put logos back

[Bob Carpenter]

I put the logos back and floated them to
the right of the text:

http://code.google.com/p/stan/

We're missing a non-small GM compiler logo and it
looks fuzzy compared to the others on the new monitor.

Daniel -- can you generate one about the same resolution as
the others?

Other quibbles:

* fonts aren't same size when displayed at width=150.

* to much interline space on the bold titles (other spacing
is OK).

* Titles aligned just a little below top of circles on all

We can do it together if you're around tomorrow.

- Bob

[Andrew Gelman]

And we will have "Gibbs local" for Stan 2.0, right?

[Bob Carpenter]

On 7/12/12 5:33 PM, Andrew Gelman wrote:
> And we will have "Gibbs local" for Stan 2.0, right?

At the very least, we'll have an adaptive slice
sampler.

Should get rid of the Gibbs logo for now?

We'll have to figure out a model syntax for Gibbs
from which we can infer conjugacy. This kind of
thing is much easier if you restrict the models to
directed graphs as in BUGS and JAGS.

- Bob

[Andrew Gelman]

I defer to you on logos, but we can discuss slice sampling once stan 1.0 exists. I'm assuming that in our first version we will just do brute force on the entire log-posterior when doing slice on discrete variables. But I'd think it wouldn't be hard to allow for some handholding by the user to indicate which portions of the model are relevant for each discrete variable.

[Bob Carpenter]

On 7/12/12 6:02 PM, Andrew Gelman wrote:
> I defer to you on logos, but we can discuss slice sampling once stan 1.0 exists. I'm assuming that in our first version we will just do brute force on the entire log-posterior when doing slice on discrete variables. But I'd think it wouldn't be hard to allow for some handholding by the user to indicate which portions of the model are relevant for each discrete variable.

That would be pretty easy to implement now.

But for models of any size it'd be very slow compared
to BUGs/JAGS.

- Bob

[Andrew Gelman]

But the idea is we'd only do Gibbs/slice for the discrete variables. We'd do HMC for the others.

[Bob Carpenter]

On 7/13/12 10:01 AM, Andrew Gelman wrote:
> But the idea is we'd only do Gibbs/slice for the discrete variables. We'd do HMC for the others.

That's simplest.

The bigger step would be to allow samplers to
be declared for blocks of variables, so we could have
two blocks of HMC, a block of HMC and Gibbs, etc.

- Bob

[Andrew Gelman]

Yes, that would be fine. My point was that if we have a big model with just a couple of discrete variables and we do Gibbs/slice with them, that this would not necessarily be slower than doing Bugs on the whole model.

The real concern, though, would be with variable such as mixture indicators that come in large batches.