How to check if a neural network has learned a specific phenomenon?

🤔 In this video, Ms. Coffee Bean and I explain how "probing" neural networks (in NLP) works. In other words, how we check if a neural net trained on task A is also able to perform task B.

We are also touching upon a very nice paper, recently published by Elena Voita and Ivan Titov. Since I strongly recommend anyone to read it, here is the reference to the explained paper:

▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀
🔥 Optionally, pay us a coffee to boost our Coffee Bean production! ☕
▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀

🔗 Links:

#AICoffeeBreak #MsCoffeeBean

[DavenH]

Informative video, I'm gonna watch all your content now! Also thank you for getting 'phenomenon' and 'phenomena' right. A pet peeve, but nearly all native English speakers mess it up!

[gergerger]

Just stumbled across your channel. Very impressed with your videos! Happy to subscribe :)

[TheRelul]

very nice Romanian accent! Subscribed

[orjihvy]

What is the difference between probing and fine-tuning in transfer learning?

[chinmayapani]

Hello Letitia Your explanations are excellent. It keeps the leaner like me watching your videos all the way through.
One doubt i have related to this video.
You mentioned, bobs pass the data in a compressed form to alice, alice decodes the message and reconstruct it to get the label and if alice able to get the right label then she gives credit to bob.
The question is how alice get to know whether she has decoded correctly?Where is the true label to compare ?

[kfliden]

Thanks, first video on probing that makes sense to me. But just wondering if probing is just for diagnostics or it's actual option for fine tuning in production?

[Youkouleleh]

Hello, I have some question:
1) 3:30, you remove the layers that were responsibles to classification et replace it with probling layer, but how do you know how much layer to remove, is there a rules of a method to precede ?
2) is it the same than transfer learning or is it difference btw what you present and transfert learning ?

For the overfitting:
- if you divide your set with a train/test, will that be enough to check if their is overfitting ?

Thansk for the video

[safaelaat]

Thank you very much, your videos are very helpful for me!! Could you activate the automatic translation please?