As someone who does ML for a living, I’m curious what insight you think this is going to give versus the models we already have from the exercise physiology literature. Xert have taken the aimless ML approach and basically ended up with approximations of existing physiological models that everyone else uses.
Back on topic, I like the interval discovery for breaking down the main work intervals. It mostly works well for me and I like seeing things like average power and heart rate across each interval, since the averages across the whole workout aren’t really useful to me.
W’ plot is interesting for anaerobic workouts, just to see how empty the tank was. There’s a chart in golden cheetah that tracks time in W’ bal zone, basically dividing equally into four levels of fatigue. Accumulating greater workload under higher levels of fatigue is difficult to visualise from power alone.
One of my most common used features is editing the ride data directly. Simple fixes like filling in HR or power where the signal dropped and there are 0 values. It’s a bit laborious clicking the next page button hundreds of times - maybe a button to jump to the next 0 HR/power/cadence value?
I have yet to see anyone in cycling attempt to use ML to identify linkages between training and performance, even though it has been used in swimming for many years.
I would say that soccer, swimming, American team sports, and the military are all doing a better job of pushing forward on how to use all of the data that modern microsensors can provide. Cycling analytics, OTOH, have been stagnant for almost 20 y.
so what exactly are they doing? do they have metrics like training stress? I know soccer players wear trackers during games but what are they doing besides tracking HR and distance they run? my impression is that cycling actually has the best performance management analytics, or at least the most accessible by lay people
thanks, but the issue is that stuff is super proprietary and I would bet really imperfect. In Boston the sports talk revolving around the Red Sox is this push and pull between the analytics and the baseball people
anyhow I feel you were a bit harsh about to dismiss intervals.icu as if creating ML is that easy (we know TR are working on it and are probably having a heck of a time trying to get it right). power and HR and CTL, TSS, etc remain popular because they work pretty darn well, and best of all it’s accessible to average joes like us
My point is simply that I don’t see anyone involved in cycling attempting to push the boundaries. Instead, everything is just a rehash of the same ol’ same ol’ HR/lactate/power (pick one) paradigm that we’ve been living with for the last few decades. There are lots of skilled coders out there, but where are the true innovators? Why does everything have to be a clone/rip off of TP/WKO4? Surely somebody out there is capable of an original thought or two?!?
Just started using intervals.icu recently and love it. I was also doing a free trail of Training Peaks and prefer the concise and easy to navigate views in icu.
Features I’d like to see:
Efficiency Factor (EF) data field: normalized power / avg HR
Variability Index (VI) data field: normalized power / avg power
Aerobic Decoupling data field for intervals
Data field legend on the left side in the activity summary page
Fitness level in the week summary on the activities page
One of my most common used features is editing the ride data directly. Simple fixes like filling in HR or power where the signal dropped and there are 0 values. It’s a bit laborious clicking the next page button hundreds of times - maybe a button to jump to the next 0 HR/power/cadence value?
I have put this on the todo list (maybe allow spreadsheet download and upload for bulk editing?). Intervals will automatically fix short drop outs in HR and power. It doesn’t change the data you see on the charts but uses the “fixed” data for calculations. It also fixes power spikes and those fixes are indicated on the charts.
I have yet to see anyone in cycling attempt to use ML to identify linkages between training and performance, even though it has been used in swimming for many years.
Alan Couzens is attempting to do exactly that using neural networks: https://humango.ai
I need to get all the basics in play in Intervals.icu first.
Awesome website - thankyou @davidtinker for producing it and keeping it simple. It has everything that I need and helps me structure my training and give me visibility of when I need to go harder and when I need to back off. Call me old fashioned but I would like to be able to export data into excel a little easier - specifically the activities over time.
The feature I love the most is getting that email when you break a power record - now doesn’t that make you happy!
And not everyone has $500 to drop on a powermeter… and some people also do other sports
Before power meters existed, people didn’t perform?
Yes, new technology and much better data can be acquired in this day and age, and it’s even made accessible to the amateur.
BUT, realistically, for a lot of people that kind of money is better invested on a coach, or better nutrition, or other things that make them perform better until they might reach a level where a power meter is almost “necessary” to see true gains. Then it’s truly worth it to have such a specialized tool
Obviously this is my opinion. So take it any way you want it !
This is a bit of a misrepresentation of what Xert is. Although advanced optimization and analytical models are used, which could potentially be classified as “ML” in a general sense, they really don’t fall into the more common definitions of ML, such as a NN for example. There are some fundamental issues with applying ML to training data (see Modelling Athletic Training and Performance: A Hybrid Artificial Neural Network Ensemble Approach). Xert’s analysis for points-of-failure together with a multi-dimensional quantification of training load is likely a better source of information for a NN. These are unique to Xert and nothing like what “everyone else uses”. Stay tuned for more from the Xert team…
yes I agree, all the features discussed above are great, in addition I forgot to add they have an awesome new “ask a coach feature” a bit like TR podcast but personalized! and maybe I am slow but I love to see gradient and altitude separate in the rider analysis. One thing I wonder though is if automatic interval detection should work on more than just power…eg cadence. For example I might do an interval at 100rpm but power stays about the same? Just a thought.
It’s a great website,can’t believe some of the negative comments on here. One feature I’d like is to be able to edit your FTP or HR threshold by date, some of us aren’t great at doing it immediately when it changes, so being able to set FTP at X amount on Y date and have the website update your activities accordingly would be fantastic. Apart from that well done on producing a great resource for us all, good luck for the future as well.
