I have been taken by a dog :(

Good morning
In the C3_Lab_1_transfer_learning I upload a photograph of myself and the model predicted with 100% probability that I am a dog.
Having certain identity controversies…, I wonder where is the error.

Maybe you look like a dog :grin::grin:, just joking!

Well the NN learns fearures and maybe what it has learned, maybe the features it has learned are such that in the case of your image predicts a dog.

You could train it longer with a larger dataset and avoid overfitting. Try to implememt a train dev test set that come from a same rich and large dataset, hopefully the model will learn better features.

Also you are doing trasfer learning I see, you probably need more images and train the extended network for a longer period as well.

Thank you for your answer.
I’ll try your proposals.
Nevertheless, I think that is kind of “OOV” token. That is, that the model has been trained on a binary basis, or a cat or a dog, and not prepared for the unknown, which, I suspect, should be a good practice. Am I ok?

Yeah the model might be trained on some kind of images and its probably good in detecting those.

The prettained model has learned high level and low level features so its capable to detect lines and other high level features present in all images. That is the point of using a pretrained model making slighlty changes to it and then retraining it on new images to learn the low level features of the new images.

If trained sufficiently enough and using its previous knowledge it can learn new low level features faster than trained from scratch.

1 Like

Hi @Hermes_Morales_Gross,

this thread could be interesting for you since it also touches upon intended use and unintended use of AI models and links to a paper that investigates bias in data and algorithms.

Best regards
Christian

1 Like