Deep Convolutional Neural Network AlexNet
From 1990 to 2001, SVM Support Vector Machine was the most mainstream method
- Feature extraction
- Select kernel function to calculate similarity
- Convex Optimization Problem
- Nifty Theorems
Computer vision problems are described as geometric problems - Extract features
- Describe the collection (e.g. multi-camera)
- non-convex optimization
- pretty theorem
- Works well if the assumptions are satisfied
The most famous method around 2005 is feature engineering - Feature engineering is key
- Feature descriptors: SIFT, SURF
- Bag of Visual Words (Clustering)
- Finally, using SVM
machine learning from 1920 to 2020, the change of data, memory, and CPU growth,
data and computing power, lead to different preferences for the selection of network methods. Maybe after a while, the neural network will decline again. Methods or new methods are on fire again. The general upgrade network is also popular for 10 to 20 years.
ImageNet In 2010, Li Feifei launched a project at StandFord. The single image is larger and the data set is larger.
AlexNet - AlexNet wins the 2012 ImageNet competition
- Deeper and bigger LeNet
- Major improvements: Dropout, ReLu, MaxPooling
- Changes in Computer Vision Methodology. Moving from SVM to deep learning, deep learning does not need to care much about computer vision.
AlexNet's small changes to LeNet took 20 years.
more details
- Activation function changed from sigmoid to ReLu (slow down gradient disappearance)
- A dropout layer is added after the fully connected layer is hidden
- data augmentation
2. Code implementation
The FashionNIST and MNIST grayscale images used here are implemented, so there is only one Channel channel.
- AlexNet test acc 0.883
- LeNet test acc 0.827
LeNet test acc 0.827
3. Q&A
-
- The Local Response Normalization in AlexNet is to generalize the data, but it has little effect later.
-
- Why is LeNet not a deep convolutional neural network? Because SVM was more popular at that time, it had a lot to do with the data magnitude and computing power at that time.
reference
https://www.bilibili.com/video/BV1h54y1L7oe?p=1