2019

  32. [32]

    2023.Unequal Error Protection Source–Channel Coding

    Andres Kwasinski and Vinay Chande. 2023.Unequal Error Protection Source–Channel Coding. 135–172. doi:10.1002/9781118693803.ch5

    work page doi:10.1002/9781118693803.ch5 2023

  33. [33]

    Young Seo Lee, Gunjae Koo, Young-Ho Gong, and Sung Woo Chung. 2022. Stealth ECC: A Data-Width Aware Adaptive ECC Scheme for DRAM Error Resilience. In DATE. IEEE, Antwerp, Belgium, 382–387. doi:10.23919/DATE54114.2022.9774775

    arXiv 2022

  34. [34]

    Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi

  35. [35]

    Guanpeng Li, Siva Kumar Sastry Hari, Michael Sullivan, Timothy Tsai, Karthik Pattabiraman, Joel Emer, and Stephen W. Keckler. 2017. Understanding Error Propagation in Deep Learning Neural Network (DNN) Accelerators and Applica- tions. InSC. Denver, CO, USA. doi:10.1145/3126908.3126964

    arXiv 2017

  36. [36]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. InICCV. https://arxiv.org/abs/2103.14030

    Pith/arXiv arXiv 2021

  37. [37]

    2014.Tagged Memory and Minion Cores in the lowRISC SoC

    lowRISC. 2014.Tagged Memory and Minion Cores in the lowRISC SoC. Technical Report memo-2014-001. lowRISC. https://lowrisc.org/docs/memo-2014-001- tagged-memory-and-minion-cores/

    2014

  38. [38]

    Mathew, Martin Schultheis, Carl C

    Deepak M. Mathew, Martin Schultheis, Carl C. Rheinländer, Chirag Sudarshan, Christian Weis, Norbert Wehn, and Matthias Jung. 2018. An Analysis on Reten- tion Error Behavior and Power Consumption of Recent DDR4 DRAMs. InDATE. IEEE, Dresden, Germany, 293–296. doi:10.23919/DATE.2018.8342023

    arXiv 2018

  39. [39]

    Quinn McNemar. 1947. Note on the Sampling Error of the Difference Between Correlated Proportions or Percentages.Psychometrika12, 2 (1947), 153–157. doi:10.1007/BF02295996

    doi:10.1007/bf02295996 1947

  40. [40]

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2017. Pointer Sentinel Mixture Models. InICLR. https://openreview.net/forum?id= Byj72udxe&noteId=Byj72udxe

    2017

  41. [41]

    Justin Meza, Qiang Wu, Sanjeev Kumar, and Onur Mutlu. 2015. Revisiting Mem- ory Errors in Large-Scale Production Data Centers: Analysis and Modeling of New Trends from the Field. InDSN. https://research.facebook.com/publications/ revisiting-memory-errors-in-large-scale-production-data-centers-analysis- and-modeling-of-new-trends-from-the-field/

    2015

  42. [42]

    Vishesh Mishra, Marcello Traiola, Angeliki Kritikakou, Olivier Sentieys, and Urbi Chatterjee. 2025. SERA-Float: A Soft Error Resilient Approximate Floating-Point Computing Format. InICCAD. IEEE, Munich, Germany, 1–9. https://inria.hal. science/hal-05333255v1/file/SERA_FLOAT_v2.pdf

    2025

  43. [43]

    Modal Team. 2025. Reverse-Engineering FlashAttention-4. https://modal.com/ blog/reverse-engineer-flash-attention-4

    2025

  44. [44]

    Sanches, Kohji Hosokawa, Scott C

    Pritish Narayanan, Alessandro Fumarola, Lucas L. Sanches, Kohji Hosokawa, Scott C. Lewis, Robert M. Shelby, and Geoffrey W. Burr. 2017. To- ward On-Chip Acceleration of the Backpropagation Algorithm Using Non- volatile Memory.IBM Journal of Research and Development61, 4/5 (2017). https://research.ibm.com/publications/toward-on-chip-acceleration- of-the-ba...

    2017

  45. [45]

    Duy Thanh Nguyen, Hyun Kim, Hyuk-Jae Lee, and Ik Joon Chang. 2018. An Approximate Memory Architecture for a Reduction of Refresh Power Con- sumption in Deep Learning Applications. InISCAS. IEEE, Florence, Italy, 1–5. doi:10.1109/ISCAS.2018.8351021

    arXiv 2018

  46. [46]

    Dreslinski, and Trevor Mudge

    Byoungchan Oh, Nilmini Abeyratne, Jeongseob Ahn, Ronald G. Dreslinski, and Trevor Mudge. 2016. Enhancing DRAM Self-Refresh for Idle Power Reduction. 13 Muhammad Husnain Mubarik, Karthik Mohan Kumar, Pedro Antonio Pena, Keshavan Varadarajan, Kunal Tyagi InISLPED. ACM, San Francisco, CA, USA, 254–259. doi:10.1145/2934583.2934632

    arXiv 2016

  47. [47]

