While I’m at this: sorry the picture on my whiteboard is wrong, but it’s better for time to be the horizontal axis and accuracy to be the vertical axis, because it’s more intuitive for higher to be better.

Two more comments about 4.2: First, please work out the example above Section 4.1 (probably give the sequence of rewrites) and use it to explain the Add and Index rules. Second, I don’t understanding the word “outermosting” and that sentence seems no verb. Thanks!

(Another part of my first comment above is that enumerate item 2 (the Index rule) should have o and n renamed to i and m)

Why do the two plots have different slopes?

Still to do: Work out example for 4.1 to illustrate things. Hopefully there will be space for that.

Is that change pushed?

The figure? Yes.

And now the worked example as well.

I have to attend to various things now (drive, meeting #1, teach grad class, meeting #2) which means I will be online quite sporadically between now and 4:00. Then I’ll have a window until 5:00, which will go back to sporadic until 8:30, then I’ll be back online solidly.

I do believe that section 4 (all of it) is in much better shape now.

Reading over things, a few thoughts

1. I think the early part of 3 should use the word “monad”

1. we should make some statement about how much of the calculus and probability theory you need to understand to get something out of the paper

and what to do if you don’t know that stuff

1. we should say why “we convert (1) into <nested integrals>” makes sense

@rjnw what’s the benchmarking status?

@ccshan @pravnar are you around LH?

Sorry I just woke up, but I’m heading that way

sounds good, I’ll see you then

I’m reading over 2

Ok, and if you have a perspective on how much math background is needed, let me know

@samth I also woke up late. I will be in LH within the hour.

sounds good

I want to say somewhere that Section 3 merely requires CALC I and the ability to apply that knowledge recursively.

I think the comment 3 above I made requires more than that

I pushed a couple of sentences there to explain the intuition behind the integrals

I’ve pushed some edits to Sec 2

as well as an additional box in the figure

@rjnw and I are in my office

Current benchmark status: we have numbers for Linear Regression, Clinical Trials, GmmGibbs

LDA and NaieveBayes have some problems

working on those at the moment

Yall want coffee?

(sorry, on my way from home)

i need to get lunch

so I will maybe see you there

I’m in my office, working on making NaiveBayes and LDA produce numbers

@pravnar sounds good, because then we’ll have at least some numbers for those

Because I did goof’ed and appended to a file instead of writing over it. So this plot is based on corrupted numbers…

I was looking over AugurV2 — are we comparing against it? Or just citing? Their “GMM” (Figure 1) is really short. Is it the same as ours?

Yes, except they take pis as an argument whereas we choose theta randomly.

We should compare against them but haven’t gotten around to it.

I was just impressed by the brevity.

You’ll see where the brevity comes from if you do a line-by-line/word-by-word comparison.

For example their lambda is shorter and omits types. They don’t need the “if” we use to encode an assertion. And they don’t draw theta from a Dirichlet, so they don’t need the definition of Dirichlet.

I’m going to drive home now, as it seems like as good a time as any. Then I’ll read/edit through all the pieces, especially the ones that I haven’t read/edited yet.

Right - so it would be interesting to use the same model (in a comparison), so as to compare apples with apples. Our Plate notation is shorter than their notation for the same, for example.

Great, thanks. It’s time to proofread for math mistakes, overfull boxes, unfulfilled promises, undefined terms, grammar and spelling.

(By the way, you might have noticed that I’m increasingly a fan of moving the point to the front (of the sentence/paragraph/section/paper).)

I’ll fix all the things that I can - margin notes for the rest?

Sure - something that doesn’t affect the formatting.

(If you put margin notes that take a long time to address, of course they won’t be addressed.)

Yes, I have seen that trend. I am not against it. Unless it turns a sentence into the passive voice, then I am.

The main use of passive voice is to move the point to the front.

Then I am not so happy. I much, much prefer the active voice, it really makes things much more readable.

For example, you wrote “Then further case expressions are translated to either…” and that’s great because the point is there are further case expressions.

I am not rabid about it :wink:

In that paragraph, there was a natural flow, and it made more sense to go with it.

Point taken.

Anyways, actually going now.

#hakaru Are we keeping the radar chart?

Also, when are we expecting the numbers and plots to be added?

I’m editing Section 1. We’re working on numbers.

Make sure to pull - I’ve tweaked section 1. I’m working on 2.

Your changes to \dbltextfloatsep and \textfloatsep don’t seem to save any pages.

No pages, but about 15 lines.

And will matter increasingly as we add more figures.

I’ve reread 3 and 4. I’ve read these so many times already, I can’t see things to fix right now. I’ve read 5 and related work - pushed super-minor tweaks.

Unfortunately, I am feeling quite unwell. I am unlikely to be able to stay up very late.

I don’t even see any lines saved. FWIW, Sam and I today discovered that your TeXLive version and TeX version/variant might affect the output. Because for example the class file backs off to using Computer Modern if it doesn’t have the font it wants. And the default integral sign changed in newtxmath versions.

Also I don’t see any changes to the related work section, btw

I think the 2 minor changes were in section 5, nothing in related work.

I am not surprised that versions matter.

Right now, it is probably best if I stop editing — the way I’m feeling, I’m way more likely to make stupid errors than to improve things.

Was waiting for plots to show up to help finalize, but seems like I won’t be able to. Sorry.

@samth Here’s a Julia LDA model: https://github.com/hakaru-dev/tcp/blob/master/lda.jl

https://github.com/wpm/Naive-Bayes-Gibbs-Sampler