This is a pioneering venture at the intersection of hardware description language programming, optimization, and deep learning. This project aims to enhance the capabilities of automatic differentiation engines critical to deep learning on microprocessors under nanoseconds. Our theme centers around Robotic Learning, where we aspire to create robots that can learn and adapt on the go.
Unlike traditional project structures, this initiative follows an inverted format where the team drives decisions and sets targets, fostering a collaborative and dynamic research environment.
Whether you're interested in robotics, AI, hardware design, or software development, this project offers a unique opportunity to contribute to community development. This project is designed for a 7-member Undergraduate +/ Masters team.
In the rapidly evolving fields of robotics and generative AI, deploying advanced models often relies on deep learning to handle a diverse range of tasks. These tasks span from basic object detection to sophisticated challenges such as predicting the next best view in a 3D scene. Central to the success of these models is the use of automatic differentiation engines, which compute the derivatives of mathematical functions. These computations, whether simple or complex, are fundamental to deep learning training and are executed millions of times, making their speed and efficiency crucial.
Our project aims to discover and implement innovative methods, techniques, and algorithms that enhance the efficiency and effectiveness of these computational processes. In artificial intelligence, improving efficiency often translates to cost-effectiveness. Thus, our research targets one of AI's most pressing concerns: sustainable and efficient deployment.
Algorithms, despite their inherent complexities, must be computationally practical to be viable. In AI, the importance of speed goes beyond theoretical performance, emphasizing real-time execution with wall-clock time as the ultimate measure. The vast and dynamic nature of this field means that even small, novel improvements can have a significant impact on both practitioners and researchers in deep learning, driving meaningful advancements in the domain.
By addressing these objectives, we aim to make significant contributions to the efficiency of deep learning models, ultimately benefiting the broader AI and robotics communities.
Gradients on the Clock embarks on the essential groundwork and prototype development phase, aimed at creating a cutting-edge computational framework(s). This project uniquely integrates microcontrollers, Field-Programmable Gate Arrays (FPGAs), and GPUs/CPUs on a standard machine to facilitate core operations in gradient descent, crucial for the process of automatic differentiation. The core objective is to track the time, energy, and FLOps necessary for performing basic gradient descent on microprocessors and find the computationally cheapest means of achieving online learning on an FPGA with a microprocessing interface. This system is designed to bridge the gap between theoretical AI concepts and practical hardware applications, providing an educational foundation in the integral aspects of hardware and software integration for AI and specifically TinyML.
Participants will gain a multi-dimensional educational experience, achieving the following enhanced learning outcomes:
In case I have overwhelmed you with details, relax, this project will be performed in phases dependending on the feasibility and teams' developing expertise. Note that I will be providing a Basys3 FPGA board and Arduino MEGA 2560 microprocessor. I will also be providing for additional components as required.
Join a student-led Open Source Project at the forefront of AI and hardware interaction. This is a unique opportunity to find and potentially publish novel insights and contribute to transformative open-source AI research. By participating, you'll gain hands-on experience in hardware engineering, deep learning applications, robotics, and the mathematics of optimization. You'll enhance your technical skills and collaborate on a public platform that bridges the academic and industrial spheres.
We're seeking applicants with a robust foundation in university-level calculus and proficiency in Python. While prior knowledge of microprocessor programming and software development tools like Xilinx Vivado and Arduino is advantageous, it's not mandatory. Familiarity with web development (HTML, JavaScript, CSS) will also be helpful for contributing to our project documentation. Be prepared to acquire a diverse set of skills over the course of two semesters.
Additional Requirements: Candidates must demonstrate strong analytical thinking and problem-solving skills and be self-sufficient in terms of learning new concepts. Experience with any machine learning frameworks like PyTorch is highly desirable but not required. Commitment to collaborative development and an eagerness to learn new technologies are crucial. Applicants should also be ready to engage with both theoretical and practical aspects of the project, including participating in discussions, writing reports, and implementing prototypes.
This project is open to ANU School of Computing students as part of the Techlauncher program for the S2 2024 intake. We're actively seeking motivated students to join the inaugural team. This project is listed in the ANU Techlauncher redmine account. For more information, please contact me.