WHOOP | Director of Machine Learning Platform Engineering | Onsite | Full-Time
At WHOOP, we're on a mission to unlock human performance. WHOOP empowers its members to perform at a higher level by providing a deep understanding of the interactions between their unique physiology and chosen behaviors.
We’re looking for a talented leader for our ML platform team. This team owns a suite of tools and services that scientists use during the research, experimentation, and production phases of ML development to enable state of the art data science. You will play a pivotal role in ensuring the reliability, scalability, and performance of our machine learning infrastructure, enabling WHOOP to continue pushing the boundaries of human performance. The ideal candidate has a proven track record leading a cross-functional team, is able to identify new opportunities to leverage our large and growing dataset, and is excited to work with physiological data.
WHOOP | Director of Machine Learning Platform Engineering | Onsite | Full-Time
At WHOOP, we're on a mission to unlock human performance. WHOOP empowers its members to perform at a higher level by providing a deep understanding of the interactions between their unique physiology and chosen behaviors.
We’re looking for a talented leader for our ML platform team. This team owns a suite of tools and services that scientists use during the research, experimentation, and production phases of ML development to enable state of the art data science. You will play a pivotal role in ensuring the reliability, scalability, and performance of our machine learning infrastructure, enabling WHOOP to continue pushing the boundaries of human performance. The ideal candidate has a proven track record leading a cross-functional team, is able to identify new opportunities to leverage our large and growing dataset, and is excited to work with physiological data.
Systems like Hadoop and Spark will run multiple versions of the same task writing out to a FS at once but to a temporary directory and when the first finishes the output data is just moved to the final place. It's not uncommon for a job to "complete" writing data to S3 and just sit there and hang as the FS move commands run copying the data and deleting the old version. It is just assumed a rename/move is a no-op in some systems.
As an avid hiker and runner I'm more excited to see what this will do for things like Garmin smart watches. I run through downtown Boston a few times a week and the accuracy is very bad at times making you look like you run much farther than you do.
That's my point, even if you run at a steady pace the instant pace reported by your device tends to jump around quite a bit, even if they smooth it out over 10s. On my Garmin watch, the instant pace typically jumps around by +-20s/mile, which is significant when you're trying to hit a specific pace target.
They don’t in real running. After seeing his page I got super excited. They’re really accurate at the pace you csalibrate them too, then drift quickly once that pace changes. So if you camlivrat them at 7:30min/mile, then run 8min/mile they’re pretty far off. Especially if you’re a distance running, or looking for accurate pacing. I’ve tested it both on marathons and tracks.
All current running watches and phone software uses single point calibrations. Now if they curved fit 2 or 3 points....
Stryd is a calibration-free technology. Motion capture technology is used to precisely track the movement of the foot at different running speeds & on different terrains.
Do you run the same path frequently? When I do, I keep the path the same, as I pre-calculate the distance, and can keep a watch to track my time. Thus I have no need for path tracking.
A GPS unit is useful for assisting dynamically changing routes (e.g. driving) or to track unknown paths (e.g. plotting your hike in snow/darkness). May I ask what you'd need a fine-accuracy GPS unit for during a run?
For runs it's mostly just for fun. I use Strava so it has sections they identify along your run that they rank you against your friends and strangers. I do end up running very similar routes most of the time.
I mostly find the GPS invaluable in hiking. It can be a great thing to have on the top of a mountain in the winter with low visibility. If you get up to a summit and the weather turns, you can have the watch send you back down the way you came. Nice and safe.
The only deficiency I find is accuracy when you are trying to workout at a specific pace on track.
On a longer run they are probably accurate to within 1% which makes practically no difference but you often find on a track that the inaccuracy multiplies on each lap so if you are doing say 1 mile repeats (4 laps) then by the end of the mile you can be 20 metres out.
It's not a biggie because visually you can see if your watch says 0.23, 0.46 etc as you go past the 402m mark you know that you need to up your pace because it's short but having a more precise distance and therefore pace would save a few brain cells when you're trying to concentrate on hitting your paces.
I try to run different paths as much as possible. My Garmin Forerunner fitness tracker is reasonably accurate at recording the path and displaying an instantaneous pace, but it still jumps around and gives erratic results at times. So better accuracy would certainly be a welcome improvement.
I've run into a severe issue with Fitbit's app on the iPhone 7.
On my older iPhone, the mile counts were reasonably accurate. On the 7, there's a spurious ~2 miles per hour added at all times.
My guess is that the 7's chip updates position much more frequently, so that small inaccuracies (a couple dozen feet west, then a couple dozen feet east on the next update) make it look like I'm sprinting all over the place in random directions at high speed. No idea why Fitbit doesn't smooth it out a little...
Did a six hour, 4.6 mile hike in the Adirondacks and it tracked me at 18 miles (edit: 13, sorry).
The Fitbit app lets you take a GPS track using the phone's GPS (as the Charge HR doesn't have a GPS chip). It's not step-based.
Standing still for a break still makes my mileage go up, and I can see the blue dot on the map in the Fitbit app jitter around a little - a few feet here, a few feet there.
(That's a ~5 mile hike in actual mileage - and you can see the issue very clearly between "miles" 4 and 5 on the map, where we likely took a long break, and in the longer distance between the downhill miles as we were moving faster)
With the existing L1 signal the multipath signals overlap each other and its hard to find out which the original signal was.
The newer L5 signal is much shorter, so multipath signals will show up as individual peaks, and the receiver can simply pick the first signal, cause that will be the direct one.
You can see how off the trace can be here: https://i.imgur.com/eHJI8FH.jpg . You could probably do some smoothing to help fit some of this to a street but I've been on runs where I zig zag down streets just to add distance or to do hills.
