AI Training - Will it work?

[Nate_Pearson]

Everything you listed should be accounted for today with AT.

But of course, if someone wants to do what they want to do, then we can’t stop them.

A far as other people doing “AI Training” I like to think of the cooking analogy.

Machine learning output is a combination of data + data engineering.

Just like a good meal is a combination of good ingredients and a good chef.

Data = Ingrendients
Chef = data engineering

You need both to have a great outcome. You can’t simply say “Well, this restaurant cooks food too! It’s already been done!” and expect it to be the same.

[Nate_Pearson]

@bloya89

Personally I think AI is useful to a format like TR for “optimizing” a product that is based on general application to a broad group of athletes. TR plans are a starting point/guidance. But the canned plans need to be individualized to the particular athlete. AI can never truly accomplish that because it can’t replace an athlete’s intuition regarding what their own body is telling them. I don’t think all the metrics and algorithms in the world can ever replace that. But I’m a skeptic by nature so open to reasonable arguments to the contrary!

What if you also told us what your body was telling you?

[Nate_Pearson]

@edistoed

What I envision as a perfect platform would be one that uses machine learning against a large set of data to set up a plan based on whatever multitude of existing factors/fitness it can gather. Then adjusts based on on-going factors, including the previous prescribed ride performance and other inputs (thinking stuff like HRV, etc., ala Whoop), all towards meeting whatever improvement goals you may have.

I.e., the next best thing to actually having a coach.

The goal is to beat a coach in terms of picking workouts.

We’ve already seen this on the closed beta. A coach had a “breakthrough” workout assigned to an athlete. They showed the progression data to the coach and they adjusted the workout to what AT suggested.

Coaches will still be very valuable for accountability, skills, racing tactics, and a sounding board. I don’t see them going away but I see them being used in tandem with AT.

[Nate_Pearson]

@splash

There are problems with TR that ML could likely help with fairly easily. For example, it could predict who over-, or under-tests with the ramp test, and then swap the test out in your training plan or adjust the FTP result to create a better match.

This already happens with the progression system. Except instead of changing your FTP it puts you in a different point in your progression.

@splash

It could also probably find out what type of workout progression usually works best to achieve a specific goal (eg raise FTP or vo2max), depending on your past training history or the type of rider you are. It could likely create types of riders and identify them.

This is happening in our internal system with FTP predictions. I suspect we’ll run future models to optimize certain levels in specific workout zones. This will automatically take into account repeatability which is awesome.

@splash

The main problem with all computer approaches is that the data to feed the model has to be good. The other issue is communication - did you stop the ramp test, because you couldn’t go on anymore, or because your child just burst into the room? How does the computer know?

This is what the survey data post-workout is for. If you stop early we get to know why you stopped. This is also why we can’t have the system tell if you’re overtraining at the moment because we can’t tell why you stopped a workout.

[Nate_Pearson]

@stevemz

Will it work? It depends on how you define “work”

I think the real question is; is it better than what we have today? And then is next month better than last month?

For TR employees, the answer is it’s way better than what we have today. And I know it will continue to be improved.

It’s not a binary “Yes” or “No”.

[Nate_Pearson]

@AlphaDogCycling

On the ramp test: TR is using a single % of max 1 minute power for everyone. It would be interesting if you could apply AIML to get a better individualized estimate of what the “right” % of max 1 minute power from the ramp test to use for me to set my FTP for training.

We just adjust your progression level so you get the correct workouts afterwards.

We like this approach because your levels during a phase will also change a bunch. This makes more sense compared to constantly changing your FTP.

You can think of FTP more as “Ramp Test Result”.

[Nate_Pearson]

@AlphaDogCycling

Similar to the above: a good first step would be individualized CTL & ATL decay parameters instead of the 42 & 7 that are used today.

CTL and ATL so far haven’t been as useful as other things in our ML model.

It’s because TSS doesn’t tell a big enough picture. IE 4x10 min at threshold is the same as 1x40 in terms of TSS but are very different workouts.

[Nate_Pearson]

@Winter2020

I agree there are probably some interesting trends in the data (for instance, I’d love to see if greater compliance with their plans correlates with greatest improvements in FTP), but I wonder how reliable the TR data will be for anything other than inferring how best to implement TRs existing plans. The TR platform is designed to have people follow their plans and its data will reflect this. This data can be used to show relationships between people who have succeeded using their training plans but it can’t show that someone would respond better to a polarized approach, a different distribution of Z2 vs SS rides, or anything else that outside the prescribed TR plans.

There are orders of magnitude more data for people not directly or not following plans at all.

