Optimized Image Classification on the Cheap
summary
Stanford Cars (≈16k images, 196 classes) compares two transfer methods: ResNet18 fine-tune versus feature extraction from ResNet50. Hyperparameter search uses multitask Bayesian optimization (SigOpt) that samples partial trainings at 10/50/100% of 35 epochs. Fine-tuning the shallow network outperforms deep feature extraction; optimization yields ~+4% for fine-tune versus ~+2% for feature-extract, with an optimized fine-tuned baseline near 87% accuracy. Doubling data to 32k via color-jitter augmentation (Bayesian-tuned) further reduces background-clutter errors and boosts final accuracy substantially.
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:55Reduce training time: Transfer Learning
- 1:37Boost Performance: Image Augmentation
- 1:58Reduce tuning time: Multitask Optimization
- 3:53Data: Stanford Cars
- 4:43Tuning: Multitask Optimization
- 5:15Experiment scenarios
- 5:59Hyperparameter Tuning Tuning SGD
- 6:10The optimization loop
- 7:47Insight: Multitask efficiency at the hyperparameter level
- 8:41What did misclassifications look like?
- 10:04Generating synthetic Images with Multitask Optimization
- 10:54Image augmentation process
- 11:50Experimental designs
- 12:16Baseline model performance
- 13:31Hyperparameter Tuning: SGD and Augmentation
- 13:45Results: a 10% boost!
- 14:13What did the augmented images look like?
- 16:01Transfer Leaming Results
- 16:21Augmentation Results
related talks
52mTransfer Learning and Finetuning Deep Convolution Neural Network
01h 31mPyTorch Workflow Mastery: A Guide to Track and Optimize Model Performance
28m🧼 From GPU-poor to data-rich: data quality practices for LLM fine-tuning
40mPutting Deep Hybrid Learning to Work
39mModel Fine-tuning For Search: From Algorithms to Infra
02h 26mPractical Deep Learning for Data Scientists
30mTransfer learning for image recognition in healthcare...
28mObject Detection using Transformers and CNNs