peakers.ai

---

Jim Manton / ERO Sports

https://forum.slowtwitch.com/...0network%20#p6836199

AI? Training? Ya mean like "practice?"

---

"What's your claim?" - Ben Gravy
"Your best work is the work you're excited about" - Rick Rubin

garmin is testing the waters with something like this for beginners. They have a 5k plan that supposedly adjusts to how your hitting the workouts I think.

---

Use this link to save $5 off your USAT membership renewal:
https://membership.usatriathlon.org/...A2-BAD7-6137B629D9B7

Personally, beyond 'real time' adjustments, I think we will find a lot of value in individual adjustments. In my view, the power of A.I. comes in personalization.

We are already seeing this in medicine (e.g https://www.tandfonline.com/...3808993.2017.1380516) The reality is that different patients respond to different treatments in different ways. This reality has largely been ignored to this point due to practical limitations in acquiring enough data to root out these individual differences and limitations in having sufficient computing power and algorithms to crunch said data once acquired. Therefore, at least to this point, patients have been given treatments that work out well for the average of the studied group.

The same is true for athletes.

I wrote a little about what I'm seeing in terms of individual differences in response to different types of training in this post https://alancouzens.com/...l_int_responder.html (are you a volume or intensity responder?) But, in summary, the same is very true. Not all athletes respond to the same type of training in the same way and we need to personalize the prescription to get the best result (in fact, to get *ANY* response at the extreme).

A.I. will (/does) provide us with the tool to crunch the mass of individual data that is now available to create an individual model of the athlete, run various options through the simulated model and tell the athlete what type of training will lead to the optimal response (i.e. highest fitness for a given load). When this goes 'mainstream', it will be a huge step forward on the current paradigm of coaching philosophies built on small studies that find an effect of a given intervention for the group average or the memory limited general recollections of an old coach's brain.

In short, personalized precision coaching will be as much of a game changer in our domain as precision medicine will be to health care.

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

If this forum had an upvote, I'd give that two.

---

Shane Miller - GPLama
YouTube | Web | Twitter | Instagram | Facebook | Strava

There has already been some level of this in recent years. The Oakley Radar Pace, for example, generated training programs for users in the app, then guided the user during the training sessions (interactive dialog system). It also adjusted the plan and thresholds to account for missed sessions or changes in fitness. But so much more could have been done in terms of individualization had Intel (who made the sensors and a large part of the app) had the stomach to stay in the market. The training engine behind it, made by the NZ company Performance Lab, lives on, however, and does have other device partnerships.

I like Alan Couzens' post a lot - while at Intel, I worked to create the fitness/wellness equivalent of precision medicine (without the regulatory hurdle of FDA approval). The idea was similar to what he articulated so much better than I could - that individuals respond to different training stimuli at different rates and with different levels of success for various mixes, so if you could automatically measure user similarity and had a history of prescriptions, compliance and outcomes, you could propose and adjust training routines on an individualized basis based on the experience of athletes similar to you rather than as a one-size-fits-all recipe.

The real win could come in the longitudinal area - the aging athlete. Current aging IM athletes, for example, are trailblazers and could be valuable real-world lab rats, especially if you could combine workout and lifestyle data with health data (securely and anonymously, of course) in some sort of fitness/wellness/health cloud that people could trust.

The ideal company to pursue this would be one who already has access to a ton of user experience - Strava, TrainingPeaks, Garmin, Apple, Google. In the sports/fitness area, those who have actual training programs (TP, TrainerRoad in its limited domain) would be better positioned with experiential data (prescriptions and outcomes) than those with purely physiological data; it disappoints me (as a TP coach user and as a machine learning practitioner) that TP has done absolutely nothing in this area. But the giants (Apple, Google and Amazon) have even more potential to unite fitness data with the much more sensitive health data.

Someone will get this right.

Ian

Much more overrated than you think.

Just look at the precision medicine field - consistently, more and more revelations come out where publications "laying the foundation for groundbreaking discoveries" are later caught doing blatant p-value hacking or just plain error statistics. That doesn't mean you can't build a multi million "business" on top of fake promises which sound good (see Theranos), but it reinforces that point where I'd trust "precision medicine" in practice is as far today as it was several years ago. Of course, if you happen to consult an idiot doctor, the outcome might be even worse, but to choose between human decision error and programmed machine decision error is an interesting question to which I personally wouldn't even try answering unless I knew exactly all the model in question parameters (which, by definition, are always "corporate secret").

