Decoupling Representation and Classifier for Long-Tailed Recognition
read the original abstract
The long-tail distribution of the visual world poses great challenges for deep learning based classification models on how to handle the class imbalance problem. Existing solutions usually involve class-balancing strategies, e.g., by loss re-weighting, data re-sampling, or transfer learning from head- to tail-classes, but most of them adhere to the scheme of jointly learning representations and classifiers. In this work, we decouple the learning procedure into representation learning and classification, and systematically explore how different balancing strategies affect them for long-tailed recognition. The findings are surprising: (1) data imbalance might not be an issue in learning high-quality representations; (2) with representations learned with the simplest instance-balanced (natural) sampling, it is also possible to achieve strong long-tailed recognition ability by adjusting only the classifier. We conduct extensive experiments and set new state-of-the-art performance on common long-tailed benchmarks like ImageNet-LT, Places-LT and iNaturalist, showing that it is possible to outperform carefully designed losses, sampling strategies, even complex modules with memory, by using a straightforward approach that decouples representation and classification. Our code is available at https://github.com/facebookresearch/classifier-balancing.
This paper has not been read by Pith yet.
Forward citations
Cited by 20 Pith papers
-
VISA: VLM-Guided Instance Semantic Auditing for 3D Occupancy World Models
VISA improves closed-set 3D occupancy mIoU on nuScenes by using VLM instance audits as reliability-weighted semantic supervisors during training of existing world models.
-
SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents
SkillRet benchmark shows fine-tuned retrievers improve NDCG@10 by 13+ points over prior models on large-scale skill retrieval for LLM agents.
-
Class-frequency Guided Noise Schedule for Diffusion Models
Proposes CFRG noise schedule for diffusion models that assigns larger noises to low-frequency classes to improve generation on imbalanced datasets.
-
Learning to Annotate Delayed and False AEB Events: A Practical System for Extreme Class Imbalance and Asymmetric Label Noise
An automated AEB annotation framework uses data augmentation and noise suppression to achieve 80% recall improvement and 50% workload reduction for rare delayed/false triggers under class imbalance and asymmetric label noise.
-
Toward Multi-Domain and Long-Tailed Quantization via Feature Alignment and Scaling
Proposes EmaQ and EmaQ-LT methods for multi-domain and long-tailed DNN quantization with CDF alignment, sensitivity aggregation, class-conditioned scaling, and convergence guarantees, showing strong low-bit results on...
-
Why Not Hyperparameter-Friendly Optimisation? A Monotonic Adaptive Norm Rescaling Approach For Long-Tailed Recognition
SAMN applies the Pool Adjacent Violators Algorithm to directly enforce monotonicity on per-class weight norms during classifier retraining, eliminating hyperparameters while integrating with other long-tailed methods.
-
Rethinking Loss Reweighting for Imbalance Learning as an Inverse Problem: A Neural Collapse Point of View
Loss reweighting is cast as an inverse problem that dynamically infers class weights to equalize per-class average losses under the Neural Collapse simplex ETF target.
-
Synthetic Data Generation for Long-Tail Medical Image Classification: A Case Study in Skin Lesions
A diffusion-based synthetic data pipeline using inpainting and OOD post-selection improves long-tail skin lesion classification on ISIC2019, delivering over 28% accuracy gain on the rarest class.
-
AdamO: A Collapse-Suppressed Optimizer for Offline RL
AdamO modifies Adam with an orthogonality correction to ensure the spectral radius of the TD update operator stays below one, providing a theoretical stability guarantee for offline RL.
-
Syntax as a Rosetta Stone: Universal Dependencies for In-Context Coptic Translation
Combining dictionary glosses with Universal Dependencies syntactic information in in-context learning produces new state-of-the-art Coptic-English translation results across model sizes.
-
CXR-LT 2026 Challenge: Multi-Center Long-Tailed and Zero Shot Chest X-ray Classification
CXR-LT 2026 introduces a radiologist-annotated multi-center dataset of 145k+ CXRs to benchmark robust multi-label classification on known classes and open-world generalization to unseen rare diseases.
-
LoFT: Parameter-Efficient Fine-Tuning for Long-tailed Semi-Supervised Learning in Open-World Scenarios
LoFT uses parameter-efficient fine-tuning of foundation models for long-tailed semi-supervised learning, supported by proofs that this reduces hypothesis complexity to minimize balanced posterior error and compresses ...
-
Prior-Anchored Debiasing for Long-Tailed Multi-Organ Pathology Report Generation
PriOrGen uses prior-anchored modules to debias visual encoding and textual decoding for improved long-tailed multi-organ pathology report generation.
-
Prior-Anchored Debiasing for Long-Tailed Multi-Organ Pathology Report Generation
PriOrGen uses a Visual-Prototype Anchored Bottleneck and Meta-Report Anchored Bank to mitigate visual and textual biases in long-tailed multi-organ pathology report generation.
-
Personalized Generative Models for Contextual Debiasing
DecoupleGen personalizes diffusion models to create images with uncommon contexts for debiasing object recognition, yielding consistent gains on scene classification tasks.
-
SemLT3D: Semantic-Guided Expert Distillation for Camera-only Long-Tailed 3D Object Detection
SemLT3D introduces semantic-guided expert distillation with a language MoE module and CLIP projection to enrich features for long-tailed classes in camera-only 3D detection.
-
Enhancing Oracle Bone Inscription Recognition via Multi-Scale Layer Attention
MSLA is a new attention mechanism that models multi-scale and cross-layer interactions to achieve more accurate OBI recognition than prior attention methods.
-
Proposal Refinement for Few-Shot Object Detection
Adds refinement loss and auxiliary RPN branch to rebalance novel-class proposals, yielding 1-6% gains on few-shot detection benchmarks.
-
Syntax as a Rosetta Stone: Universal Dependencies for In-Context Coptic Translation
HAMR combines bi-level meta-learned instance reweighting with KNN-based neighborhood resampling to improve NLP performance under class imbalance, evaluated on six NER and classification benchmarks.
-
Novel Dynamic Batch-Sensitive Adam Optimiser for Vehicular Accident Injury Severity Prediction
DBS-Adam, which scales learning rates by batch difficulty from EMA gradient norms and loss, reaches 95.22% accuracy on Bi-LSTM accident severity prediction and shows statistically significant precision gains over AMSG...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.