Research map / Large language models / AI and society
Defog Artificial Intelligence Glasses: Neural Networks for the Imperfect Real World
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a short doctoral-style position paper on an image-based attention mechanism that fuses four pretrained CNN feature extractors via attention weights and metric learning to retrieve clothing images from noisy web data.
Problem
Neural networks degrade on real-world noisy and low-quality image data, and existing methods discard low-quality samples rather than using all available data.
Approach
Use an attention mechanism over four isolated pretrained models (ResNet50, VGG16, EfficientNet-B3, DenseNet121) that weights their feature vectors and learns metric embeddings so the network can extract useful knowledge from low-quality images; the paper also analyzes that the learned model weights are largely redundant.
Key results
- The attention-based fusion reportedly outperforms isolated models on clothing image retrieval, but no numeric mAP@10 values are given in the text.
- Average attention weights were about 0.43 for VGG16 and 0.25 for EfficientNet-B3, with ResNet50 at 0.12 and DenseNet121 at 0.20, suggesting the model contributions are repetitive/redundant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use an attention mechanism combining four models | clothes image retrieval · clothes image retrieval | image retrieval results | Helped |
| Combine EfficientNet-B3 and DenseNet121 with ResNet50 | image-based attention mechanism · clothes image retrieval | feature redundancy | Mixed |
Surprising findings
- Learned attention weights are concentrated on VGG16 and EfficientNet-B3 and are nearly uniform across clothing categories, which the author reads as redundancy among the fused models.
Limitations
- No quantitative comparison or numeric results are reported for the main method against baselines.
- Evaluation is limited to a single image retrieval task on a small, self-collected noisy dataset.
- Broader claims about generalization across tasks and benefits for resource-limited countries are not empirically supported.
Experimental setup
Datasets: clothing images collected from the internet (noisy, in the wild)
Benchmarks: internal clothing image retrieval dataset (noisy, web-collected)
Metrics: mAP@10
Baselines: isolated models (ResNet50, VGG16, EfficientNet-B3, DenseNet121)
Closest papers by meaning
- To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets (ICLR 2024 · similarity 0.83)
- Neural Collapse is Globally Optimal in Deep Regularized ResNets and Transformers (NeurIPS 2025 · similarity 0.83)
- MixEval-X: Any-to-any Evaluations from Real-world Data Mixture (ICLR 2025 · similarity 0.83)
- Deep Networks Always Grok and Here is Why (ICML 2024 · similarity 0.83)
- Algorithm Development in Neural Networks: Insights from the Streaming Parity Task (ICML 2025 · similarity 0.82)
- Do We Always Need the Simplicity Bias? Looking for Optimal Inductive Biases in the Wild (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.