The juror who found herself guilty
In 1990, Estella Ybarra was a juror on a trial that convicted Carlos Jaile, a vacuum cleaner salesman, of raping an eight-year-old girl. Ybarra had doubts about his guilt from the start, as the victim's description of Jaile's appearance and vehicle did not match his. Despite this, Ybarra eventually succumbed to pressure from other jurors and voted to convict Jaile. For years, Ybarra carried the burden of her decision and wondered about Jaile's life in prison. In 2017, after finding an old certificate of appreciation she received for her jury service, Ybarra started to question her past actions. She felt a shift and began to question whether she had made the right decision. She learned more about the flaws and biases in the criminal justice system, especially when it comes to wrongful convictions. This realization led Ybarra to seek out Jaile and apologize for her role in his wrongful conviction. They eventually met in person and Ybarra took steps to help Jaile gain his freedom. Her journey highlights the importance of questioning one's actions and seeking justice for those wrongfully convicted.
Gene therapy allows an 11-year-old boy to hear
Aissam Dam, an 11-year-old boy who was born deaf, has become the first person in the United States to receive gene therapy for congenital deafness. Aissam, who previously lived in a poor community in Morocco, used a sign language he invented to communicate and had never received any schooling. However, after moving to Spain, his family took him to a hearing specialist who suggested he might be eligible for a clinical trial using gene therapy. On October 4, Aissam underwent the treatment at the Children's Hospital of Philadelphia, and it was a success. The gene therapy aimed to replace a mutated gene called otoferlin, which causes deafness in about 200,000 people worldwide. Although the therapy is still in the early stages and more patients need to be involved, researchers believe that this groundbreaking study could lead to gene therapies for other forms of congenital deafness.
Why is machine learning 'hard'? (2016)
Machine learning has become more accessible in recent years, with online courses, textbooks, and frameworks making it easier to implement models. However, machine learning remains a relatively difficult problem. Implementing existing algorithms and models for new applications requires a deep understanding of available tools and their trade-offs. Debugging machine learning algorithms is particularly challenging, as there are often multiple dimensions along which things can go wrong - algorithmic issues, implementation bugs, bugs in the data, and limitations in the model's capabilities. This makes the debugging process more complex and time-consuming. Additionally, there can be long delays between implementing a potential fix and seeing the results, especially in deep learning models that require hours or days to train. This necessitates a parallel experimentation paradigm, where multiple experiments are run simultaneously. Building intuition for debugging machine learning is crucial, as it allows developers to associate certain behavior signals with potential issues in the debugging space. Ultimately, fast and effective debugging is a critical skill for implementing machine learning pipelines successfully.
Direct pixel-space megapixel image generation with diffusion models
The Hourglass Diffusion Transformer (HDiT) is a new image generative model that can train at high resolutions directly in pixel-space. It combines the efficiency of convolutional U-Nets with the scalability of Transformers, allowing for linear scaling with pixel count. Unlike traditional methods that require additional techniques like multiscale architectures or latent autoencoders, HDiT is able to achieve successful training without them.
The HDiT model is built upon the Transformer architecture, which is known for its ability to handle billions of parameters. By leveraging this scalability, the model is able to bridge the gap between efficiency and scalability in image generation.
In experiments, the HDiT model showed competitive performance with existing models on ImageNet-256, a common benchmark dataset. Additionally, it set a new state-of-the-art for diffusion models on FFHQ-1024, demonstrating its effectiveness in high-resolution image generation.
Overall, HDiT presents a promising approach to generating high-resolution images without compromising on efficiency or scalability.