CS231n-LAB

This is the accompanying assignment for Stanford University's computer vision course.

assignment1

Assignment 1 in CS231n-LAB focuses on fundamental concepts in machine learning and neural networks. In this assignment, we implemented several key components:

• K-nearest neighbors (KNN) classifier for classification tasks.
• Fully connected layer for dense neural network architecture.
• Rectified Linear Unit (ReLU) layer for introducing non-linearity.
• Support Vector Machine (SVM) layer for classification.
• Softmax layer for multiclass classification.
• Constructed a two-layer fully connected neural network for learning basic neural network principles.

assignment2

Assignment 2 in CS231n-LAB delves into advanced topics in deep learning and neural network architectures. This assignment includes the implementation of the following techniques:

• Batch normalization layer, Layer normalization layer, Spatial-Batch normalization, Spatial-Group normalization layer for normalizing activations within features.
• Dropout layer for regularization and preventing overfitting.
• Convolutional layer for learning spatial hierarchies of features.
• Built a high-performance deep learning network using PyTorch framework, integrating the aforementioned techniques for optimal results.

assignment3

Assignment 3 in CS231n-LAB explores feature extraction, adversarial examples, and advanced machine learning frameworks. This assignment involves the following tasks:

• Generated saliency maps, fooling images, and class visualizations to analyze network activations and vulnerabilities in neural networks.
• Recurrent Neural Networks (RNN): Suitable for sequential data processing tasks such as natural language processing and time series analysis.
• Long Short-Term Memory (LSTM): A type of RNN architecture with improved ability to capture long-range dependencies.
• Transformer: Known for its effectiveness in handling attention mechanisms and transformer-based architectures like BERT for language understanding tasks.
• Generative Adversarial Networks (GAN): Used for generating realistic data samples, such as images, by training a generator and discriminator network simultaneously.
• SimCLR (Contrastive Learning): Focuses on learning effective representations from unlabeled data using contrastive loss functions, enabling better generalization and transfer learning capabilities.

PS: Some files such as cifar-10-python.tar.gz, coco_captioning are not uploaded. If you need to run the code in one click, please download the corresponding files from the address provided in the ipynb or terminal before running.