Show HN: NeuralFlow – Visualize the intermediate output of Mistral 7B

by valine • 2 years ago

134 20 2 years ago

7734128• 2 years ago

That's pretty, but it's not "useful" in any way right? Or can people use these images to better understand the inner working somehow?

valine7• 2 years ago

An individual snapshot of the model’s output isn’t very useful. What I’ve found enormously helpful is watching how the structure of the output changes over time as I fine-tune Mistral.

There will typically be visual artifacts in the heat-map that appear right around the time the model starts to go off the rails. Due to the nature of residual layers, problems at lower layers cascade and light up when visualized this way.

Difwif• 2 years ago

Fascinating. Makes me think there's still room for new architecture/training ideas like dropout or weight decay to improve model performance.

aantix• 2 years ago

How do you determine “ right around the time the model starts to go off the rails”?

iandanforth• 2 years ago

If you do this for gradient updates you should get a leading indicator.

albertzeyer• 2 years ago

When you frequently look at some visualization, you will get a feeling when things are normal, or also when things are unexpected.

If it is the first time you look at this, or first time for this specific visualization, or first time for a new type of model, it will not tell you much, and you cannot really tell when things are wrong. You really need to compare it.

E.g. in the readme, it shows the example of overfitting, where you see that activations became too large. But to tell whether this is unexpectedly large or not, you need to compare it to some other plots where everything is fine.

lappa• 2 years ago

A few days ago I saw a post using NeuralFlow to help explain the repetition problem.

https://old.reddit.com/r/LocalLLaMA/comments/1ap8mxh/what_ca...

I’ve done some investigation into this. In a well trained model, if you plot the intermediate output for the last token in the sequence, you see the values update gradually layer to layer. In a model that produces repeating sequences I almost always see a sudden discontinuity at some specific layer. The residual connections are basically flooding the next layer with a distribution of values outside anything else in the dataset.
The discontinuity is pretty classic overfitting. You’ve both trained a specific token to attend primarily to itself and also incentivized that token to be sampled more often. The result is that if that token is ever included at the end of the context the model is incentivized to repeat it again.

...

Literally just plotting the output of the layer normalized between zero and one. For one token in mistral 7B it’s a 4096 dimension tensor. Because of the residual connections if you plot that graph for every layer you get a really nice visualization.
Edit: Here's my visualization. It’s a simple idea but I've never personally seen it done before. AFAIK this is a somewhat novel way to look at transformer layer output.
Initial output: https://imgur.com/sMwEFEw
Over-fit output: https://imgur.com/a0obyUj
Second edit: Code to generate the visualization: https://github.com/valine/NeuralFlow

This is nearly identical to the overfitting example in the repo, only really representing a binary, but it's a good start. Perhaps some transformations can be applied to help further?

vermarish• 2 years ago

Any explanation on the choice of the colormap? It can have a powerful effect on how we perceive this visualization.

valine7• 2 years ago

Trial and error. I tried a bunch of different color maps, this one had the best contrast.

My process is to periodically prompt the model as I fine-tune. The features that seem to correlate with the model losing coherence are highlighted nicely with this color mapping. More of an art than a science.

mft_• 2 years ago

Now we just need to train a model on this visualisation to help us to interpret it...

nmitchko• 2 years ago

This reminds me of the matrix movie scene when they look at the encrypted thoughts of the matrix.

https://cdn.swisscows.com/image?url=https%3A%2F%2Fi.pinimg.c...

greasegum• 2 years ago

You get used to it, though. Your brain does the translating. I don't even see the code. All I see is blonde, brunette, redhead.

Show HN: NeuralFlow – Visualize the intermediate output of Mistral 7B

Related Stories