Beyond RAG: New Continual Learning of LLM w/ InCA

Поділитися
Вставка
  • Опубліковано 1 січ 2025

КОМЕНТАРІ • 43

  • @code4AI
    @code4AI  4 дні тому +9

    With the automatic audio dubbing from UA-cam /Google you hear a synthetic voice in your regional language.
    To hear my original voice in English, switch to "Default" or "English" in the settings. Thank you.

    • @wwkk4964
      @wwkk4964 4 дні тому +2

      Thank you so much for all your hard work! Do you believe the INCA setup could be reminiscent of what a recurrent neural network might do sometimes in some narrow domains (of course, in a highly tortured and deformed metaphor only)

    • @javaboy6581
      @javaboy6581 Годину тому

      Por qué no hay idioma Español en las traducciones?

  • @IvarDaigon
    @IvarDaigon 4 дні тому +31

    Checking account balances is a bad example for a RAG system or InCA because it can be achieved a lot simpler and more securely with function calling directly on your data source.
    RAG is better suited to the task of fact checking and not data lookup.
    Having people's private data like account balances, names, address etc encoded directly into a model or a RAG vector store is a very bad idea because once it's in there it is extremely insecure and hard to remove.
    A much better example would be a comparison of how it accurately retrieves ambiguous lookup terms like "Kiing Charles" on sites like wikipedia of which there can be many many (historical) individuals who all share the same name.

  • @_ariosto1519
    @_ariosto1519 4 дні тому +7

    Regarding 2:35, you can split electronic health records in identifiable (PHI) and not. A good solution is to de-identify PHI, for example, replacing dates by “age at”, limiting address to zip code, etc. By doing so, you remove the risk because all data in the RAG is non-PHI. Thank you for your great videos and continuous posting! 😊

  • @mohamedkarim-p7j
    @mohamedkarim-p7j День тому +1

    Thank for sharing👍

  • @ArulKogilan
    @ArulKogilan 4 дні тому +18

    Can you include links to download the relevant papers mentioned in your videos

    • @code4AI
      @code4AI  4 дні тому +5

      Links are included in the video.

    • @CaridorcTergilti
      @CaridorcTergilti 4 дні тому +11

      ​@@code4AIThey should also be in the description to be easier to click

    • @luiztomikawa
      @luiztomikawa 3 дні тому +2

      @@code4AI Where?

    • @rmt3589
      @rmt3589 3 дні тому +2

      Just Google the title, more detailed than MLA format.

  • @HunteronX
    @HunteronX 3 дні тому

    Really like this idea, thank you.
    Is like a concept abstraction over any text that comes in. If you create enough classes, each can represent an action to take place in a workflow, dynamically.

  • @fdavis1555
    @fdavis1555 4 дні тому +1

    Very helpful and delightfully presented, as usual!

  • @jancijak9385
    @jancijak9385 3 дні тому +2

    As an research engineer working on similar systems, I can see how ECL, can improve existing prompt input of users. The issue is that this in no way solves catastrophic forgetting in long context. It would just improve the short term alignment to user query. Maybe I am missing something. Can you reference the paper?

  • @r.c.z.7139
    @r.c.z.7139 4 дні тому +5

    RAG is insecure: 2:30 - no this is not true. While you will get the nearest result using cosine or another metric you can and should always check if the data you get, should be processed in the current security context. Leaving out that step is simple bad design and not something inherent to rag.

  • @stefano94103
    @stefano94103 4 дні тому

    Your information is always next level! Awesome!

  • @araobenjamin
    @araobenjamin 2 дні тому

    Great video! I've just subscribed, thanks for the work, I'll definitely read the paper and try to implement it.
    The only thing is that this model looks great as an evolving classification model, not as an intelligence enhancement model.
    Not a bad thing, just that it has its place that is not of a RAG.

  • @actorjohanmatsfredkarlsson2293
    @actorjohanmatsfredkarlsson2293 3 дні тому +1

    Problem with InCA is that its only for classification tasks. It only work with supervised data sets. So it doesn't replace RAG yet. If it could be modified to classify it's input in an unsupervised fashion it could be extended to do "RAG" type of retrieval (it still be RAG but wouldn't need a vector store nor reply-play).

  • @TheEtrepreneur
    @TheEtrepreneur 4 дні тому +1

    so InCA is basic (almost deterministic) programming. 💪💥 Great exercises by the way!

  • @michaeltse321
    @michaeltse321 3 дні тому +1

    This is like short and long term memory. Long term memory is built up in layers and during sleep, memories from short term memories are added to long term memory in a way that doesn't destructively alter previous memories.

  • @claxvii177th6
    @claxvii177th6 4 дні тому +1

    it's warm and sunny here in the the tropical coast of south america! cheers!
    maybe today there is something i can try out!

  • @antoniovieiradasilvajunior666
    @antoniovieiradasilvajunior666 3 дні тому +2

    If INCA+ECL was made to text classification how these can substitute a RAG application ?

    • @code4AI
      @code4AI  3 дні тому

      I explain this point in my video ....

