Cupertino, CA, December 18, 2023 – Aizip, Inc., a technology and market leader in artificial intelligence (AI) for edge applications, today announced the leap forward of building AI models with AI models.
The recent release of foundation models and generative AI (such as ChatGPT) has led to worldwide attention and exciting applications. Content generation produces high-quality results for business and consumer use. AI-produced coding has made amazing progress and is now becoming a reality. However, the dream of having AI develop AI has always been for science-fiction movies—until now. Aizip has teamed up with scientists from MIT, UC Berkeley, UC Davis, and UC San Diego to successfully demonstrate a fully automated AI-design pipeline, Aizipline, from data generation to model design to model testing, which involves a new application of the foundation AI models: to develop other AI models.
This interdisciplinary team of researchers has pioneered an artificial intelligence paradigm that is capable of automated efficient-AI model design. Utilizing a base of foundation models, the automation pipeline employs advanced techniques such as unsupervised representation learning, generative models, transfer learning, efficient neural architectures, neural architecture search, hardware-aware training, and sim2real testing to build different efficient AI models automatically in a scalable fashion. Through a fusion of data-centric, model-centric, and system-centric approaches, high efficiency can be achieved for many AI applications. The technical prowess of this AI-design automation is exemplified by the fully automated keyword spotting (KWS) design pipeline, which synthesizes diverse acoustic datasets, employing scalable neural architecture search in ultra-efficient neural architecture spaces. This AI-generated knowledge can then be packed into an economical model to achieve maximal efficiency. The ultimate goal of the AI-design automation paradigm is to create an “AI Nanofactory” where millions of specialized, efficient AI models can be generated with minimal human intervention to power the future of pervasive AI and AI engineering.
Aizip decided to first tackle the pervasive intelligence market. “With the help of large foundation models, small models will evolve faster than big ones, so the trend of improvements favors the edge. This is a space where Aizip has been an early adopter, and Aizip now leads in leveraging the big to dramatically improve the small,” said Brian Cheung, an AI scientist at MIT and Chief Scientist of Aizip. Over the past three years, Aizip has developed a wide selection of tools leveraging generative AI, foundation models, vector databases, self-supervised learning, and efficient neural architectures, among many other state-of-the-art AI advances. With these tools, the AI-model development cycle can be significantly shortened or even fully automated.
“Let’s take keyword spotting as an example,” explained Boltzmann Li, Principal Machine Learning Architect, who led the first demonstration at Aizip. “To develop a customized and ultra-efficient KWS model that can perform reliably under a variety of conditions, Aizip leveraged generative models to synthesize keywords with strong diversity in speaker identity, acoustic environments, noise conditions, and other factors.” Together with other AI-design tools such as ultra-efficient neural architectures, the KWS model-development pipeline in Aizip is now fully automated. Li noted that while human-in-the-loop (HIL) is still needed for many applications at present, Aizip’s ultimate goal is to achieve complete AI design automation (ADA).
“Nature evolved intelligent systems that operate with remarkable efficiency,” noted Bruno Olshausen, a professor at UC Berkeley and Director of the Redwood Center for Theoretical Neuroscience. “Just as the tiny brains of small animals with fewer than a million neurons must utilize efficient wiring and neural algorithms to perceive and act in the world, the tiny AI systems powering tomorrow’s edge computing devices will require clever, efficient solutions to operate with minimal power and footprint. There is much progress to be made in this direction, and Aizip has taken an important step in leading the way.” KWS, for example, is so efficient that it can be deployed onto hardware that costs as little as $0.10.
“We’re witnessing a revolution in human-machine interaction and brain-computer interfaces fueled by advances in brain- and body-sensing technology. Making sense of the massive data streaming from these sensors despite the high levels of variability and noise in their biological operating environments is a major challenge that calls for powerful AI, down to the physiological interface,” said Gert Cauwenberghs, a professor at UC San Diego and Co-Director of the Institute for Neural Computation. “Brain and body sensing in a wearable format requires efficient AI models that can be deployed at the edge. The technology at Aizip enables transformative applications in bio- and neuro-engineering.” Leveraging the Aizip ADA tools, Cauwenberghs’s group is working with Aizip to demonstrate efficient AI models for in-ear electrophysiological brain-state monitoring.
To produce AI models for these devices at scale, ADA is key. “At Aizip, we’re really building an AI Nanofactory—Aizipline” described Yubei Chen, CTO of Aizip and a professor at UC Davis. “We need to simplify every stage and build ADA tools accordingly. By leveraging foundation, visual-language, and generative AI models, the data collection, preparation, and testing process can be accelerated by orders of magnitude. Leveraging self-supervised learning, efficient architecture space, hardware-aware architecture search, scalable compilation, and many other techniques, our model design and development cycle can be further shrunk significantly.’’ Over the past two years, AI-design pipeline efficiency at Aizip has improved more than tenfold.
“We’re addressing many fascinating questions along the way. How do we find the right data among massive data? How to generate data from limited data? How to define and ensure real-world robustness? How to maximize the capacity of a neural network under resource constraints? How to debug an AI model? We try to simplify and optimize every stage of AI design with minimal human-in-the-loop,” reflected Chen, who also holds an affiliated position with UC Berkeley.
“Customized and efficient AI is essential for pervasive AI in real-world applications,” noted Yan Sun, Chairman and CEO of Aizip. With the support of AI tools, Aizip has developed and tested tens of thousands of models and delivered hundreds of models to customers worldwide. Some now support volume shipping for both consumer and enterprise markets, benefiting society and promoting a sustainable world. Aizip is committed to bringing pervasive AI to anything, anywhere, and any time.
The broad portfolio offered by Aizip consists of Aizip Intelligent Audio (AIA), Aizip Intelligent Vision (AIV), and Aizip Intelligent Time-Series (AIT) products. Aizip provides TinyML and edge models to its IC partners, integrators, service providers, and users worldwide. For more information, please contact email@example.com.
* * *
About Aizip, Inc.
Aizip, Inc., develops AI models for edge applications. Based in Silicon Valley, Aizip provides models with superior performance, quick turnaround time, and excellent ROI. These models can be used in a wide range of applications for an intelligent, automated, and connected world. For more information, visit www.aizip.ai.