Principal Investigator: Virginia Smith, Assistant Professor, Electrical and Computer Engineering, College of Engineering
Co PI: Ameet Talwalkar, Assistant Professor, Machine Learning, School of Computer Science
We have received funding from the Carnegie Bosch Institute for Machine Learning for Connected Intelligent Systems. Modern networks of Internet-of-Things (IoT) devices, such as mobile phones, wearable devices, and autonomous vehicles, generate massive amounts of data each day. Due to privacy concerns with shipping personally identifiable information, as well as the growing computational capacities of IoT devices, it is increasingly attractive to keep data local and push more computation to the edge. The burgeoning field of federated learning explores training machine learning models directly at the edge. Federated learning can be naturally cast through the lens of optimization, a key component in formulating and training most machine learning models.
Ultimately, we envisage a world in which federated learning over massive IoT networks is as seamless as performing linear algebra operations in MATLAB on a laptop. We aim to take a step towards this lofty goal by developing a class of distributed optimization primitives for federated learning, akin to the underlying linear algebra methods that comprise the LAPACK package. To achieve this, we must tackle the statistical and systems challenges inherent to federated learning, which require a fundamental departure from methods designed for traditional data center environments.
Our proposed work has the potential to cause a paradigm shift for data-driven applications in IoT networks. In addition to our methodology, our proposed benchmarking framework will be a crucial asset to the broader research community to concretely define the challenges in federated learning and promote reproducibility in empirical evaluations. We have a strong record of mentoring students from underrepresented populations, and will continue to do so as part of this work.