AI Paradox Explained - How Overfitting Makes Your Model "STUPID"
Hilarious real-life example of overfitting
I’m going to share an interesting, real-life example of why too much training can be bad for an AI model.
It is easy to assume that the more information you use to train a model, the better it will be.
But this is totally wrong.
My goal was to train a custom AI image model on a particular brand of vodka.
Here is the real image of the vodka bottle.
This is what the vodka bottle is supposed to look like.
I trained 2 custom image models on it.
For the first model, I used 10 different images of this bottle of vodka to train the model.
In the second model, I used 30 different images to train the model.
Which model was better?
It depends.
The second model (the one with 30 training images) was able to generate a more accurate bottle. The label and logo were more likely to be correct for example.
BUT…
The first model was able to generalize better in unique situations.
Look at this side-by-side comparison.
What do you notice?
This is a black-and-white photo, so the bottle should be in black-and-white too.
The first model (trained on only 10 images) was able to correctly generalize and make the bottle appear in black and white.
The second model (trained on 30 images) was “too stupid” to know what to do.
The second model made the vodka bottle look more accurate (look at the training label).
But the first model was “smarter” because it understood that it needed to generalize to a new situation.
👉 Sign up to our free 5-Day email course to grow 🚀 and earn💲in the AI age
This is a perfect example of “overfitting”
Overfitting is when a model learns too much.
A model that is overfitted or overtrained can’t generalize in new situations. It simply memorizes too much.
This example shows that more training isn’t always better.
We want the model to learn what the bottle looks like, but we also need the model to “think” about what the bottle would look like in a new environment.
If I want to train a perfect custom model on this vodka, I would need to find a sweet spot between 10 to 30 images of training.
Also, I should take a photo of this bottle in different lighting and different situations so the model could learn.
If I included a black and white photo in the training data, then the model (even a 30 image model) would “know” what to do in that situation.
Here is an example of the “sweet spot”. The bottle looks great AND the model was able to generalize to a new lighting situation. Look at the purple light coming in through the bottle.
This lesson in overfitting and overtraining is a perfect example of overfitting.
It is easy to see what overfitting is when you see it visually like this.
But overfitting applies to all kinds of training, even with text.
It is a good reminder that more training is not always better. And a wide variety of good data is necessary if you want a model to be “smart”.
This entire example is from our new project. We are making the best software to train a custom AI model on images of yourself or your products.
You can try a few examples for free at Cartario.
Check out Cartario — The Best Custom AI Image-Generator
Follow us on Instagram
👉 Sign up to our free 5-Day email course to grow 🚀 and earn💲in the AI age
Follow me on LinkedIn
👉 Sign up for our newsletter to get the latest tips on how to use ChatGPT and boost your productivity.
Check out our new YouTube Channel
Follow us at our website: AI Growth Guys
Interesting. The first thing I noticed about the two images was that the one on the right had too much clarity in the background, to the point of distraction. A human knows to focus the eye of the audience on the product, not the "anonymous" city behind it. The bars on the window seemed obnoxious, too.