Use this thread for questions/discussion of today’s lesson. Please do not use this for questions/comments about topics we haven’t covered yet - use the Lesson 6 further discussion thread for that. Also remember to watch the official updates thread.
I hope FB releases windows (anaconda) version of pytorch 1.0 during NeurIPS
Beautiful landing page for platform.ai @jeremy!
what do the two axes in there represent?
Nothing really I think, just a 2D projection from a much larger space. A bit like PCA I guess. Well rather than nothing it’s more that “we don’t know” and we have to figure it out, that’s the whole point of it
Have you tried peterjc123 builds. Working great for me.
I think I must have missed this - but what tool is Jeremy using to show the PCA-like projection?
This is platform.ai
Maybe I missed first couple of minutes. Dumb question: What is projection ?
where do these categories come from? are they pre-defined? or is there a set I can choose from (like if I want to find a projection that separates sedans from SUVs, is it possible?). Thanks.
A projection is a way of going from a higher dimensional space to a lower dimensional space. Although the new, low dimensions don’t have inherent meanings, often you can observe a meaning or pattern.
Often you many want to look at several possible projections to find something that seems “meaningful” or that separates your data in a way that interests you.
It’s there to help you label a dataset, so yes, you can have the labels you want.
I was part of the fellowship this past summer some of my fellow fellows built this! Pretty cool to see it live!
Does platform.ai employ active learning ?
The projections that Jeremy was demonstrating a couple of min back – are they some intermediate layer taken from a trained network or similar?
Or is this a standalone tool which just takes all images, similar to doing PCA on a high-dimensional data. ?
It involves taking intermediate layers from a network.
Yes it is.
Does adding many more features always positive for a neural net, or can it lead to a kind of “information overflow” and actually be detrimental to the model ?
Making life easier for your model will generally result in getting better results.
That being said, to quote last-week-Jeremy, try bla