Join a cohort of fellow learners to discover the principles that underly ChatGPT and other LLMs. I have adapted the material from the aforementioned book into a series of four notebooks covering topics ranging from probabilistic text generation to neural nets, machine learning, embeddings and even transformer models.
Giulio Alessandrini, our manager of machine learning at Wolfram, will join to answer your questions on the day we cover neural networks. On the final day of the study group, Alan Joyce—who directs content development at Wolfram|Alpha—will join us to explain how the computational powers of Wolfram Language can be integrated with ChatGPT, and why this integration is much more than the sum of its parts.
Stephen Wolfram’s book is aimed at anybody who is curious about these ideas, and this study group follows the book’s lead. Therefore, no prior Wolfram Language, machine learning, or even coding experience is necessary to attend this study group.
Please feel free to post any questions, ideas and/or useful links in this thread between sessions—we always love to continue the discussion here on Community!
This is a one-week study group that will run from September 9 through September 13, 2024 at 11:00am Central US Time each day.
I was curious to see that the sample from the 4-gram model in today’s session/notebook was the smallest to generate actual English words. Is that likely related to the fact that 4.7 letters is the average word length in English? (Source)
My mistake, the 2-gram generated ‘wore’ and ‘hi’ and the 3-gram generated ‘was’.
Though I imagine the increase in the proportion of valid words in the sample is related to the distribution of word lengths in English?
Hey Arben,
I have already learned more than I expected to from Wolfram’s perspective on AI, Machine Learning and ChatGPT. Looking forward to your approach to teaching.
But could I really use Wolfram Language alone, or WL combined with (?), to do industrial grade things like spot defects in images of circuit boards or components soldered down on circuit boards? Like cracks, malformed parts, missing traces, debris, etc? Or would that still require developing a specialized tool? Or look at images of objects from different angles and determine if they are images of the same thing? Like an aerial photo of a city or a farm?
Or better yet, after recognizing whether they are images of the same thing, highlight what has changed in the thing after correcting for the different look angles and lighting? Or would that be too specialized an application?
(I am certain that if I wanted to do that I would at least look at WL as my experimental laboratory for learning how to do it)
That may well be true. For the words → sentences analogy, sentences can have extremely variable lengths, so it might not just be the mean word length but the tightness of that distribution!
Hey Carl! Thanks for the kind words for our materials :). While I’m not a machine learning pro, I think that your example use case is quite doable (and depending on the image resolution, you may need fewer images than you imagine—high resolution images by definition have a lot of information stored in them, and IME even training on the order of tens of examples can be sufficient to do binary categorization with high accuracy in those cases).
As for perspective transformations, that could be a little trickier. I’m not sure what the extent of the different angles and viewpoints might be, but it could be that your training data could include an image and various transformations of it created with something like ImagePerspectiveTransformation or similar?
For spotting differences, theoretically you could train a neural network to do that by providing training data consisting of two input images with the output being a selected region where differences occur, but I’m not sure that that’s the direction I’d take. If you can do something like FindGeometricTrransform to get your two “perspectives” roughly aligned, you might be able to do something like explicitly compute image differences and find the region over which they’re maximized. This is kind of on the cusp of “better with traditional methods” vs “better with machine learning”, to my (admittedly somewhat ignorant!) mind.
Hi everybody—I’ve just uploaded to the materials folder a slightly updated version of today’s notebook with a few fixes we encountered during today’s session.
AI has so many ways to describe it: AI, Machine Learning, Deep Learning, and Generative AI. Which term is the best way to describe Wolfram’s approach? Why is Wolfram’s purpose to get into AI world?
“Machine learning” is generally used to describe the process of training neural nets
“Deep learning” is a subset of machine learning referring to the training of “deep” neural networks, i.e. those with many layers
“AI” is super general and encompasses the above and more—it’s so general that it can lead to miscommunications and mystification, though
“Generative AI” refers to “AI” (usually based on neural nets which have been trained via machine learning) which can “generate” novel text, images, audio, and so on
As for what we’re doing at Wolfram, we’ve supported neural nets and machine learning for a long time now, and our fully symbolic approach is really well-suited to the construction and understanding of neural nets. Our new LLM (one example of generative AI) functionality takes advantage of the benefits of this new technology to do things like create helpful personas, write code, check or rephrase writing, and so on inside notebooks. Alan will address this in more detail on Friday!
“Machine learning” is generally used to describe the process of training neural nets
“Deep learning” is a subset of machine learning referring to the training of “deep” neural networks, i.e. those with many layers
“AI” is super general and encompasses the above and more—it’s so general that it can lead to miscommunications and mystification, though
“Generative AI” refers to “AI” (usually based on neural nets which have been trained via machine learning) which can “generate” novel text, images, audio, and so on
As for what we’re doing at Wolfram, we’ve supported neural nets and machine learning for a long time now, and our fully symbolic approach is really well-suited to the construction and understanding of neural nets. Our new LLM (one example of generative AI) functionality takes advantage of the benefits of this new technology to do things like create helpful personas, write code, check or rephrase writing, and so on inside notebooks. Alan will address this in more detail on Friday!
Hey Arben,
I had an open session of 14.0 running on my Mac and I double clicked the Machine Learning notebook after downloading it; my Mathematica session went off to LaLa land. Wouldn’t even obey the interrupt Cmd + Option+ [.] or CTRL C. Had to force quit from MacOS. I started a new session and opened the notebook from within Mathematica and it opened as I expected. But I have not “enabled dynamics”… It should have opened the same by double clicking on the file, I think, yes?
Hey Carl—there shouldn’t be any issues with either way of doing things, and they should indeed be equivalent. (Now, Day 3’s notebook is so big that it might cause an issue here or there, but even that one ought to be fine.)
Hi! I have a question regarding the implementation of AI. In my city, Guatemala (in Central America) there is a project to use IA to control the traffic lights. I haven´t see examples of this kind of implementation in any other places. And since in todays session was mention the problems that a neurological networks can have with algorithms, do you think this implementation could be effective? I belive is not a generative IA what is going to be used, but I wonder what would be the results.
Hi Ana! I could see that being a possibility. I think the most likely thing is that traffic cameras could recognize how many cars there are along each direction and what sorts of speeds they’re seeing, then change the lights dependent on those inputs. There are traditional ways to do this measurement, but with neural nets you can just plug in any camera and a cheap circuit board and very quickly get reasonable estimates for these parameters (as well as, say, bikes and pedestrians). You could then use this data to determine when it’s time to change the lights, in something like the following fashion:
Camera uses neural nets to detect the number of cars, people, and cyclists, as well as their speeds
This data is fed into a neural net which has been trained on data that looks like “if you have this many cars and people and cyclists and these are their speeds, switch the light”
This decision is fed into some kind of traditional failsafe, as you couldn’t guarantee with a neural net alone that you wouldn’t wind up having an intersection have green lights in two perpendicular directions, which would be very bad!
So, you’re right that you would need some kind of non-neural network “final check” for a case like this, and also right that this isn’t something you’d need (or want) to use the newer generative-style AI for.
Thank you Arben. In this scenario do you think the neural net could start to memorize the traffic information? If that happen I starting to see the problems that would trigger.
