Research

Research areas

Professor Iyer leads the DEPEND Group at CSL, with a multidisciplinary focus on systems and software that combine deep measurement driven analytics and machine learning with applications in two important domains: i) trust (that spans resilience and the security of critical infrastructures) and ii) health (that spans computational genomics and health analytics focused on personalized medicine). The Depend Group has developed a rich AI analytics framework that has been deployed on real-world applications in collaborations with industry, health providers, and government agencies including NSF, NIH, and DoD.  Our overarching interest spans several related aspects of developing artificial intelligence and machine learning methods that apply to predictive health-related analytics both in online and offline analysis. The work in our group brings together statistical analysis with both Bayesian and deep-learning methods, often integrated into joint hardware and software implementations.

Broadly, our research is on two broad areas of AI and Systems.

The first is in Systems for AI. We innovate AI-supported software (SW) systems, addressing major gaps in hardware (HW) accelerators to significantly enhance the performance and resilience of the underlying HW while addressing the significant impact of accelerator failures that stifle system performance objectives, resulting in substantial financial losses for both vendors and customers. This AI systems work is primarily supported by the IBM/Illinois Discovery Accelerator Institute (IIDAI) and we have consistently been among its top grant recipients. Our team including, ECE student Archit Patke and recently graduated CS student Haoran Qiu, partnering closely with ECE Prof. Tamer Basar and his students, developed QLM, a multi-GPU, LLM serving system that maximizes HW usage while optimizing application-specific performance objectives. QLM has become the de facto standard for distributed LLM serving. Major cloud hyperscalers, including Google, IBM, ByteDance, and Snowflake, have adopted it. QLM is open-sourced as AIBrix and is supported by leading open-source organizations, including the Cloud Native Computing Foundation (CNCF), the PyTorch Foundation, and several other open-source houses.

The second is AI for Systems and AI innovations in systems management and control, having substantial applications in healthcare, spanning individualized medicine for cancer treatments like immunotherapy, AI twins in cardiology, pharmacogenomics in AI-driven drug efficacy, and various other diseases, enhancing patient-specific drug efficacy.  The group’s interest spans several related aspects of developing artificial intelligence and machine learning methods that apply to predictive health related analytics both in online and offline analysis. The work in our group brings together statistical analysis with both Bayesian and deep-learning methods, often integrated into joint hardware and software implementations. Working collaboratively with clinicians and medical researchers we combine omics and patient specific data to build predictive machine learning models and algorithms that have been transformational in-patient diagnosis and care. The intent is that the tools and technologies will translate into clinical practice. We also develop methods to perform measurements and to benchmark the accuracy of our findings. In health, we primarily work with and are supported by Mayo Clinic and grants from NIH and OSF, UIC, and Carle, while we collaborate with major drug companies like Roche, Eli Lilly, and Genentech, often through our Mayo connections. We also work in neuro-centric diseases like epilepsy, dementia, and autism. Our recent work in developing AI-managed brain implants to control epileptic seizures and stave off dementia has garnered joint patents with Mayo and is being incorporated into their patient care. This kind of research that co-joins science and engineering innovations with sometimes sparse data is a domain where we excel and have had a major impact.

Reliability and Security
Measurement driven analysis and modeling of dependability in cyber-physical systems, high-performance computing systems, cloud computing systems.