School of Engineering Department of Computer Science and Engineering 103 Generative AI Supervisor: CHEN Qifeng / CSE Student: LI Jiajun / ELEC Course: UROP 1000, Summer For this UROP1000, at the beginning, I was assigned tasks about data collection, learned some techniques for collecting data and crawling, and learned how to use Linux. Then, I did some reading on the paper about sketch-to-image and learned the principle of the diffusion model. In this process, I learned how to read a paper, the advantages and disadvantages of the methods between different papers, discuss what I learned in the paper with my mentor, and communicate with her about what I think can be improved. Finally, I tried the implementation methods in the papers and experienced more deeply the shortcomings and benefits of their methods. Generative AI Supervisor: CHEN Qifeng / CSE Student: MAHAPATRA Amadika / MATH-CS Course: UROP 1100, Fall In the traditional GAN architecture, the generator ( ) takes a latent code ∈Ƶ where Ƶ is the input latent space. We then define the discriminator ( ), where ( ) represents the probability that came from the training data rather than the distribution produced by G. G and D are then pitted in an adversarial minimax game, with G being trained to maximise the likelihood that D mis-classifies a generated sample for a training data sample. This is equivalent to stating that G seeks to minimise log (1− � ( )�). StyleGAN, published in 2019 provides a style-based generator that maps the input to another latent space by a non-linear mapping network. A sample ∈ is fed into learned affine transformations, which return style vectors . StyleGAN uses a progressive GAN architecture which requires the addition of noise to each output generated by the synthesis network to increase the dimensionality till the desired output size is achieved. From the discussion above we see two examples of generator architectures that operate in different latent spaces. This report aims to summarise the research on GAN inversions, including the novel and exciting method proposed by StyleCLIP which was tested personally during the course of this project. Finally, this report proposes a potential new mechanism using denoising diffusion models as a sub-process in a new GAN inversion method which is still under construction. Generative AI Supervisor: CHEN Qifeng / CSE Student: PANG Yukun / PHYS Course: UROP 1000, Summer This paper explores fundamental algorithms in computer vision, an important domain within Artificial Intelligence (AI). By tracing the field’s development from its early explorations to current algorithms, we provide insights into the revolutionary techniques that have shaped this discipline. Our analysis begins with foundational methods such as the k-Nearest Neighbor (kNN) algorithm and progresses to advanced neural networks such as Convolutional Neural Networks (CNNs). We discuss how each algorithmic advancement has contributed to overcoming specific challenges in image recognition, object detection, and scene understanding. Additionally, we explore the integration of CNNs with Transformer architectures.
RkJQdWJsaXNoZXIy NDk5Njg=