Transfer Learning and Finetuning Deep Convolution Neural Network
summary
Claims reuse of an ImageNet-pretrained GoogLeNet by fixing the first two convolutional layers and fine‑tuning upper layers to transfer features to domain tasks: fluorescein angiographic eye images for diabetic retinopathy and ACS fashion images for clothing/type captions. Uses image augmentation (rotations), 24x24 inputs, trains ~50 iterations on a single 4 GB GPU taking about a week, then pairs DCNN features with an LSTM for captions. Results report ~96% accuracy on the DR set and markedly better captions after fine‑tuning.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:02Traditional vs deep learning
- 4:58Deep learning for unstructured data
- 8:32CNN history and rise
- 10:18Transfer learning and fine-tuning
- 13:24How CNNs work
- 21:43Transfer learning in CNNs
- 24:52Diabetic retinopathy case study
- 34:06Image augmentation and training
- 41:04Model reuse across domains
- 42:30Fashion captioning case study
- 49:41Healthcare applications
- 51:30Q&A
related talks
30mTransfer learning for image recognition in healthcare...
08mDeep Transfer Learning for Computer Vision
56mTutorial on Deep Learning II
55mDeep Learning in Medical Image Diagnostics
47mInterpreting Deep Neural Networks (DNNs)
59mDeep Neural Networks for Video Applications
42mDeep Learning for Computer Vision
17mOptimized Image Classification on the Cheap