show how gradient is penalized

[jeisner]

Recall that the log-likelihood bar shows the unregularized log-likelihood and how part of it has been removed for regularization.

The gradient bars could do the same: show the unregularized gradient and how part of it has been removed for regularization. In fact, the gradient may change sign as a result, so we'd have to think about how to display this. The tooltips might also change.

This shows more clearly how regularization pulls the weights back toward 0 ("weight decay"), and it more clearly links the gradient to the log-likelihood bar.

[fmof]

There are actually places in the code (stubs, not implemented) to do this. The biggest challenge is the shading. (Granted, that is what is required by a solution to this...). I can expand on this in a few hours, but it seems to do this would require an entire overhaul of the gradient shading bit. Rather than using browser-supported shading as I'm doing now, I think the easiest would be to position a whole bunch of colord divs in a very careful manner.

[jeisner]

Confused by the reply because I wasn't suggesting shading and I don't think there is any shading now.
Only something like what we have on the log-likelihood bar.

[fmof]

The log-likelihood bar is an SVG element, which means I can just draw rectangles on it as I choose. (If you look carefully, you'll see that the regularization penalty is actually an overlay on the unregularized log-likelihood.)

The slider bar (what's colored) is a div, and the gradient components are colored in with a gradient shading function (at 2AM I conflated the two). Doing an overlay with the shading function would be pretty tricky. To really do this, either I'd need to replace the shading with overlay-divs, or replace the slider bar div with an SVG element. The first option requires playing with the CSS widths, positions and z-indices. The second option requires hacking an external jQuery library (though using SVG elements could bring tighter integration with d3).

[jeisner]

Hmm, SVG sounds like it would be the ideal solution but would be a pain --
do I understand correctly?
Alternatively, could you just shade part of the gradient bar in a different
color?

I haven't even figured out what the look should be, since sometimes the
regularization lengthens the gradient bar instead of shortening it.
Possibly we just want some kind of vertical line showing where the bar
would have ended up without regularization.

On Sun, Jun 9, 2013 at 12:54 PM, fmof notifications@github.com wrote:

The log-likelihood bar is an SVG element, which means I can just draw
rectangles on it as I choose. (If you look carefully, you'll see that the
regularization penalty is actually an overlay on the unregularized
log-likelihood.)

The slider bar (what's colored) is a div, and the gradient components are
colored in with a gradient shading function (at 2AM I conflated the two).
Doing an overlay with the shading function would be pretty tricky. To
really do this, either I'd need to replace the shading with overlay-divs,
or replace the slider bar div with an SVG element. The first option
requires playing with the CSS widths, positions and z-indices. The second
option requires hacking an external jQuery library (though using SVG
elements could bring tighter integration with d3).

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/96#issuecomment-19169395
.

[jeisner]

I haven't even figured out what the look should be [...]

Another issue about how to display gradient bars:

With L1 turned on, the gradient hint bar often stops at 0 because that is where a jump would take us. I tend to think that the hint bar should show the true gradient that corresponds to O - E - C_R, not the truncated version. Even if we choose to truncate the *actual update_ in that dimension so that it stops at 0.

[fmof]

Hmm, SVG sounds like it would be the ideal solution but would be a pain -- do I understand correctly?

To both parts, I'd say "most likely." I know that fixing #17 was a bit of a pain, involving intercepting and redirecting function calls. I don't think replacing the slider div with SVG should have ripple effects but I'd want to think about it some more before proceeding.