Cifar-10 Exploratory Data Analysis
summary
Claims that CIFAR-10 can be trimmed without hurting accuracy, citing both a paper that finds 10% subsets retain performance and results from semantic clustering that report up to 55% reduction. Methods include agglomerative clustering on ResNet feature vectors, PCA/t‑SNE, hash filters (wavelet and average hash) to drop identical frames, and active‑learning queries (least confidence, margin, entropy). Evaluation runs on VGG16 with fixed hyperparameters and reports comparable test accuracy after the pruning.
AI-written summary · about
chapters (8)
- 0:00Cifar-10 data analysis intro
- 1:15Reduce datasets for AI training
- 5:50Active learning for optimization
- 10:42Corset selection for labeled data
- 15:46Remove redundant images with hashes
- 18:17Semantic redundancies in Cifar-10
- 20:42Cut data by brightness
- 25:00Q&A on hashes and coresets
related talks
50mTraining on the Test Set and Other Heresies
17mOptimized Image Classification on the Cheap
34mDimensionally reducing data squeezing out the good stuff
56mTutorial on Deep Learning II
40mTools and Tricks from a Pragmatic Data Scientist
47mDeep Learning with PyTorch
45mScale Learning and Reasoning Across Heterogeneous Gradients and Semantics
30mTransfer learning for image recognition in healthcare...