Robust learning doesn’t occur in perfect conditions.
Exercise performance depends on a chain of factors that work together to make the body deliver. However, the more factors involved, the more fragile the output becomes in the absence of those factors. You can become strong inside a very specific configuration of gym, timing, and running route until one element shifts: a different rack, forgotten headphones, a new running route, and suddenly the session feels off. This failure mode is called overfitting and is well studied in machine learning systems. Fitness has the same vulnerability. Let us examine how.
Whether you hit your performance goals typically depends on a small number of structural variables: the stimulus from training, adequate sleep, sufficient nutrition, and recovery. Their importance varies, but training stimulus and nutrition carry significant weight. If we represent this as an equation, it might look like:
Performance = β1(Training Stimulus) + β2(Sleep) + β3(Nutrition) + β4(Recovery) + ε
Each feature has a coefficient. The larger the coefficient, the more influence that feature has on the output. This mirrors the linear equation that underpins much of machine learning:
Y = β0 + β1X1 + β2X2 + β3X3 + ... + ε
In practice, however, your performance equation often expands. Music enters. The exact rack position. Time of day. Crowd energy. Supplement timing. Familiar equipment. The model of you starts to resemble:
Performance = β1(Stimulus) + β2(Sleep) + β3(Nutrition) + β4(Recovery) + β5(Music) + β6(Rack) + β7(Time) + β8(Crowd) + ... + ε
If high-output sessions consistently occur with music playing, the coefficient on music quietly increases. If you always train at 6am, time of day gains weight. If rack three feels right, it becomes predictive. Then one day the music is gone, the rack is taken, or the schedule shifts. The equation changes not because your capability disappeared, but because your performance had become dependent on variables that were never structural to adaptation. If your output collapses in their absence, the system has overfitted to its environment. The larger the coefficient, the greater the instability when that feature disappears.
If your performance collapses outside routine, you trained the routine, not the body.
Overfitting in training operates at two levels. The first sits close to the movement itself. You squat in the same rack every week. You run the same route. You train at the same hour, after the same meal, following the same warm-up. These patterns reduce friction, which is useful, but they also narrow the conditions under which your strength expresses itself. When someone else occupies your usual rack and you move across the gym, the lift feels unfamiliar. The load is unchanged.
The second level is peripheral. Forgetting your AirPods. Training without your usual playlist. Missing a single dose of creatine and assuming performance will drop. Using a different bottle. Physiologically, little has shifted, yet the prediction of performance recalibrates downward. The body learns that output usually co-occurs with certain cues. When those cues disappear, confidence and drive adjust accordingly. Strength becomes conditional.
How overfitting is tackled in machine learning is almost identical to how it should be treated in fitness: you either reduce your system’s dependence on non-key variables, or you deliberately zero some of them out entirely, or both. In ML, this is called regularisation, making the system adaptable such that a change in data features, gym variables such as environment and peripherals, does not alter your output. Run a different route. Lift in a different area. Train without music. Shift your time from mornings to nights. When output drops sharply after a small contextual change, you have identified an inflated coefficient. If you cannot perform at night because you always trained in the morning, your system has over-adapted to mornings. Training more nights reduces that dependency. Over time, with repeated exposure to varied conditions, the body recalibrates. Peripheral features may still assist, but they stop dictating.
There is a paradox here. One might remember the jack of all trades idea and assume that range weakens mastery. Except that it does not apply in the same way to the body. The body is too physical, too materially responsive, not to adapt when presented with controlled range. Mastery is still king, but range wires into grounded mastery rather than competing with it. Mastery in strength comes from range, not customisation.