So, I stopped my previous post after my thoughts on the DEI talk. The actual post wasn't all that long, the reason I said for stopping the post but rather because writing that post was emotionally draining. Can I talk about that stuff? Will people interpret it the way I intended. Will people who I don't want to offend and don't think I'm offending, will they be offended? Will the people I want to offend not be offended (okay, just joking here). It's really hard for people to have these conversations and too many people these days don't allow for nuance but I felt it important to get my thoughts out there. Not so much because my thoughts are anything special or that they're right or wrong but one, because the session was terrific and I am documenting my conference experience and two because having these conversations won't get easier unless people are willing to put themselves out there.

Anyway, back to our regular programming.

Thursday afternoon I sat in on "AI Lessons using Machine Learning for Kids." Machine Learning for Kids is a site/platform that allows people to set up and train machine learning models. The speaker's slide deck with links to resources can be found here.

The speaker walked us through how she used ML for kids with her class. She basically teaches a typical Machine Learning pipeline, that is:

collect data –> prepare/clean –> select model –> train –> validate –> predict

Her specific project was sentiment analysis of text. The platform can also work with images, sounds, and numbers. After logging in (or clicking to try without registering) you start a project. You select the model you're using, "recognizing text" in this case. You then use their interface to add data. The instructor, of course, would have to provide guidance so that a student would know what and how to enter.

When finished, you click the train button and the platform builds the model. I'm pretty sure you can then play with the model but I didn't actually test that. The cool thing, to me was that you can then have the platform generate code - scratch, python, and a couple of other platforms. I clicked the generate button and all of a sudden I got a chunk of Python code that I could embed in a program. Specifically, the code let you use your now trained model as a library (or more specifically, web based API) and make calls with data and get your results.

While this is very cool, it raised some questions. The big one for me is that you can go through all of this and have almost zero knowledge of the underlying machine learning. Contrast that with the session I wrote up a couple of days ago where the students learned all the little details of gradient descent for linear regression (more on this later).

Is this really the right way to go?

The instructor noted that when we use a tool like ChatGPT we're using an already trained model so this is no different. That's true but in a class learning about, say sentiment analysis as an ML technique, I think students should learn how the sausage is made.

Now, this approach of leaving the actual ML as a black box does still allow for some good learning to go on. Students will learn the pipeline. They can also discuss issues like false positives and true negatives - the effects of different training data, over training and much more. They could also see, from a user point of view if different models produce different results. A teacher might also have constraints to keep them from diving deeper - insufficient time, toolset limits, student backgrounds and I'm sure others.

That said, I'm a firm believer in seeing what's under the hood.

Now, going back to the Linear Regression talk from Wednesday, that might be too much for an average high school class and almost certainly too much for a pre APCS-A class so where do you draw the line?

It turns out that in a CS0 class using a text based language like Python, students can pretty easily write a simple sentiment analysis system. You can also scaffold pretty thoroughly.

The basic gist is:

1. Make two lists, one of positive words, one of negative
2. Read in the text you want to test
3. Split the text into words
4. Count how many words are in the positive list and how many in the negative.
5. If there are more positive than negative, it's a positive piece, otherwise negative.

This is extremely bare bones and has issues. For instance if a document has "not bad" in it it'll just see the word bad and classify as negative but that's something the class can talk about. The project can also be easily extended in any number of ways. You can also include all of the good things from the speakers lesson design - the pipeline, different training sets, etc.. The important thing is that the students will have a stronger understanding of how sentiment analysis actually works.

While this was a strong session, I'm a big believer in students understanding the inner workings of things and I think this is going to be even more important with the current AI craze - this is something one of the end of day keynote panelists actually pointed out (more on that down below). Of course, you have to know your population, where they're at, and what your restrictions are.

In middle school with scratch, the black box approach is most likely the way to go, particularly since ML for kids hooks in to scratch. Teaching a High School class using Python? You should probably do a unit where your class codes the whole thing.

Just my opinion and approach.

One final note before moving on to the rest of the day. I've noticed that while there are plenty of sessions that, by description, talk about ChatGPT tools - what's considered the cutting edge of the AI fad, there have been a few sessions on Machine Learning and more traditional AI. This has been great to see as so much of what runs and what will run the world operates using these techniques and technologies and it's important not to drop them in education for the flavor of the month.