Very rigid plan compliance over a season is extremely rare. And half of our users don’t follow a plan at all. Then we have all of the outside data plus people leaving and rejoining TR seasonly.

[Nate_Pearson]

@hdas

TR staff (Chad) confessed that there’s no evidence whatsoever that that plan is any good, but people requested it. This is one of those instances where you see comercial interest getting in the way of proper training.

What?! Please link me where we said this, and I will preempt and say if we did say this we were wrong.

I’m going to share some SSB HV data tomorrow on the podcast. It’s very good at making some athletes faster. It’s also only for around 7% of our athletes.

We definitely have a problem with athletes picking plans that have too much volume for them. We need to put more guard rails into the system to help them self select down.

There is a vibe on the internet that “if the plan didn’t work for me, then it doesn’t work for anyone.” That is not the case.

[Nate_Pearson]

@Brennus

The ramp test is a great diagnostic test…but it is almost true by definition that it gives many riders the wrong FTP. The ramp test gives you your MAP. Your FTP is estimated as a portion of that MAP. @Nate explicitly said on the podcast that FTP is distributed around 75%. If you’re a rider that is a few standard deviations away from the mean that’s bad news.

The level system takes this into account.

I think you should think of it as a “Ramp Test Result” rather than your FTP. That helps us get out of all of these “what is your FTP and how to measure it” debates.

It’s more of “What can you do, and how can you improve it”. The ramp test gets you on the way to getting there and it’s fairly repeatable.

[stevemz]

Your quote makes it seem like we disagree when in reality we are more or less saying the same thing.

My follow up post covers it AI Training - Will it work? - #105 by stevemz

[Nate_Pearson]

@Bbt67

@iamholland

It’s explained in the podcast and explicitly how TSS and IF don’t work. The example given is how a 5x10 is the same as a 1x50

But the TSS and IF are not the same in this or any other similar example… this did make me laugh during the podcast, smoke mirrors and marketing. It was a statement that hadn’t been thought through imo, because it clearly is not true.

PS That isnt what was said.

If you have a 90 minute workout and all the non-interval periods are the same, then 5x10 is the same IF/TSS as 1x50. At most there might be a 2-3 difference depending on the NP window and where you put the workouts.

It’s definitely not enough to demonstrate the difference between 5x10 and 1x50.

[Nate_Pearson]

@simonicusfacilis

Different NP

It’s the same or very close NP.

[Nate_Pearson]

@Bbt67

This might be true but what percentage is useful for AT and more importantly how do you identify and create a subset of data that can realistically be used for AT? It must be pretty difficult or at least time consuming.

As of now, all of it is used for AT or at least looked at. We try to identify and throw out rides with faulty equipment; ie sustained 2000 watt spikes. Stuff like that.

And yes, it was totally time consuming! This is one of the reasons why we’ve been doing this for three years.

@Bbt67

It also worth noting that pretty much none of this data has subjective feedback, that will start to be collected going forward. I would suggest this is true for a lot of ML features particularly the more advanced ones and it is going to be a fair while before there is a decent data set for these.

Yes, subjective feedback should improve the model. But that doesn’t mean there’s no value in the data today.

[Nate_Pearson]

@Bbt67

My point is there is a lot of potential ML features they have no data for at all, they are going to have to find ways to collect it. I think it has been mentioned in a round about way during the youtube video and in various threads.

The nice thing is we have a lot of athletes and can operate at scale.

[AlphaDogCycling]

Interesting. I would love to see what you can share on the predictive (or it sounds like non-predictive) value of these two metrics, as they are so embedded in the cycling training culture. So weaning people off of these will take a lot of education.

[Nate_Pearson]

Yes, we 100% agree.

The weird part is your TSS might not be as high with AT but you’ll end up being faster. This has been a mind warp for some internal athletes.

[AlphaDogCycling]

Exactly this. I’ve seen this other places with ML: the answer can be very counter intuitively, and it isn’t always obvious how to explain the answer / direction it gives you to draw human understandable “heuristics”.

This is going to be the other challenge: AT’s recommendations will be a true “black box”, without human explainable reasons for the recommendations. So comparing what it is recommending vs. what a human would recommend and why isn’t going to be easy, or probably possible at all.

[HMG]

We all need a mind warp occasionally

[Zaaaach]

I might have missed the explanation in the podcast, but I see TrainNow, I do not see the “achievable/reaching” bubble and I don’t see the “alternatives” beside “variants”. Is that because they’re coming in the future, or because some of those are reliant on having used TrainNow to select workouts previously, meaning I have to use it at least once to get it to start working for me?