You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
CS 198-126: Modern Computer Vision Fall 2022 (University of California, Berkeley) Lecture 1 - Intro to Machine Learning
CS 198-126: Lecture 1 - Intro to Machine Learning
In this lecture on machine learning, the instructor covers a broad range of topics, including an introduction to the course, an overview of machine learning, different types of machine learning, machine learning pipeline, labeling data, and loss function. The concept of bias-variance trade-off, overfitting, and underfitting is also discussed. The instructor emphasizes the importance of choosing the right function during the process of machine learning and the role of hyperparameters in the process. The overall goal of machine learning is to accurately predict new data, not just fit the training data. The lecturer encourages students to attend the class and make an effort to learn about machine learning and deep learning.
CS 198-126: Lecture 2 - Intro to Deep Learning, Part 1
CS 198-126: Lecture 2 - Intro to Deep Learning, Part 1
In this YouTube lecture on Intro to Deep Learning, the instructor discusses the basics of deep learning models and how to train them using gradient descent, covering different building blocks for neural networks and why deep learning is such a prevalent technology. The lecture introduces the perceptron and stacking multiple perceptrons to create a more complex and sophisticated neural network, and explains how to compute the output by matrix multiplication and a final addition, with the middle layer using a ReLU activation function. The speaker addresses the use of the Softmax function and ReLU activation function, using loss functions as metrics for evaluating how well the model is performing, and the concept of gradient descent optimization. Finally, the instructor discusses the idea of deep learning and how a big neural network prompts low loss despite its ability to memorize the data. Also the lecturer introduces the concept of hyperparameter tuning in neural networks to improve their performance with specific datasets. He notes that there are no universal values for hyperparameters and suggests exploring different options such as layer numbers and activation functions. Due to time constraints, the lecture ends abruptly, but the lecturer assures students that the upcoming quiz will not be overly difficult and will be accessible on the GreatScope platform.
CS 198-126: Lecture 3 - Intro to Deep Learning, Part 2
CS 198-126: Lecture 3 - Intro to Deep Learning, Part 2
In this section of the lecture, the concept of backpropagation is explained, which is a faster way to get all partial derivatives required for the gradient descent algorithm without performing redundant operations. The lecturer also discusses how to improve upon vanilla gradient descent for deep learning optimization and introduces momentum, RMSprop, and Adam as optimization methods. The importance of keeping track of a model's training history, the use of batch normalization, and ensembling as a technique to improve model performance are also discussed, as well as techniques commonly used in deep learning to help decrease overfitting such as dropout and skip connections. Finally, the lecturer briefly touches on the ease of use of PyTorch and opens the floor to questions.
CS 198-126: Lecture 4 - Intro to Pretraining and Augmentations
CS 198-126: Lecture 4 - Intro to Pretraining and Augmentations
In this lecture, the speaker explains the evolution of feature extraction in machine learning, the advantages of deep learning, and how transfer learning can be used to improve the accuracy and speed of models. They also discuss the concept of freezing and fine-tuning layers in neural networks and the importance of embeddings in reducing the dimensionality of categorical variables. The lecture introduces self-supervised learning and its different tasks, including the jigsaw, rotation, and masked word prediction tasks, which can be used to pretrain models and transfer learned representations to downstream tasks. Finally, the renewed interest in self-supervised learning in computer vision is discussed, and the lecture encourages students to complete the homework on the high Crush notebook.
straightforward for tabular data, but complex for data like text, audio, or images. However, for images, there are specialized feature extractors available in classical computer vision.
CS 198-126: Lecture 5 - Intro to Computer Vision
CS 198-126: Lecture 5 - Intro to Computer Vision
This lecture on computer vision covers various topics, including the history of computer vision and its development over the years. The instructor also explains deep learning and how it improves on classical computer vision methods. The lecture delves into the concept of convolutions and how they are used as feature extractors in computer vision, leading to the creation of convolutional neural networks (CNNs). Additionally, the lecture discusses the role of receptive fields and introduces pooling layers as a method to increase the receptive field of CNNs. Overall, the lecture provides an overview of computer vision as a field and the techniques used to extract information from images. In the second part of the lecture, various techniques for preserving the size of an image during convolutions are discussed, including padding and same padding. The concept of stride in convolutional layers is also covered, demonstrating how it can mimic the effect of a pooling layer. The anatomy of a CNN and its hyper-parameters, including kernel size, stride, padding, and pooling layers, are explained, with emphasis on how a convolutional layer acts as a feature extractor that passes low-dimensional blocks of features to a fully connected network for classification. The lectures also cover the LeNet network architecture for classifying handwritten digits and the importance of normalizing image data before passing it through a neural network. Finally, data augmentation is discussed as a technique for creating additional training data, and the importance of model checkpointing while training is emphasized.
