3 ms·
I've been using it all day and have the following feedback. 1. I like the design, and it works very cleanly. Nothing appears to be heavy, and the UI is intuiti
by AndrewKemendo 10y ago
I've been using it all day and have the following feedback.
1. I like the design, and it works very cleanly. Nothing appears to be heavy, and the UI is intuitive.
2. It only does one thing - track food. I think this is it's biggest strength. It doesn't do fitness or anything else right now, which it shouldn't.
3. The default goal to lose weight is simple to use and I think captures the predominant use case - by calorie reduction. I think the goal breakdown with consumed/remaining should be the most prominent UI element on top with the breakdown by calorie type being second. Having the challenge details (# meals X days) at the top isn't data I need each time I open the app. I can see why that would be a design challenge.
4. I have only used it a day so I can't say how well or not it displays trends about calories/nutrition over time, but I know that I would like to be able to break things down more.
5. The biggest challenge I think you have is with CV. Correct me if I am wrong, but my guess is that you are trying to use users to do reinforcement learning on your Deep Vision Nets. I am a Deep Vision guy myself (which is why I downloaded it by the way) and my guess is that you are going to have a hard time doing training this way. Here is why:
A. If the results of the object classification are good enough to always be result #1 (because it's obvious you are using a probabilistic return set (imagenet?)) then over time people will be annoyed at having to select the object/food in addition to taking the picture.
B. If the results are not good, then people will get annoyed with having to take the picture AND ALSO enter the food type. They will just resort to entering it manually each time. For example I made vegetable curry for dinner, and didn't take a photo because I knew it wouldn't know what it was.
So as a result, your training set will stagnate and won't learn any better than if you did it with a team of people. If you want it to really learn you're going to have to incentivize or force people to always take a picture and always tag it. Even better if you can have them bound each item right?!
By the way, crowd sourcing Machine Vision training is I think the right way to do things (that's what we do with interior home objects FYI).
I look forward to seeing iteration here. Best of luck.
- m_ke 10y agoHey, thanks for the thoughtful response. We're not doing any reinforcement learning, we just fine tune the net as we get more data (and occasionally train from scratch when we add a lot of new classes). In regards to (A), we plan to start skipping the selection steps for predictions that we're really confident in and will just add them by default. I think once we have enough data we might even be able to predict what users will eat before they take a picture. I eat practically the same thing for breakfast every day of the week so it could just log it for me without requiring me to do any work. Same goes for things like coffee shops, we don't ask for location right now but if you always get the same thing when you walk into a coffee shop, we could just log it for you based on the fact that you were there. B. We keep track of our predictions and what users end up logging so we can tell what our weaknesses are. When we train new models we prioritize the weakly performing classes, especially if they're popular among our users.