Research Themes

Integrating hardware, software, algorithms and data for scalable analytics and security is at the core of what we do.

Learning Based System Design and Optimization
Our customized performance optimization engine is automated and works across platforms. The process starts with automated measurement of the hardware. The system then abstracts out the key hardware characteristics. These abstracted hardware details are then integrated with the machine learning algorithms as well as the data in order to optimize the machine-learning computations based on the limits of the hardware being used as well as the characteristics of the data.

This approach offers research collaborators and industry partners opportunities to get significantly better customized hardware acceleration results than are available elsewhere, and to get these results without the expensive, slow process of manually customizing machine learning algorithms based on the particulars of a hardware system.

When the characteristics of the data change, the systems reassess and rebalance how and where the computations take place on the hardware.

“We can cope with dynamics of the data and dynamics of the hardware platform,” said electrical and computer engineering professor Farinaz Koushanfar, co-director of the UC San Diego Center for Machine-Intelligence, Computing and Security.

The customized and automated hardware-software integration approach provides researchers with the ability to do serious data analytics on mobile and embedded platforms, often in real time.

These solutions often integrate adaptive data collection processes with training, learning, and inference in real-time and streaming applications.


Real-Time and Interactive AI

The MICS Research Center stands at the forefront of cutting-edge technological advancements, driven by a team of distinguished faculty members who are luminaries in the domains of statistical learning, information theory, signal processing, and data science. Our research center is dedicated to exploring and innovating in the realm of Real-Time and Interactive AI, where the convergence of adaptive data collection processes, advanced training methodologies, and generative AI models heralds a new era of transformative applications.

The cornerstone of our research theme lies in the seamless integration of feedback information theory with the intricate domains of training of generative AI models. This fusion gives birth to a paradigm that is not only real-time and predictive but also well-suited for active learning applications. 

“We can cope with dynamics of the data and generative models by incorporating what we know from theory of active learning and sequential information theory,” said electrical and computer engineering professor Tara Javidi, co-director of the UC San Diego Center for Machine-Intelligence, Computing and Security. Our interdisciplinary team of experts collaborates intensively to unlock the full potential of Real-Time and Interactive AI across various sectors.


Robustness, Security, Ethics, and Privacy for AI Systems
Through their work at the interface of hardware, software, algorithms and data, the Center is uniquely positioned to engage with industry partners on a wide range of urgent security challenges that are getting more complicated every day. Now that the boundaries of the Internet have extended well beyond traditional computing devices to include the Internet-of-Things (IoT), intelligent vehicles, smart grids and more, the attack domain has expanded further into the physical world and includes even more critical infrastructures.

With the Center’s deep expertise in hardware design as well as hardware-software-algorithm-data integration, the researchers are working on privacy preserving computing systems that can determine “who does what computation where” – which is a key aspect of secure computing.

Secure embedded medical devices and handheld DNA analytics are just two possible outcomes of cutting edge research on privacy preserving computing systems focused on the micro-managing of computations, an approach that is also relevant for IoT security.

“It’s about knowing, and controlling, who holds which permissions to each part of the system. Privacy preserving computing is hard to generalize. There is currently no general way to address IoT security,” Koushanfar said.

Internet-of-Things (IoT) security is one area in which the Center is currently working – and it is extremely challenging. “You need to know who is doing what computations and who holds which permissions when you’re thinking about IoT security,” said Koushanfar. She noted that IoT security is often ad hoc and extremely hard to generalize, meaning there is significant room for innovation. Regrettably, most of today’s solutions are platform dependent and based on adversarial models.


Future Applications and Projects
The automated, customized hardware acceleration tools that the Center has developed are currently tailored for machine learning and deep learning applications. But their approach to automated and customized hardware software integration can be expanded to a much broader group of challenges facing industry. In fact, the approach can efficiently solve inverse problems in areas such as surgical robots, imaging systems, and low-power sensor networks.


Additional New Projects Include:
Interactive machine learning
Optimized integrated sensing
Rapid software development and testing


Industry Partners