CS 198-126: Lecture 6 - Advanced Computer Vision Architectures
CS 198-126: Lecture 6 - Advanced Computer Vision Architectures
This lecture on advanced computer vision architectures focuses on convolutional neural networks (CNNs) and their various techniques. The lecturer explains the architecture of AlexNet and VGG before delving into advanced techniques such as residuals to maintain backward residual values for higher accuracy and simpler architectures. The use of bottlenecks and one-by-one convolutions are discussed, as well as the importance of being able to learn the identity in computer vision architectures. The lecture also covers the issues of vanishing gradients in neural networks and how it can be alleviated with batch normalization and residual networks. Techniques such as global average pooling and depth-wise separable convolution are explained in-depth, followed by discussion of the mobile net architecture and its benefits.
Also the lecturer examines advanced computer vision architectures and focuses on optimizing convolutional neural network models by using step local convolutions and one by one convolutions. He emphasizes the importance of understanding these optimizations and the problems that may arise with certain optimizations in building future networks efficiently. The lecture concludes with a discussion on the tradeoff between accuracy, performance, and model size, highlighted by the comparison of the efficient net model to other networks. Students are informed of an upcoming quiz and a homework assignment due the following Friday.
CS 198-126: Lecture 7 - Object Detection
CS 198-126: Lecture 7 - Object Detection
The lecture discusses object detection, specifically adding localization to a simple classification CNN, the IOU method for object detection, the R-CNN system, and optimizing object detection algorithms to minimize processing time with YOLO. The video explains YOLO by chopping up an image, and discusses the challenges with YOLO object detection, including using anchor boxes to eliminate ambiguity. Finally, the YOLO architecture is explored, which is a fully convolutional neural network for object detection, and the storage of a large number of classes for classification is presented as an ongoing research question. The speaker recommends reading "The Yellow Paper" while advising against RCNN due to unreadability.
CS 198-126: Lecture 8 - Semantic Segmentation
CS 198-126: Lecture 8 - Semantic Segmentation
The lecture discusses image segmentation, including semantic segmentation and instance segmentation. The main goal of segmentation is to detect all objects in an image and separate them out. The lecturer explains how a convolutional neural network (CNN) can be used for semantic segmentation and how downsampling can help with computationally expensive full resolution images. Different approaches to transform a small volume back into an image size are also discussed. The lecture introduces the U-Net, a model for semantic segmentation that combines previous improvements with skip connections, and explains how it can be expanded to instance segmentation using the Mask R-CNN approach. A pre-trained semantic segmentation model is demonstrated, and the speaker talks about pre-training and upcoming course assignments.
CS 198-126: Lecture 9 - Autoencoders, VAEs, Generative Modeling
CS 198-126: Lecture 9 - Autoencoders, VAEs, Generative Modeling
In this lecture, the concept of generative modeling is introduced, which involves using machine learning to create new images based on a dataset. Autoencoders, a type of neural network used for feature learning, are explained, focusing on their structure and how they can learn features of input data through compression and reconstruction. The lecture also covers variational autoencoders and their benefits, as well as the use of structured latent spaces in autoencoders to interpolate between images. The importance of vector quantization for working with discrete data is discussed, and the loss function for a variational autoencoder is explained, which includes a reconstruction loss and a commitment loss to prevent hardcoding of the input data. The lecture ends with a recap of the topics covered.
CS 198-126: Lecture 10 - GANs
CS 198-126: Lecture 10 - GANs
The lecture on GANs introduces the concept of two networks, the discriminator and the generator, competing against each other in a game theory-esque setup. The generator's input is random noise, which it assigns meaning to generate real-looking images, and the discriminator's job is to judge if the image is real or fake. GANs use a loss function that corresponds to negative cross-entropy loss, with the generator wanting to minimize and the discriminator wanting to maximize it. The value function represents how well the generator is doing and needs to be maximized by the discriminator by correctly classifying fake and real data. The lecture also covers issues with training GANs and the non-saturating loss that allows the generator to have more agency to change.