  • @wiktorm9858
    @wiktorm9858 4 дні тому +3

    Can someone write in short, how this method work and if really works? 😮 Thank you! 🎉

    • @code4AI
      @code4AI  3 дні тому +5

      You mean to watch a video is too complex? To read the scientific papers presented in the video is too boring? To understand the content of this new method is not worth it? You just want a one-liner? Okay. This video introduces InCA, a novel method that leverages Large Language Models through continual in-context learning (C-ICL) with a unique external module (ECL) for dynamic adaptation. Current AI models struggle with continuous learning, often forgetting old tasks when learning new ones.
      By using statistical models of semantic tags, InCA achieves efficient class selection, avoids catastrophic forgetting without storing any previous data, and presents a unique alternative to conventional methods based on parameter updates and even retrieval-augmented approaches. How does it work and is it better than RAG? Find out.
      A key takeaway from paper is that the most important aspect is to have more effective and optimized prompts rather than focusing on gradient based optimization methods with parameter updates. It emphasizes that the proper design, modeling and formatting of prompt by the usage of the ECL method, improves the results more than any model parameter updates.

    • @ThePARKBENCH
      @ThePARKBENCH 3 дні тому

      OMG, too funny... I adore your reply and sense of humor, my friend! ❤ I will be restructuring my UA-cam and website topics/apps (available) for my viewers. Thank you, Sir 🎉

  • @Swooshii-u4e
    @Swooshii-u4e 3 дні тому +1

    And ecl is coming from Intel? I heard you say "Intel" a couple times

  • @CharlotteLopez-n3i
    @CharlotteLopez-n3i 4 дні тому +2

    Intriguing: InCA focuses on prompt design over parameter updates. Could solve key AI model issues and avoid catastrophic forgetting. Worth keeping an eye on!

  • @dgpreston5593
    @dgpreston5593 4 дні тому +3

    Firstly, thank you for your attention to the paper and for engaging with this topic in such depth. However, I find myself uncertain about the assertion that "No External Knowledge Base is Required." I will need to review the referenced papers to provide a fully informed response, but on the surface, this claim seems to stretch credulity. It appears to imply that the model weights have somehow been transformed-by what could only be described as an alchemical process-into a kind of crystal ball capable of embodying all future facts beyond the training cutoff date.
    Such a proposition seems unlikely. Instead, I suspect the mechanism involves projecting the user's original query into a conceptual space before retrieval. This process might serve to minimize overly broad activation of similarly shaped, yet contextually irrelevant, facts. Without such filtration, these non-germane facts could inadvertently represent privacy leakage if injected directly and unfiltered into the primary LLM.
    That said, your interpretation may well be correct, and I acknowledge that a deeper review of the papers is necessary to evaluate this fully.

    • @KManAbout
      @KManAbout 3 дні тому +2

      This was something that confused me. I think you are right.

    • @code4AI
      @code4AI  3 дні тому +3

      A statistical method is a crystal ball for you. What an interesting view.

    • @dgpreston5593
      @dgpreston5593 3 дні тому +1

      @@code4AI as a metaphor, "crystal ball" in the sense that the InCA process purportedly would allow one to make true statements, at points in time "into the future" wrt the training cutoff date.
      My quibble might be with the notion that "no external knowledge base is required". I think this could be reasonably qualified. Otherwise, we are adventuring into the realm of prophesy. Not that such is necessarily a bad thing, if a less-bold term would be intuition, prediction, anticipation, or such. It depends on your use case, and whether your training set and inference domain are intentionally grounded in objective historical records, or if speculation is admissible.

    • @dgpreston5593
      @dgpreston5593 3 дні тому +1

      BTW, thanks for drawing the parallel to Meta's LCM. These ideas, while distinct, use greater degrees of abstraction from two different points of origin to obtain similar gains in results. I think both are worthy of exploration, but it may become too difficult to resist over extending the frontier of opportunities.
      😮😢😂😅😊

    • @HunteronX
      @HunteronX 3 дні тому +1

      Exactly.

  • @bernardoramos9409
    @bernardoramos9409 День тому

    This is not "continual learning", it is just another way to classify and store/retrieve information (modern database)

  • @supermold
    @supermold 4 дні тому

    I feel like the granite llms might be well suited to being used as an ECL.

  • @BuFu1O1
    @BuFu1O1 3 дні тому

    🤯🤯🤯🔥

  • @nadavnesher8641
    @nadavnesher8641 4 дні тому

    Great!

  • @yuriy6249
    @yuriy6249 2 дні тому

    Hey, please stop pulling RAG into each and every of your videos here and there. Looks like you don't understand the whole point of RAG, what it does and what problem it aims to solve, nor how it works. Your example with "check deposited" clearly shows that you don't have a clue what kind of data retrieval augmenting is for.
    RAG is solving exactly the scalability problem, which is actually pointed out clearly for the ICL on the first page of the paper you're citing in video! Prepare better.

    • @code4AI
      @code4AI  2 дні тому

      Thank you for this comment. I will use it in my new video. Great stuff.