Where Does the Semantic Gain Come From? A Reproduction and Extension of Semantic Knowledge-driven Contrastive Learning for Long-Tailed Recognition

Sushrut Ghimire

cs.CVOct 2, 2026

Where Does the Semantic Gain Come From? A Reproduction and Extension of Semantic Knowledge-driven Contrastive Learning for Long-Tailed Recognition

Authors: Sushrut Ghimire

Organizations: Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Germany

Abstract

Semantic Knowledge-driven Contrastive Learning (SKCL) uses a language model to decide which classes are related, and pulls each image towards the prototypes of its semantic neighbours. On CIFAR-100-LT (beta = 100) it reports 54.02% top-1 accuracy, 2.01 points above Balanced Contrastive Learning (BCL), the method it builds on. The code and the class descriptions are not public. I reimplement SKCL, BCL and ConCutMix in one framework, check it against the public baseline code, and run every configuration with three seeds. The two baselines reproduce within 1.5 points, but SKCL built on BCL, as the paper describes it, ends up 0.56 points below BCL. To find out why, I add SKCL to the authors' own ConCutMix code. Trained for the paper's 300 epochs, it reaches 53.69, only 0.33 below the published number. At the same budget, however, the semantic graph adds just 0.23 points over ConCutMix, while training ConCutMix for 100 more epochs adds 1.04. Together with ConCutMix's published lead over BCL (1.15), this explains the claimed gain. A BCL model that never sees the graph already shares 41.2% of the graph's top-2 neighbours with its own most-confused classes (2.0% by chance), which shows why the graph adds so little on these benchmarks. I also test several changes to SKCL. Combining it with the CutMix branch improves it by 1.06 points, and an adaptive version of the graph improves it slightly (+0.34 and +0.28 in two codebases), although these gains are within seed noise.

Figures & tables

	CIFAR-100-LT		CIFAR-10-LT
Method	Ours	Published	Ours	Published
BCL	$50.85\pm 0.81$	52.01	$83.53\pm 0.23$	84.31
ConCutMix	$51.71\pm 0.80$	53.16	$83.84\pm 0.41$	86.07
SKCL	$50.29\pm 1.51$	54.02	$83.19\pm 0.10$	87.50

Table 1: Top-1 accuracy (%),

\beta=100

, ResNet-32, my framework; mean

\pm

SD over three seeds of the last-10-epoch mean. Published values from [ 1 ] .

Figure 1: Per-seed results (open circles), mean

\pm

SD (bars) and published values (diamonds). Numbers give ours minus published. SKCL + adapt. is the adaptive graph of Section 4.5 .

Figure 2: SKCL variants added to the official ConCutMix code, CIFAR-100-LT,

200

epochs, best epoch. Filled markers: mean of two seeds; open markers: single seeds; shaded band: the two ConCutMix seeds.

	Best epoch	Last-10 mean
ConCutMix, 200 epochs (2 seeds)	$52.42\pm 0.30$	$52.20\pm 0.27$
ConCutMix, 300 epochs	$53.46\pm 0.38$	$53.23\pm 0.38$
ConCutMix + SKCL, 300 epochs	$53.69\pm 0.25$	$53.44\pm 0.29$
SKCL, published (300 epochs)	$54.02$

Table 2: ConCutMix with and without SKCL on the official ConCutMix code, CIFAR-100-LT, three seeds. Top-1 accuracy (%).

Step	Source	Gain
BCL $\rightarrow$ ConCutMix	published ( $52.01\rightarrow 53.16$ )	$+1.15$
$200\rightarrow 300$ epochs	measured, official code ( $52.42\rightarrow 53.46$ )	$+1.04$
+ semantic graph	measured, official code ( $53.46\rightarrow 53.69$ )	$+0.23$
Sum		$+2.42$
Claimed by SKCL	published ( $52.01\rightarrow 54.02$ )	$+2.01$

Table 3: Breakdown of SKCL’s claimed gain over BCL on CIFAR-100-LT (best epoch, as in the original papers).

Figure 3: Best-epoch accuracy against training epochs on CIFAR-100-LT. Open markers: published values (official ConCutMix logs at

200

and

400

epochs, SKCL paper at

300

). Filled markers: my runs of the official code.

Figure 4: Effect of the graph and of interventions on it, CIFAR-100-LT (and CIFAR-10-LT where marked), mean

\pm

1 SE of the difference over three seeds. The shaded band marks

\pm 1

point of seed noise; the red marker is the published claim.

Change	Setting	Reference $\rightarrow$ changed	Gain
+ CutMix branch	own, C100	$50.29\rightarrow 51.35$	$+1.06$
+ CutMix branch	own, C10	$83.19\rightarrow 84.13$	$+0.94$
paper’s member sets	own, C100	$50.29\rightarrow 50.92$	$+0.63$
adaptive graph	own, C100	$50.29\rightarrow 50.63$	$+0.34$
adaptive graph	official, C100	$52.10\rightarrow 52.38$	$+0.28$
graph weight $0.5$ instead of $1$	official, C100	$52.10\rightarrow 52.69$	$+0.59$

Table 4: Changes that improve SKCL. Own: my framework, last-10 mean, three seeds. Official: official ConCutMix code, best epoch,