    Palframan, Nam Sung Kim, and Mikko H

    David J. Palframan, Nam Sung Kim, and Mikko H. Lipasti. 2014. Precision- Aware Soft Error Protection for GPUs. InHPCA. IEEE, Orlando, FL, USA, 49–59. doi:10.1109/HPCA.2014.6835966

    arXiv 2014

  48. [48]

    Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. LibriSpeech: An ASR Corpus Based on Public Domain Audio Books. InICASSP. https://us.openslr.org/resources/12/about.html

    2015

  49. [49]

    William Peebles and Saining Xie. 2023. Scalable Diffusion Models with Trans- formers. InICCV. https://www.wpeebles.com/DiT.html

    2023

  50. [50]

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. arXiv:2307.01952 https: //arxiv.org/abs/2307.01952

    Pith/arXiv arXiv 2023

  51. [51]

    Qureshi, Dae-Hyun Kim, Samira Manabi Khan, Prashant J

    Moinuddin K. Qureshi, Dae-Hyun Kim, Samira Manabi Khan, Prashant J. Nair, and Onur Mutlu. 2015. AVATAR: A Variable-Retention-Time (VRT) Aware Refresh for DRAM Systems. InDSN. IEEE, Rio de Janeiro, Brazil, 427–437. doi:10.1109/ DSN.2015.58

    2015

  52. [52]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust Speech Recognition via Large-Scale Weak Supervision. arXiv:2212.04356 https://arxiv.org/abs/2212.04356

    Pith/arXiv arXiv 2022

  53. [53]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language Models are Unsupervised Multitask Learners. OpenAI technical report. https://cdn.openai.com/better-language-models/language_ models_are_unsupervised_multitask_learners.pdf

    2019

  54. [54]

    Brandon Reagen, Udit Gupta, Lillian Pentecost, Paul Whatmough, Sae Kyu Lee, Niamh Mulholland, David Brooks, and Gu-Yeon Wei. 2018. Ares: A Framework for Quantifying the Resilience of Deep Neural Networks. InDAC. ACM, San Francisco, CA, USA, 17:1–17:6. doi:10.1145/3195970.3195997

    arXiv 2018

  55. [55]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. InCVPR. https://arxiv.org/abs/2112.10752

    Pith/arXiv arXiv 2022

  56. [56]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexan- der C. Berg, and Li Fei-Fei. 2015. ImageNet Large Scale Visual Recognition Challenge.International Journal of Computer Vision115, 3 (2015), 211–252. doi:10.1007/s11263-015-0816-y

    doi:10.1007/s11263-015-0816-y 2015

  57. [57]

    Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. 2024. Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillation. arXiv:2403.12015 https://arxiv.org/abs/2403. 12015

    Pith/arXiv arXiv 2024

  58. [58]

    Bianca Schroeder, Eduardo Pinheiro, and Wolf-Dietrich Weber. 2009. DRAM Errors in the Wild: A Large-Scale Field Study. InSIGMETRICS. https://research. google/pubs/dram-errors-in-the-wild-a-large-scale-field-study/

    2009

  59. [59]

    Noam Shazeer. 2019. Fast Transformer Decoding: One Write-Head is All You Need. arXiv:1911.02150 https://arxiv.org/abs/1911.02150

    Pith/arXiv arXiv 2019

  60. [60]

    Manning, Andrew Ng, and Christopher Potts

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive Deep Models for Seman- tic Compositionality Over a Sentiment Treebank. InEMNLP. Association for Computational Linguistics, 1631–1642. https://aclanthology.org/D13-1170/

    2013

  61. [61]

    Vilas Sridharan, Nathan Debardeleben, Sean Blanchard, Kurt Ferreira, Sudhanva Gurumurthi, and John Shalf. 2015. Memory Errors in Modern Systems: The Good, The Bad, and The Ugly. InASPLOS. doi:10.1145/2694344.2694348

    arXiv 2015

  62. [62]

    Aaron Stillmaker and Bevan M. Baas. 2017. Scaling Equations for the Accu- rate Prediction of CMOS Device Performance from 180 nm to 7 nm.Integra- tion, the VLSI Journal58 (2017), 74–81. http://vcl.ece.ucdavis.edu/pubs/2017.02. VLSIintegration.TechScale/

    2017

  63. [63]

    David Stutz, Nandhini Chandramoorthy, Matthias Hein, and Bernt Schiele. 2021. Bit Error Robustness for Energy-Efficient DNN Accelerators. InProceedings of MLSys. MLSys, Virtual, 1–21. https://proceedings.mlsys.org/paper_files/paper/ 2021/file/2f9b1b6b29361118f630783c19891ea0-Paper.pdf

    2021

  64. [64]

