Adeoluwa Adekanmbi
MIT Department: Electrical Engineering and Computer Science
Faculty Mentor: Priya Donti
Research Supervisor: Arsam Aryandoust
Undergraduate Institution: Howard University
Hometown: Ibadan, Nigeria
Website: LinkedIn
Biography
Adeoluwa Adekanmbi is a rising junior majoring in electrical engineering at Howard University. Growing up in Nigeria, his diverse upbringing instilled in him a profound understanding of both privilege and disparity, shaping his perspective on global challenges. At MIT, Adeoluwa works under Dr. Priya Donti, where he is conducting research to unify machine learning (ML) tasks and dataset characteristics to create better ML models that can perform well in diverse domains. At his undergraduate institution, he conducts research under Dr. Charles Kim on radial pulse monitoring, utilizing embedded systems. Adeoluwa is interested in developing innovative solutions to mitigate climate change. As such, Adeoluwa’s vision extends to graduate studies, where he aims to explore the intersection of artificial intelligence and renewable energy, hoping to make a lasting impact on society. He is a pianist and enjoys connecting with people, playing games, and engaging in sporting activities during his leisure time.
Abstract
A definition and comparison of unified machine learning tasks and data representations
Adeoluwa Samuel Adekanmbi1, Arsam Aryandoust2, Priya L. Donti2
1Department of Electrical Engineering and Computer Science, Howard University
2Department of Electrical Engineering and Computer Science, Massachusetts Institute
of Technology
Artificial General Intelligence (AGI) requires the design of machine learning (ML) models that are capable of processing diverse data types, such as images, audio, and text, while simultaneously handling multiple tasks, such as chatting, robot control, or playing Atari games. Achieving this requires integrating these tasks and data types into a unified framework. Currently, there is no clear definition of a unified ML task and data representation. The lack of a unified representation results in inconsistent data handling and integration, challenges in reproducibility and benchmarking, and suboptimal model performance. Therefore, we introduce the first such definition in this work. For this, we review, refine, and compare different unified formulations. We demonstrate how various ML tasks, and their associated datasets can be incorporated into the frameworks we propose. We then generate performance benchmarks using a simple multi-modal, multi-tasking ML model for each proposed framework. Lastly, we present an application that can transform new ML tasks and their associated datasets into our unified frameworks using prompt engineering. Overall, introducing a unified framework for ML tasks and data types can significantly advance the field of AGI, improve ML models’ performance and reproducibility, and enable a wide range of innovative applications across different domains.