Every black box model, call it AI, machine learning or whichever buzzword you like needs either thousands of observations or, if you choose a more classic approach, very strictly controlled experimental design to minimize non-controlled impact on the response you're measuring (and this is assuming you know how to isolate / account for the unwanted impacts). Note, we're talking about "personalized models" instead of population level effects here.

Now ask yourself, how much variation do you have in your day-to-day activities, which you can recall, and how much variation is in stuff you've actually not measured - either because you simply don't do it, or nobody knows you should do it. Even more, your feelings today can have (and do have) tons of lag-type parameters, many of which could be important... or could be not at all; some could manifest themselves in hours, others in days, weeks or even months. On top of it, add the fact while humans like binary answers ("do this workout / don't do it"), the actual impact / response is definitely continuous, and for dichotomizing it to the binary scale, you need to know the loss function (what happens short/long term if I don't do the workout when the model said that I should do with 60% probability?) which, again, is highly individual, but can impact the result of the "AI suggested" program dramatically.

All that said, you can build interesting models ("all models are wrong, but some are useful") anyway, and the best of them probably would perform close to experienced human. But to show that they outperform humans guiding other humans in exercise science instead of image recognition (trained on millions of images) or chess play, we're still quite a journey away from it.

---

----------------------------
Need more W/CdA.

Thanks, Ian, for the kind words and additional perspective. I always get a lot out of hearing from those with experience in the real world ML 'trenches'.

In my way of thinking (and would love to hear your thoughts on the below) applying ML to sports has some advantages and disadvantages over applying ML to medicine...

Advantages...
* A lot of athletes are already data geeks! :-) So, they often bring to the table a lot of data specific to them as an individual Thus, the reliance on 'clustering' athletes into groups of similar athletes may be a little less important than it is in health care/wellness, enabling us to 'jump the line' to more powerful, more individual, models. In health care, (fortunately) most patients will have limited individual histories with a particular disease. In sport, athletes often come 'pre-loaded' with a data history of many individual outcomes and, more often than not, many different strategies employed on the way to those outcomes. I see this as a big advantage.

* Additionally, as you said, we don't have the same regulatory constraints that health care does. In theory, this should lead to a much quicker development of the application of ML to sports. In theory....

Disadvantages
* 'A.I.' in sports isn't regulated. I fear this is leading more to a marketing game than a model accuracy game and athletes are left to determine the true validity of all of these 'proprietary models' in today's coaching software. If companies can push out a simple if-then 'expert' model without having to go through the steps of collecting massive data and developing 'true' validated models & athletes don't care about whether the model is accurate/validated, they will thwart the development of better ML. When we get to the point that athletes care enough about model accuracy that they demand accuracy metrics before 'buying into' the latest, greatest software, (in the same way that, for instance, the use of ML in tumor detection is based on how accurate the computer is at identifying tumors vs a human doctor rather than a snazzy marketing pitch ), I firmly believe the application of 'good' A.I. in sports will skyrocket.

Thanks again for sharing your experience and perspective.

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

One platform not mentioned here is 2PEAK. I have tried Peakers AI and 2Peak and am using 2peak now because it is significantly cheaper and seems to be basically the same idea. Now that I have used 2PEAK more I prefer it.

there's an easy fix to the problems of trust in the cloud and sensitive data, at least for health care - M4A. Once health care is used to make people better, rather than as a profit center, most of the difficulties evaporate. But it hardly matters, since Google has all your healthcare data already, and will do as it will with it.

The other two of the Big 3 have healthcare databases too. So too does Experian, the credit management company hacked by the Chinese to expose everyone's credit rating, which is now also managing your health records. According to their announcement earlier this year,
'every person in the U.S. population, of an estimated 328 million Americans, have been assigned a unique Universal Patient Identifier'.
The ID is 'not intended' to be 'patient-facing' and is 'not known to the patient'.
That's bound to go well.

I've been following the healthcare database story since the earlies, and it's been a shitshow the whole way.. In particular since HIPAA was written before these databases, none of them is subject to HIPAA regulations. So all your data is out there, available to paying subscribers: but not available for fitness applications, since it is after all worth money and why would they let the mere subject have profitable data ?

Bah.

Alan's work is fascinating though.. again the problems are not technical but political..


how many athletes have any idea how to evaluate accuracy metrics ?
Most of us are going to be vulnerable to AI 'special sauce' marketing..

Theres also trainasone....but its correct that its not regulated at all so you have no idea if they have any kind of data governance at all, which is critical when talking about AI.

