OpenAI CLIP Explained | Multi-modal ML
Vložit
- čas přidán 29. 08. 2024
- OpenAI's CLIP explained simply and intuitively with visuals and code. Language models (LMs) can not rely on language alone. That is the idea behind the "Experience Grounds Language" paper, that proposes a framework to measure LMs' current and future progress. A key idea is that, beyond a certain threshold LMs need other forms of data, such as visual input.
The next step beyond well-known language models; BERT, GPT-3, and T5 is "World Scope 3". In World Scope 3, we move from large text-only datasets to large multi-modal datasets. That is, datasets containing information from multiple forms of media, like both images and text.
The world, both digital and real, is multi-modal. We perceive the world as an orchestra of language, imagery, video, smell, touch, and more. This chaotic ensemble produces an inner state, our "model" of the outside world.
AI must move in the same direction. Even specialist models that focus on language or vision must, at some point, have input from the other modalities. How can a model fully understand the concept of the word "person" without seeing a person?
OpenAI's Contrastive Learning In Pretraining (CLIP) is a world scope three model. It can comprehend concepts in both text and image and even connect concepts between the two modalities. In this video we will learn about multi-modality, how CLIP works, and how to use CLIP for different use cases like encoding, classification, and object detection.
🌲 Pinecone article:
pinecone.io/le...
🤖 70% Discount on the NLP With Transformers in Python course:
bit.ly/3DFvvY5
🎉 Subscribe for Article and Video Updates!
/ subscribe
/ membership
👾 Discord:
/ discord
Thanks for reporting, explaining and lastly opening up recent ML!
I found clip to be very interesting since I always frowned at the lost potential of two different embeddings being arbitrary and methodically separate. This is huge!
yes there will be plenty more on CLIP and other similar models very soon - some of stuff I've built (and will demo) is awesome and nothing more than zero-shot CLIP, excited to share!
This was really excellent - some of the pieces are starting to make sense
Thank you so much for this great walkthrough! Looking forward to more
Great video! Looking forward to your next video diving more into using CLIP for zero-shot classification!
Me too, it's fascinating. Thanks for watching!
Nice video and explanation! I think on min 28:45 you plotted cos_sim instead of dot_sim!
This is amazing James. Thanks for the detailed explanation. I am excited for the future CLIP videos 🙂.
Thanks Ashraq! As you know, I'm excited for them too
Thanks James, very good video about CLIP. Funny thing is that you display twice the cos_sim, so the second time it is not the dot_sim which is displayed. And you fighted to find any difference between the two similarity matrices. LOL 🤣
ah did I do that, oops 😅
Great video. I think you may be plotting the same graph twice though (cos sim). In practice it is almost the same though it would seem.
Great video really ! I have just one thing to say, you should let the images longer in the screen I had to pause the video multiple times to be able to understand them
Thanks Gabriel, I head the same from another viewer - will do this going forwards :)
Excellent video
Really liked the content...thanks for sharing
Thanks for watching!
Excellent content! As a suggestion, can you please keep the images/diagrams a bit longer? They move pretty fast in the video, which means I'll have to rewind the video every now and then.
Sure that’s great feedback, thanks!
Excellent explanation! We can build a CZcams video search engine powered by clip, perhaps you can iterate on the Nlp CZcams search video you did?
That's a great idea, but it might be difficult for CZcams videos where it is just someone talking, as the image embedding would just be something like "a person talking"
Possibly it could be interesting to embed both the text + images with CLIP, and maybe even an averaged text+image embedding for parts of videos where both the speech + image are important.
I will think about this more, it's a great idea so thankyou!
Thanks. It is very informative. Can you pls explain and teach us how to do fine tunning on the custome dataset. Pls
Thanks a lot!
Thank you for the good explanation, if we have 2 different embeddings like texts and 3D images, we can use CLIP to predict images?
fantastic stuff!
Nice explanation!
is there a hosted API for clip where you can provide your image data and get the vectors instead of having to host it yourself, kinda like how you give an input to `ada-002`?
10:23 I believe CLIP is an abbreviation of Contrastive Language Image Pretraining
Plz post on Deep Reinforcement Learning tutorials & projects with python !
Eventually I’m sure I will, RL is very cool
can you give the github code please
Transitions are too flashy and triggering to my eyes. Good explainer however.