Wow, that's pretty horrible, must be giving you close to twice the actual distance. I think Google Maps on my phone applies some smoothing that does this a lot better, but then it can lag for a while when I'm going sharp right in an intersection.
Maybe try telling it that you're biking instead, my guess is they do more smoothing the higher speed they expect.
It's pretty much raw GPS if you're in walking or biking mode. Driving mode uses sophisticated filtering algorithms, and doesn't actually make any assumptions about your speed, though it does assume you're on a road.
I have a question for you. If GPS locations in cities are affected by reflections off buildings, wouldn't the effect be deterministic and could you take advantage of that to reduce the noise?
What makes anyone confident in the accuracy of wearable tracking devices? I haven't seen it verified and I wouldn't assume it: Accuracy is expensive, generally speaking, and few consumers will pay for it or even question it - how often have you heard someone mention it?
Or, for example, have you wondered how accurate your simple bathroom scale is? I looked into it a little, wanting to collect accurate health data: IIRC +/- 0.1 kg (~0.2 lbs) is available in <$100 scales, but for real precision you need the $500 scales at your doctor's office. And how consistent is it? I tested mine, a decent one with good reviews, and getting it to report the same weight in immediately consecutive measurements was a challenge; I had to stand on it in just the right way.
Here are a couple articles that found the accuracy of activity trackers wanting. I've seen other articles in places like ZDnet that found the same problems, though based on less rigorous research.
I'm not sure what the value of a more accurate scale would be for a bathroom scale - your weight varies by far more than 100g just based on what you've eaten and how much water you've retained.
A guess: Two uncertain data points makes it even worse. If you can at least depend on the scale, the error is lower.
I tested our old Nintendo wii (or whatever it's called) balance board and so far it gives the same value every time I step on it. The value was also only 0.2 kg off from the doctors expensive scale.
The > at the end of your sporttechie link is making it 404, but here's the important part:
“What the plaintiffs’ attorneys call a “study” is biased, baseless, and nothing more than an attempt to extract a payout from Fitbit. It lacks scientific rigor and is the product of flawed methodology. It was paid for by plaintiffs’ lawyers who are suing Fitbit, and was conducted with a consumer-grade electrocardiogram – not a true clinical device, as implied by the plaintiffs’ lawyers. Furthermore, there is no evidence the device used in the purported “study” was tested for accuracy.”
Fitbit’s research team rigorously researched and developed the technology for three years prior to introducing it to market and continues to conduct extensive internal studies to test the features of our products. Fitbit Charge HR is the #1 selling fitness tracker on the market, and is embraced by millions of consumers around the globe.
Consumer Reports independently tested the heart rate accuracy of the Charge HR and Surge after the initial lawsuit was filed in January and gave both products an “excellent” rating. We stand behind our heart-rate monitoring technology and all our products, and continue to believe the plaintiffs’ allegations do not have any merit. We are vigorously defending against these claims, and will resist any attempts by the plaintiffs’ lawyers to leverage a settlement with misleading tactics and false claims of scientific evidence.
Good point (and sorry for the link, which I'm too late to fix). However, if we take Fitbit's statement at face value, we should infer that Fitbit is as biased and unreliable as the other side, so I wouldn't write off that research.
I would say it can be used at it's simplest as a replacement for cron. It supports running programs on a schedule and you can set concurrency rules, SLAs, and triggers around even single commands or programs. You also get nice graphs of task run time and email alerts if jobs take longer than the SLA you've set.
In Cambridge, MA one of the local bus' has the power to keep a light green longer than normal with the same/similar system: http://pb.cambridgema.gov/1bus .
I might be making this all up, I'll need to actually test it out now. I swear when watching Netflix on a computer, you start an episode and your cursor is auto-hidden. At some point, maybe even when the second episode starts, I swear the cursor becomes visible again on the screen. I then instinctively giggle my finger on my touchpad to make it hide again. After watching couple episodes in a row and doing that, if I don't I feel like it pauses and ask if I'm still there within 20 or so seconds.
Has anyone else noticed that? My theory is that they reveal the cursor after every new episode to try to get you to move it and hide it. If you do that each time, they could assume you're not paying as much attention.
I was reading through the wikipedia page for the building and it mentions that when it was built there was a bylaw that required kitchens and bathrooms to have a window in them.
I've just started using Deis at work recently. Dokku looks like it was a small project sponsored by the company behind Deis. I haven't looked much at dokku-alt.
The thing that I really like about Deis is it supports deployments off of docker images in a private repo on top of git push deploys.
One of the things we do is host a REST service backed by data we generate. We can have jenkins automatically build and publish a docker container to our internal repo, then it's two REST calls to Deis to update the docker container that is running. We generate the data in Hadoop through a longish oozie workflow weekly and we can make two REST calls to repoint the REST service to some new database tables.
I also like that each layer or part of Deis can be swapped out. Want to run haproxy for the load balancing layer instead of nginx? Someone probably has a container that can be dropped in instead. Want swap a more advanced scheduler for etcd, you can swap in Mesos, YARN, or even Flynn's scheduler.
We’re looking for a talented leader for our ML platform team. This team owns a suite of tools and services that scientists use during the research, experimentation, and production phases of ML development to enable state of the art data science. You will play a pivotal role in ensuring the reliability, scalability, and performance of our machine learning infrastructure, enabling WHOOP to continue pushing the boundaries of human performance. The ideal candidate has a proven track record leading a cross-functional team, is able to identify new opportunities to leverage our large and growing dataset, and is excited to work with physiological data.
https://jobs.lever.co/whoop/18467795-9a63-4208-924d-34eb1d62...