---

"see the world as it is not as you want it to be"

// artificial is a cultural perspective with a blind spot defined by the core .. intelligence is something we can use in the best case and we don’t invent .. being individual on both sides of communication channels is a spread out field of energy for existence into a open future created on interactions .. machine learning can help us to understand how we teach our self and "qualitify" today and it's not a struture capable to improve intelligence in general .. to get faster the key is using time the shorter way .. & many thanks to all in this thread for the inspiration to the last sentence ..

---

*
___/\___/\___/\___ the s u r f b o a r d of the K u r p f a l z is the r o a d b i k e .. oSo >>

I fully believe it will happen, and that it will eventually have to be regulated in racing.
For instance: Your computer has a preloaded map of the course and knows the profile. It knows the weather (temp), your current output, and your physiological stress (current HR, HRv, and historical models of your body's abilities based on training time) and, if we're feeling fancy, your CDa from a pitot. It then pumps out a real time power number that optimizes all of the above to hit your theoretical best overall number. That's a massive advantage.

I know that might seem unrealistic, but 1) we're already capturing almost all of this data, 2) machine learning is growing at incredible rates (it's in practice, it's not just some theoretical) and 3) 20 years ago we barely had color screens on cell phones. Don't think in 2-3 year leaps, think in 5 or 10. If the market asks for it, this isn't limited by tech.

---

JustinDoesTriathlon

Owner, FuelRodz Endurance.

.. this instance I call experience .. to know & be sure how to prepare best to an selected aim in the time there is before an future event within my all day life by computing calculation is what I'd give the words ai edit: ore ae ..

---

*
___/\___/\___/\___ the s u r f b o a r d of the K u r p f a l z is the r o a d b i k e .. oSo >>

Isn’t this what triathlon taren is providing with his training plans?

No. And I do not believe Taren would be claiming this either.

This is as close to I have seen and used to AI being incorporated into training plan adjustment in real time.

http://Www.tridot.com

Thanks Doug,

I think that is the key to the development of ML. Companies must expose their accuracy and consumers must become a little more familiar with accuracy metrics.

Honestly, if these metrics are provided, the athlete is in an even better position than most ML researchers to assess how good the algorithm is. If I tell you that our model is able to predict your next Ironman power output within 10 watts, you're in a very good position to assess the worth of that algorithm vs a data scientist (whose only reference to watts might be high school physics & light bulbs :-) &, ultimately, to determine whether that error is 'good enough' for you to buy in. This level of visibility is what we are shooting for with our platform & I'm hopeful that others will follow suit.

I'm also hopeful that athletes will see the value of an algorithm that "calls its shots" vs one that doesn't.

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

Not me but some on ST are, which is why I am here :)

---

808 > NYC > PDX > YVR
2024 Races: Taupo

.. reading the descriptions & blog posts on HumanGo.ai it would be interesting if there is an effect on reflection capabilities and easier organization of training aims with help of NN models and ML ..

---

*
___/\___/\___/\___ the s u r f b o a r d of the K u r p f a l z is the r o a d b i k e .. oSo >>

I have had a long time interest in both multi linear modelling and AI I have done some reading and one thing I came across but have no reference to back it up is that with Machine Learning, it does become a black box. Even the original programmers loose the ability to see inside the model once the machine starts to alter the algorithms with time and new data. So as a ML program learns the algorithm becomes more opaque. So the only way to assess the outcomes is against known results I suspect and that would be an interesting feat in the area of human performance.

This will be an interesting area to follow. I have often wondered why people were not doing more multivariate modelling of performance and training. Others on here have made some interesting observations in that area.

We had an interesting meeting this week on that very topic of "organization of training aims". While the training aim for the majority of this population seems obvious...

Race as fast as I can at race xyz given my current life constraints...


there is no reason that our AI model can't be optimized to other ends, e.g.

- select the actions that will improve my mood to the highest level
- select actions that will minimize my chance of getting sick
- select action that will maximize my daily energy etc.

And, as far as 'reflection capabilities', we have a firm belief that "What If?" beats "WTF!" every time i.e. the ability to run simulations of "what's likely to happen if I string these series of workout choices together" is superior to having to physically try out a bunch of different strategies and seeing which ones actually work. Especially for long course athletes looking to reach their potential, there's simply not enough time for a real world trial and error approach to come up with the optimal solution.

---

Alan Couzens, M.Sc. (Sports Science)
Exercise Physiologist/Coach
Twitter: https://twitter.com/Alan_Couzens
Web: https://alancouzens.com

Although many methodologies in deep learning and machine learning are black box approaches, a huge portion of AI is not, and is fully interpretable.

There are many issues with current ML/DL approaches: lack of interpretability, lack of generalizability being the main ones. But there are also new methods being investigated to address such issues (e.g. transfer learning)