AI Model Search Pipeline Identifies Top Ensembles for Efficient Accuracy
In brief
- AI researchers have developed a new automated pipeline that systematically explores and evaluates thousands of complex neural network architectures.
- The system, running for 28 days on an NVIDIA RTX 4090, generated and tested over 1,000 candidate models.
- A key issue was discovered: the search process unintentionally focused too much on one family of networks called AirNet due to a flaw in the model selection code.
- This bias skewed the results but led researchers to identify the most reliable combinations-ShuffleNet and MobileNetV3-which achieved the highest accuracy (up to 63.2%).
- The pipeline's findings provide valuable insights for building efficient AI models, especially for resource-constrained applications like mobile devices.
- While some network families proved less effective, this work offers a more systematic approach to model selection.
- Moving forward, researchers plan to fix the bias issue and release their tools open-source to help others improve their AI systems.
- This advancement could make it easier for developers to find optimal neural network configurations without manually testing countless possibilities.
Terms in this brief
- Neural network architectures
- Different designs or blueprints for artificial neural networks that determine how data flows through the model. These architectures can vary in structure and depth, affecting the model's performance and efficiency.
- Ensembles
- A group of multiple machine learning models whose predictions are combined to improve accuracy and robustness. Using ensembles can lead to better results than relying on a single model alone.
Read full story at arXiv CS.LG →
More briefs
AI Researchers Shift Focus to "World Models
Scientists are moving away from language-based AI models to create "world models" that teach AI systems to react in physical environments. This shift matters because current AI models are limited to processing text and cannot understand the physical world. Top researchers like Fei-Fei Li are working on world models that can learn the statistical structure of space and time. New companies are forming to develop world models, with investors committing trillions of dollars to the effort. The future of AI will depend on these world models.
Meta Pauses Employee Tracker
Meta has stopped a program that tracked what employees did on their work computers. The program tracked keystrokes, mouse clicks, and what was on the screens. The program was meant to help train Meta's AI models. Over 1600 employees did not like the program and signed a petition against it. They said it was not private and they did not want their computer use data collected. Meta will now look into the program to see if anything went wrong. The company is spending a lot of money on AI, up to $145bn this year, and this program was part of that effort. Meta will now investigate and decide what to do next.
OpenAI and Broadcom Develop Custom AI Chip
OpenAI and Broadcom announced a new custom AI chip called Jalapeño. It provides better performance per watt than current state-of-the-art chips. The new chip is important because it helps OpenAI get the computing power it needs. OpenAI is a large buyer of Nvidia's chips but has to compete with other companies to get them. By developing its own chip, OpenAI can optimize its systems and get faster access to computing power. Other companies like Amazon and Google are also working on their own custom AI processors. The development of Jalapeño is part of a larger trend in the AI industry. Companies are looking for alternatives to Nvidia's high-powered processors. OpenAI will begin to roll out its new chip later this year and into the years ahead. OpenAI will continue to develop new chips to improve its computing power.
Microsoft's AI Speeds Up Rare Disease Diagnosis
Microsoft has introduced Talos, an AI system designed to accelerate rare disease diagnosis. The tool significantly reduces the time spent on manual reviews by automatically analyzing genomic data. It accurately identifies 90% of relevant diagnoses while only presenting 1.3 potential variants per patient for expert review. This means doctors can focus more on treatment rather than sifting through data. The system's efficiency is a breakthrough in genomic medicine, where human review is often the slowest step. By automating repetitive tasks, Talos allows experts to work faster and more effectively. Microsoft’s research shows that this approach could drastically improve diagnosis rates for rare diseases, which currently affect millions worldwide. The tool is already being used in pilot programs with notable success. Looking ahead, further integration of AI in healthcare could lead to even faster and more accurate diagnoses. Talos highlights the potential of technology to solve complex medical challenges, paving the way for a future where AI supports doctors in making quicker decisions.
OpenAI Unveils Custom Chip for Next-Level AI Processing
OpenAI has developed a new custom chip called "Jalapeño" in collaboration with Broadcom. This specialized hardware is designed to enhance the performance of large language models, which are used in tasks like chatbots and text generation. The chip aims to improve efficiency and speed, making it particularly suited for handling complex AI computations at scale. This development could be a game-changer for AI researchers and developers, as it provides a more powerful tool for training and deploying advanced models. By tailoring hardware specifically for large language models, OpenAI is addressing one of the key bottlenecks in AI processing-computational power. The chip is expected to become available by late 2026, which could significantly accelerate progress in AI capabilities across various industries. As AI technology continues to evolve, the integration of specialized hardware like Jalapeño will likely become more common. OpenAI’s move signals a shift toward optimizing AI systems not just through software advancements but also through purpose-built hardware. The future of AI processing may lie in such custom solutions, promising faster and more efficient computations for next-generation applications.