A photo of the royal institution lecture hall originally uploaded by uploaded by AnaConvTrans to Wikipedia

Royal Institute Christmas Lecture on AI 2023

I’m excited and very proud to see my colleague and mentee Safa appear on this year’s Royal Institute Christmas lecture. In the video, Safa is assisting Prof. Mike Woolridge (who has some excellent and sensible takes on AI by the way) with some experiments.

In general this is a well put together lecture that gives a lay summary of some of the techniques and acronyms that you’re likely to come across in conversation about AI. I’d recommend it to anyone interested in learning more about what these systems do or to any relatives who might have brought up the subject of AI at the Christmas dinner table

  1. glyn avatar

    @jamesravey I watched the recording earlier. Safa did great.

  2. Jamie Jones avatar
    Jamie Jones

    A completely disconnected list of relevant topics is absolutely NOT an explanation:
    (1) Neural networks
    (2) Training
    (3) Large Language Models
    (4) Language translation (e.g. English to German)

    Then there’s some of the problems with AI, not all of them technical, which were not mentioned:
    (5) Neural networks generally cannot explain their conclusions
    (6) Defect #5 gets worse when the neural network is learning AFTER training
    (7) Large Language Models may contain an unknown number of falsehoods
    (8) Only hugely wealthy organisations can own AI tools…price of entry problem

    I’m only an amateur observer….maybe someone can tell me what’s wrong with this critique.

    1. jamesravey avatar

      Thanks for your comment Jamie! I completely agree in principle but I also think that as a primer on “where to get started” with AI in 2023, the content was reasonable. If it ventured too far into some of these topics it would have to have been a series of lectures. I think many of the issues that you raise are complex and multifaceted so you couldn’t do justice to them without running a full additional lecture.

      Addressing some of the problems you mentioned:

      Neural networks generally cannot explain their conclusions

      and

      [this] gets worse when the neural network is learning AFTER training.

      Neural networks don’t learn after training but, I’m assuming you’re referring to fine-tuning which is when you take a model that was already trained and modify/update it. It’s true that neural networks are typically considered “black boxes” but there are methods that can be used to introduce explainability to neural networks and these can be applied post-training or post-fine tuning to an end to end system. There is lots of great research going on into applying explainability tools to larger models. Perhaps the lecture could have at least nodded to the existence of these techniques and tools but I think it could quickly become confusing and off topic – better to broach once the fundamentals have sunk in.

      It would have been nice to have seen some material discussing other types of models than neural networks. In my professional practice, I always recommend starting with the simplest model that will solve the problem before progressing to complex systems. Why should we use an LLM with all the fiscal and environmental cost that entails when a logistic regression classifier could solve the problem with 95% accuracy?

      Large Language Models may contain an unknown number of falsehoods

      I think the dangers of over reliance on models and our tendency to anthropomorphise things that behave like humans are definitely worth discussing. However, I think there are some more fundamental philosophical things that need addressing too.
      “All models are wrong, but some are useful” is a core tenet of data science. Model performance largely comes back to the quality of the training data or “garbage in, garbage out”. I think as a general rule of thumb we should be educating people about the fact that no ML model is going to be a “silver bullet” for solving every task. If the model is trained on garbage, it will produce garbage. If the model was not trained to solve a problem, chances are it won’t be able to solve it.

      LLMs got people excited because they appear to generalise to new unseen problems in ways that historically ‘traditional’ models couldn’t. However, recent work suggests that they can’t and that they are actually likely memorising and regurgitating training and pre-training data. The generative process that LLMs use to create new outputs is essentially based on random chance but using weighted dice. They produce content that is “likely” to be correct but not guaranteed (in fact they have no concept of ‘correctness’).

      Only hugely wealthy organisations can own AI tools…price of entry problem

      This one I respectfully disagree with. AI encompasses more than just LLMs and its cheaper and easier than ever to build out useful models. Even if you do just mean LLMs, there is a thriving community of developers working on open source models like Llama and Mistral which can be run by anyone and fine-tuned for next to nothing. The barrier to entry is technical know-how and access to the internet (where you can use resources like Google Colab to fine-tune models for free). The expensive part is typically building and cleaning the dataset. That said, most organisations that have been around for a while and have an established process or system that they want to augment with AI will have some legacy data lying around that they can use to get them started.

      I think overall the content was a bit LLM heavy and as you say, there were some omissions that ideally would have been covered, but it was a really nice intro to the key concepts involved in AI/ML and a great starting point for young persons looking to study the subject!

Leave a Reply

Your email address will not be published. Required fields are marked *