Research

Benchmarking platforms for AI, including Big-Bench, SuperGLUE, and MMLU (Massive Multitask Language Understanding), have provided valuable metrics for assessing the general performance of Large Language Models. However, these benchmarks reveal an evident gap in understanding the models' potential in sports performance tasks. Upon review of recent technical papers and reports from leading AI research entities such as OpenAI, Microsoft, Google DeepMind, Anthropic, Meta, Stanford and UC Berkeley [1, 2, 3, 4, 5, 6], it is apparent that the sports performance domain is yet to be explored in earnest.

The commonplace practices of curating and improving datasets (pre-training, fine tuning steps, including RLHF), such as data annotation methods, as well as scraping and filtering web-based content, which contain athletic training programs [7, 8, 9, 10] proves insufficient for gaining a comprehensive understanding of high-performance sports training. While this approach may suffice for recreational training pursuits such as weight loss, general health maintenance, and physical aesthetic enhancements, it proves insufficient for optimizing key athletic performance indicators such as maximal strength, reactive strength, peak power, rate of force development (RFD), velocity, ground contact, speed strength, V02 Max, change of direction (COD), range of motion, fatigue management etc... [11, 12, 13, 14, 15] in real world sporting contexts.

A radical transformation is needed in the way we collect, process, and evaluate sports performance data. Understanding the vast landscape - or "search space" - where all conceivable training programs exist is crucial. This search space can be thought of as the sum total of all possible combinations of exercises, rep ranges, set counts, intensity, volume, frequency, and rest periods that could constitute a training program. However our conventional data collection methods fall short when it comes to enabling Large Language Models to fully grasp the intricacies inherent in crafting training programs that effectively address an athlete's specific goals.

Consider this search space as a distinct plane of existence, home to a wide range of training programs per sports domain, from the exceptional to the mediocre, and even the downright ineffective. While some may find this comparison far-fetched, it could be likened to how AlphaFold [16] navigates the enormous search space of potential protein configurations, leading to breakthroughs in predicting protein structures. In both cases, the objective is to traverse a highly complex landscape and pinpoint the optimal solution. Just as AlphaFold illuminated new frontiers in the field of structural biology [17], this sports training "search space," once effectively navigated, could inspire similar breakthroughs. It holds the potential to unveil superior training methodologies that could far outstrip the current ones [18, 19, 20], thereby propelling the field of sports science to new frontiers. I anticipate moves similar to AlphaGo's Move 37 in respect to the design philosophy of athletic training programs [21, 22].

I'm still in the middle of understanding this "search space" and will talk more about it in the future as I make more progress. What I can tell you now is that this search space is "computationally very expensive" and will most likely need some kind of Deterministic or Stochastic Heuristic Search Method (such as Genetic Algorithms, Reinforcement Learning, Tree Search algorithms etc...) [23] to sample it effectively [can gain more insights about my work, from my MSc thesis - not the search space though, thats a new revelation]. In short, current Large Language Models don't know how to sample from the search space properly. All of this necessitates the establishment of new sport performance benchmarking methods specifically tailored to Large Language Models, to assist competitive athletes and coaches reach new peaks of performance pinnacles with AI support [24].

To conclude, consider this: would you entrust your athletic journey to a coach who lacks proficiency in their craft? Pursuing a training regimen advised by such an individual would undoubtedly be perilous and could lead to severe injuries. This runs counter to the central objective of any proficient strength and conditioning coach, whose primary mandate is to reduce the risk of injury, whilst optimizing performance gains [25] - not to encourage injuries from happening.

The AI community has been rigorously emphasizing the importance of AI safety/risk/ethics (red teaming) in recent times [26, 27, 28], and it's paramount that this principle is extended to its application in sports and fitness coaching. The goal is not only to enable athletes and coaches to enhance their performance but to ensure that these improvements are achieved in a manner that prioritizes health and safety. The implication here is clear: AI, like a good coach, needs to understand and respect the boundaries of safe practice while simultaneously offering opportunities for unprecedented growth and development. It's a factor that must be deeply integrated into its training, ensuring that any advice or training programs it produces are not just effective but also safe. In doing so, we inch closer to truly optimizing AI applications in sport. As a side note, I have been thinking about the implications of AI relative to performance enhancement drugs (PED's), as this is an issue that needs to be considered and addressed as well to foster fairness in competitive sports [29, 30].

Recognizing this gap in the AI application landscape, I am stepping forward to voice this concern and contribute towards filling this gap in AI applications in sports performance, starting with the development of Olyup.

Final Remarks: To all fellow researchers who may be grappling with obstacles in their work, I offer you this enduring sentiment: "Where there's a will, there's a way" (Just wanted to mention this because the last 4 years working on this project, which is interdisciplinary in nature - have not been smooth sailing). Keep these words close to your heart and let them guide you. Through the power of resolve and tenacity, you have the ability to surmount any challenge that comes your way, even during the most trying times. Remember, the darkness is always deepest just before the dawn. Your steadfast determination will lead you to breakthroughs and enlighten your path to discovery.

Author: Mbongeni "Bo" Ndlovu

Editor: ChatGPT