200

epochs, two seeds.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Run	Data	Last-10	Best	Many	Med.	Few
FIX_adapt_c100_s0	C100	51.34	51.60	68.77	51.54	30.78
FIX_adapt_c100_s1	C100	48.83	49.09	65.95	49.25	28.37
FIX_adapt_c100_s2	C100	51.72	51.84	66.65	52.05	33.92
FIX_adapt_c10_s0	C10	83.14	83.72	85.03	75.56	—
FIX_adapt_c10_s1	C10	82.70	83.14	84.64	74.95	—
FIX_adapt_c10_s2	C10	83.25	83.98	84.87	76.79	—

Appendix

Table 5: Per-seed results of all corrected-pipeline runs,

\beta=100

. Last-10: mean top-1 over the final ten epochs (the reported metric); Best: maximum top-1 over all epochs (shown for comparison only); Many, Med. (Medium) and Few are last-10 group accuracies. C100/C10: CIFAR-100-LT/CIFAR-10-LT, which has no Few classes at

\beta=100

. Run-name prefixes: bcl = BCL, ccm = ConCutMix, ccm_g0 = ConCutMix with

\gamma=0

, skcl = SKCL, adapt = SKCL + adaptive graph, cutmixcombo(_nowt) = frequency-gated (binary) SKCL graph + CutMix, full/surprising/redundant/random = graph-filtering variants; PF_joint, PF_nocc, PF_graph, PF_full = paper-faithful paper-text variants: joint member sets, no class-complement, both, both + CutMix branch; s

n

= seed

n

Run	Best	Last-10	A	B	Many	Med.	Few
OCCM200_s0	52.63	52.39	52.69	52.08	64.91	49.91	40.68
OCCM200_s1	52.21	52.01	52.19	51.84	64.45	50.22	39.60
SK_full05_200_s0	52.74	52.61	52.32	52.90	64.68	50.29	41.25
SK_full05_200_s1	52.64	52.47	52.22	52.71	65.60	49.27	40.87
SK_coarse_200_s0	52.90	52.54	52.84	52.25	65.14	49.51	41.39
SK_coarse_200_s1	52.34	52.17	52.90	51.44	64.53	50.12	40.13

Appendix

Table 6: Per-seed results on the official ConCutMix code (official ConCutMix code, CIFAR-100-LT,

\beta=100

). Best: best-epoch top-1 on the full test set; Last-10: last-10 mean; A/B: last-10 mean on the two test halves; Many, Med., Few: last-10 group accuracies. OCCM = ConCutMix; SK_ = ConCutMix + SKCL variant (full: coarse head, joint member sets,

w=1

; full05:

w=0.5

; coarse: coarse head only; fine: fine-level Term 1; adapt: full + adaptive graph; ls: graph label smoothing; all: full + ls); 200/300: epochs; s

n

: seed.

Contrastive Learning

Contrastive learning trains representations by bringing related examples together and separating unrelated ones for recognition, retrieval, and transfer.

See all 318 papers

Bayesian Gated Non-Negative Contrastive Learning
May 27, 2026Peng Cui, Jiahao Zhang, Lijie HuContrastive Learning Disentangled Representation Learning
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI).
Contrastive vision-language learning with paraphrasing and negation
Nov 20, 2025Kwun Ho Ngan, Saman Sadeghi Afgeh, Joe Townsend +1VLM Robustness VLM Reasoning
Fujitsu Research of Europe · United Kingdom · City St George’s, University of London
Contrastive Semantic Projection: Faithful Neuron Labeling with Contrastive Examples
Apr 24, 2026Oussama Bouanani, Jim Berend, Wojciech Samek +2Neural Network Interpretability
Fraunhofer Heinrich Hertz Institute, Berlin, Germany · Technische Universität Berlin, Berlin, Germany · BIFOLD – Berlin Institute for the Foundations of Learning and Data, Berlin, Germany +1

May 27, 2026J/K move · Enter open · S save

Bayesian Gated Non-Negative Contrastive Learning

Peng Cui, Jiahao Zhang, Lijie Hu

Mohamed bin Zayed University of Artificial Intelligence (MBZUAI).

Contrastive Learning Disentangled Representation Learning Feature Selection Self-Supervised Visual Representation Learning

Open paper AskPDF arXiv

While Contrastive Learning (CL) has revolutionized self-supervised representation learning, its latent representations remain highly entangled and opaque, limiting their interpretability in safety-critical applications. We identify that a fundamental cause of this entanglement is the reliance on deterministic similarity measures, which treat all feature dimensions equally. In compositional scenes, this creates an Optimization Conflict: common background features, such as, "blue sky", are encouraged to align in positive pairs but simultaneously repelled in negative pairs, causing gradient oscillations that hinder precise semantic disentanglement. To address this, we propose BayesNCL (Bayesian Gated Non-Negative Contrastive Learning). Unlike standard approaches, BayesNCL introduces a probabilistic gating mechanism that dynamically filters out task-irrelevant, high-frequency common features while selectively retaining discriminative semantics. By formalizing feature selection as a variational inference problem with a sparse Bernoulli prior, our method effectively resolves the optimization conflict. Empirical experimental results on Imagenet-100 demonstrate that BayesNCL achieves a remarkable 142.1% improvement in semantic consistency compared to state-of-the-art baselines, yielding highly interpretable representations without compromising downstream task performance. Code is available at https://github.com/Cui-Peng-624/BayesNCL.

Abstract

Contrastive Learning

Explore similar work

Bayesian Gated Non-Negative Contrastive Learning

Contrastive vision-language learning with paraphrasing and negation

Contrastive Semantic Projection: Faithful Neuron Labeling with Contrastive Examples