    Sullivan, Mohamed Tarek Ibn Ziad, Aamer Jaleel, and Stephen W

    Michael B. Sullivan, Mohamed Tarek Ibn Ziad, Aamer Jaleel, and Stephen W. Keckler. 2023. Implicit Memory Tagging: No-Overhead Memory Safety Using Alias-Free Tagged ECC. InISCA. ACM, Orlando, FL, USA, 1–14. doi:10.1145/ 3579371.3589079

    arXiv 2023

  65. [65]

    Sullivan, Nirmal Saxena, Mike O’Connor, Donghyuk Lee, Paul Racu- nas, Saurabh Hukerikar, Timothy Tsai, Siva Hari, and Stephen W

    Michael B. Sullivan, Nirmal Saxena, Mike O’Connor, Donghyuk Lee, Paul Racu- nas, Saurabh Hukerikar, Timothy Tsai, Siva Hari, and Stephen W. Keckler. 2021. Characterizing and Mitigating Soft Errors in GPU DRAM. InMICRO. ACM. doi:10.1145/3466752.3480111

    arXiv 2021

  66. [66]

    Mingxing Tan and Quoc V. Le. 2019. EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. InICML. https://research.google/pubs/ efficientnet-rethinking-model-scaling-for-convolutional-neural-networks/

    2019

  67. [67]

    Hoyoung Tang and Jongsun Park. 2016. Unequal-Error-Protection Error Cor- rection Codes for the Embedded Memories in Digital Signal Processors.IEEE Transactions on Very Large Scale Integration (VLSI) Systems24, 6 (2016), 2397–2401. doi:10.1109/TVLSI.2015.2497368

    arXiv 2016

  68. [68]

    Kim, Juan Gómez Luna, Mohammad Sadrosadati, Nika Mansouri Ghiasi, and Onur Mutlu

    Yaohua Wang, Lois Orosa, Xiangjun Peng, Yang Guo, Saugata Ghose, Minesh Patel, Jeremie S. Kim, Juan Gómez Luna, Mohammad Sadrosadati, Nika Mansouri Ghiasi, and Onur Mutlu. 2020. FIGARO: Improving System Performance via Fine-Grained In-DRAM Data Relocation and Caching. InMICRO. IEEE, Virtual (originally Athens, Greece), 235–249. https://ieeexplore.ieee.org...

    2020

  69. [69]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. 2004. Image Quality Assessment: From Error Visibility to Structural Similarity.IEEE Transactions on Image Processing13, 4 (2004), 600–612. doi:10.1109/TIP.2003. 819861

    doi:10.1109/tip.2003 2004

  70. [70]

    Frank Wilcoxon. 1945. Individual Comparisons by Ranking Methods.Biomet- rics Bulletin1, 6 (1945), 80–83. https://sci2s.ugr.es/keel/pdf/algorithm/articulo/ wilcoxon1945.pdf

    1945

  71. [71]

    Rui Xie, Yunhua Fang, Asad Ul Haq, Linsen Ma, Sanchari Sen, Swagath Venkatara- mani, Liu Liu, and Tong Zhang. 2025. Making Strong Error-Correcting Codes Work Effectively for HBM in AI Inference. arXiv:2512.18152 doi:10.48550/arXiv. 2512.18152

    doi:10.48550/arxiv.2512.18152 2025

  72. [72]

    Rui Xie, Asad Ul Haq, Yunhua Fang, Linsen Ma, Sanchari Sen, Swagath Venkatara- mani, Liu Liu, and Tong Zhang. 2025. Breaking the HBM Bit Cost Bar- rier: Domain-Specific ECC for AI Inference Infrastructure. arXiv:2507.02654 doi:10.48550/arXiv.2507.02654

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2507.02654 2025

  73. [73]

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a Machine Really Finish Your Sentence?. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, 4791–4800. doi:10.18653/v1/P19-1472

    doi:10.18653/v1/p19-1472 2019

  74. [74]

    Jeff Zhang, Kartheek Rangineni, Zahra Ghodsi, and Siddharth Garg. 2018. ThUn- derVolt: Enabling Aggressive Voltage Underscaling and Timing Error Resilience for Energy Efficient Deep Learning Accelerators. InDAC. ACM, San Francisco, CA, USA, 19:1–19:6. doi:10.1145/3195970.3196129

    arXiv 2018

  75. [75]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang

  76. [76]

    Ruohuang Zheng and Michael C. Huang. 2017. Redundant Memory Array Architecture for Efficient Selective Protection. InISCA. ACM, Toronto, ON, Canada, 214–227. doi:10.1145/3079856.3080213 14

    arXiv 2017

  77. [2018]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. InCVPR. http://richzhang.github.io/PerceptualSimilarity/

  78. [2024]

    arXiv:2402.17245 https://arxiv.org/abs/2402.17245

    Playground v2.5: Three Insights towards Enhancing Aesthetic Quality in Text-to-Image Generation. arXiv:2402.17245 https://arxiv.org/abs/2402.17245

    Pith/arXiv arXiv