Find the Duck - image limits

13 views

Skip to first unread message

Aubrey Sherman

unread,

Apr 21, 2025, 3:40:13 AMApr 21

to Machine Learning for Kids

Hi Dale,

I've been going through all the projects in the No Starch book and am learning so much -- thank you for building ML for Kids!

While working on Find the Duck, I hit the 50 image limit, but I don't see an option to change this in the Restrictions section of my teacher account. With the 50 images, my training model correctly identifies the duck, but still misidentifies ~4 other cells as ducks. I haven't been able to make my model more accurate with fewer images. Is the limit for images unchangeable? If so, do you recommend a different approach for improving my model's accuracy?

I'm also wondering: Given the scope of the project, is it better not to include images in the training data that contain only small features of the duck (e.g. just its feet, or just its tail feather), and aim instead for the largest possible area? I noticed my model will recognize more than one cell as the duck if the image is split (the cell with its head and torso, say, and the cell below with its feet). As I write this out, I'm realizing it's probably good that it catches both...!

Thanks so much,

Aubrey

Dale Lane

unread,

Apr 25, 2025, 8:29:01 AMApr 25

to Machine Learning for Kids

> I don't see an option to change this in the Restrictions

> section of my teacher account. ...

> Is the limit for images unchangeable?

If you store your project in the cloud, I apply limits on the amount of data you can add to the project (as a way of keeping my running costs under control).

If you store your project on your own computer, I have no need to apply any limits, so you can add as many training images as you like.

> As I write this out, I'm realizing it's probably good that it catches both...!

It's subjective - there's no "right" way to do this, and you could absolutely train a model to only recognise squares with the duck covering a majority of the space, or you could train a model to recognise squares with any sign of the duck in it. Both are valid, and potentially useful. But personally I tend to agree with where I think you ended up here, that a model that identifies all the squares the duck is touching is a good approach.