Artificial Intelligence, Algorithms, and Deep Athletics – Part 1
These days we hear a lot about both artificial intelligence (AI) and algorithms. Without this becoming a novel, let’s spend a little bit of time understanding what is meant by each term and why (or if!) it matters in the context of your fitness routine.
Artificial intelligence (AI) is a revolutionary technology that will have – and in many cases already is having – a huge impact on the world around us. In order for an AI engine to accomplish its intended task, it relies on being trained on a large set of data, and for what we’re discussing here, the overall training approach is known as Supervised Learning. When examined individually, each member of the training data set exhibits either the desired outcome or an undesirable outcome.
For example, consider self-driving automobiles, and specifically pedestrian avoidance. A training data set could consist of 30-second video segments of cars driving in an urban setting. In this context, each segment can be categorized into meeting either the desired outcome (the car did NOT hit a pedestrian) or an undesirable outcome (the car DID hit a pedestrian). An enormous training set could be built using data from actively driven cars, with the rare occasions in which a collision did occur being flagged as producing the undesired outcome. Using this approach, an AI engine can be trained to recognize if a nearby pedestrian does or does not pose a collision risk and, subsequently, if a collision does appear possible, to relay that critical information to a higher-level control system to take appropriate action.
An algorithm, on the other hand, takes a fundamentally different approach. Algorithms are best thought of as a sequence of steps – including both logical “if/then” steps and mathematical equations – that are rigorously defined by their creator. There is no ambiguity in an algorithm, as each branch on the logic tree and each equation is a result of the expertise of the algorithm’s creator. As such, building a successful algorithm requires a fundamental understanding of the system to which the algorithm is applied, which can be both a blessing and a curse depending on the situation. The rigorously-defined nature of algorithms provides an additional benefit in troubleshooting unexpected or undesired outcomes, as the precise path that created the unexpected outcome can be traced. Tracking down the reason for the outcome usually identifies a flaw in the algorithm itself, but it may also reveal a limited or incorrect understanding of the system itself. In this case, the very task of debugging an unexpected result can expand the creator’s knowledge of the system!
I know what you’re thinking -- “That’s interesting and all, but what does this have to do with my workout later today?” I’m glad you asked!
Let’s start with the AI approach to fitness programming. As we’ve already seen, training data is a crucial element in developing an AI engine that produces appropriate results. To create new workouts means first developing a large database of “good” workouts that can be used to train the AI engine. The difficulty with this approach comes in defining what exactly constitutes a good workout. Building a training database consisting of millions of workouts scraped from a myriad of sources on the web will satisfy the training database size requirement. There is at least one app out there producing fitness programming with this exact approach. But if they are just being automatically retrieved, what’s the quality of those workouts as it pertains to a unique individual (you)? If half the workouts in the training database are of lesser quality, then the AI engine will produce workouts that reflect that.
Now consider a training database that consists solely of workouts written by top coaches for their athletes, coaches that are training Olympians, and other elite athletes. This solves the input quality problem, but a new problem arises – the fitness level and goals of the sample athlete don’t necessarily match the intended user. Because what is good for an elite athlete is rarely appropriate for someone looking to maintain a solid fitness level for overall health. And what’s good for that person is similarly not appropriate for those returning to regular exercise after years of a sedentary lifestyle.
In that same vein, a program designed to build strength should look different from one focused on increasing endurance. While many aspects may be similar, the nature of the AI process requires the two programs to be handled relatively independently, building each one separately from their corresponding training data. The AI process’ dependence on its training database also complicates the troubleshooting process. If an undesired pattern begins to emerge, the training database needs to be adjusted so this pattern is eliminated – either by adding content that avoids this pattern or flagging the content as an undesired pattern.
It’s worth noting that there are areas in which an AI approach may make a lot of sense. Niche specialties where day-to-day variations in workout parameters are more limited represent one such case where a reasonable training database is more easily built, allowing the AI engine to produce acceptable daily workouts.
Ok, so what about using an algorithmic approach to fitness programming? The truth is most high-level coaches employ their own algorithm when they program for their athletes. They likely don’t think of it as such, but they have their own set of rules, typically adopted through both experience and research. Those rules might be fairly basic – for example, the Bulgarian weightlifting program of the ‘70s and ‘80s consisted of maxing out your snatch, clean and jerk, and front squat twice a day, five days a week – or maybe a coach might build a spreadsheet to ensure their athletes hit an exercise frequency and volume that they believe is required to be successful over the coming months or even years.
Regardless of the simplicity or complexity of the programming, unless a coach is pulling workouts out of thin air or randomly finding ones on the web (and yes, both of these happen), a coach is building their athlete’s workouts based on a sort of internal algorithm. The most knowledgeable coaches can elaborate at length on why they make specific choices – the equivalent of your math or physics teacher requiring you to show your work. Others – including very talented coaches that have trained athletes to very high levels of athletic success – are less in tune with the mechanism of their internal algorithm and take a more holistic approach. (“He needs more back squats right now!”). But regardless of how a coach perceives their own process, they are employing an algorithm of their own making to train their athletes.
A big drawback to relying on a coach’s internal algorithm is that, as humans, we all have biases and make mistakes. For instance, we might stop programming wall balls because someone dropped one on our head during a workout, and now we hate them. Or we have a rough week due to personal circumstances and forget to reference the athlete’s exercise spreadsheet (be it in our mind or on the computer), and now the intended frequency and volume targets are off.
As you might have guessed by now, Deep Athletics uses the algorithmic approach to generate fitness programming. The seed that eventually grew into what is now Deep Athletics was Aaron’s observation of his own fitness programming patterns and mistakes. Like any experienced coach, Aaron developed a basic approach to programming that he applied to each of his athletes, including laying out the entire 16-18 week cycle on a (very large) whiteboard.
But he was keenly aware that, being human, he would at times screw things up. Like forgetting to program certain equipment or having a planning session take a week instead of 3 hours because of constant work and family interruptions. Deep’s algorithm was built to harness the computer’s computational abilities as well as limit failure by external factors (other than a power outage, which we would hope is temporary).
In a follow-on blog article, we’ll look at how we went from that original seed to where Deep Athletics is now. Until then, get out there and get after it!