That was it for the Thursday sessions for me, I spent the rest of the session time looking at the posters and the hallway track.

The Thursday keynote was a panel moderated by education journalist Natasha Singer. The panelists were (as seated left to right):

• Christy Crawford of the New York City Board of Ed's CSforAll initiative
• Maggie Johnson, Global head and VP, Google.org
• Mehram Sahami, Chair of the CS department, Stanford University
• Kip Glazer, Principal, Mountain View High School

Before diving in to the panel, I've got to express more pride in my Hunter CS Teacher Cert program graduates. While delivering some opening comments, Natasha Singer gave a shout out to another Hunter CS Certified teacher, Marisa Shuman citing some of the work Marisa's done bringing AI to her students in New York. Here's Natasha's article about it: link

Now to the panel. Overall it was a great session but I did notice one thing missing. There were no actual teachers on the panel. Maggie Johnson's from Google and he background's in higher ed as is Mehram Sahami's. Christy Crawford doesn't show any classroom teaching experience on her linkedin. Only Kip Glazer listed any K12 teaching experience on linkedin and that ended back in 2016 - about 11 years total which is a nice chunk but in my mind generally not anywhere near the time needed for mastery.

It would have been nice to have at least one actual K12 teacher up there. Even better if it were an experienced CS teacher who also had strong content knowledge.

The first question was on the overall impact of AI in education. Mehram was first to respond and hit the nail on the head - AI can't transform education, only people (teachers can transform education. Bravo. He added that AI can enhance teachers ability to transform education but it's just a tool. As a side note, I really felt that Mehram was pretty much spot on throughout. I was very impressed with his responses.

Christy chimed in with a common teacher's fear - will AI replace them.

I think this was an interesting juxtaposition. You have the academic speaking the truth of how things should be and the person immersed in K12 noting the unfortunate reality that the politicians and private interests are always looking to save money and turn a buck by cutting teachers and directing money to private companies to "compensate."

The next question asked were coding should live in K12 with tools like copilot and other code generating LLMs.

Maggie noted that even if AI can generate code, it still needs to be understood, checked, tested, extended, and embedded. She said she didn't know how much coding would be necessary but certainly it's something we much teach in K12.

Mehran noted that teaching coding syntax is still important and that we can use tools like chatGPT like a code library or API to enhance our student's work. This echoed something I wrote about when I wrote up this year's SIGCSE experience.

In response to Natasha Singer asking if there should be limits on AI on AI Maggie spoke on how it's something that Google talks about top to bottom every day. When she followed up with "what about in the classroom" Christy chimed in yes but added an oh so important "but not on the backs of the teachers." I do lament that there wasn't an actual teacher on the panel but applaud Christy for consistently having teacher's backs throughout the panel. It's not a surprise, she's always had teacher's backs in her role at CSforAll. I wish others in the NY DOE would follow her lead more.

Natasha also asked about corporate influences on curriculum as this is something somewhat unique to CS. As a new field and a new field in K12, few teachers and schools knew what to teach so for better or worse, outsiders filled the void. That's the college board, code.org, organizations like PLTW, companies liked CODEHS. This doesn't happen in other older fields. A friend of mine later pointed out that it sort of does but the outside influencers are the book publishers. Anyway, Maggie posited that technology changes so fast that industry is needed to keep curriculum up to date. I think she misinterpreted the question though. I believe Natasha was specifically talking about content providers like those I listed above while Maggie was thinking about pure tech companies informing curriculum which, if they did and if they "stepped back" to go back to yesterday's post would be great.

I think Mehram also misinterpreted the question as he talked of the symbiotic relationship between industry and education. I think he was thinking more about the university industry relationship.

Once again, though, Christy had the teacher's backs when she on the one hand noted how important industry can be - they can inform, provide opportunities like mentoring, and support but there has to be a hard line. They should not be driving curriculum or education. Again, bravo.

At this point, we ran out of time. Great panel. As I said, I particularly liked Mehram's comments which included a number I didn't mention here as well as Christy's constant defense of teachers but everyone was terrific.

That's it for day 3.

In the next day or so I'll get to day 4, the exhibitors, and some closing thoughts.