This is such a rich and helpful illustration! Thanks for building it.
I would love to see a user interface capability where you could just load some parameters that are known to produce a really great outcome in the classification for each dataset. To motivate this, I spent a lot of time experimenting with the "two spirals" example. It took me a while to converge on a set of parameters that did something reasonable. Every step of the way was really valuable in learning, especially seeing the sensitivity of the classification to different activation functions, number of {layers, neurons}, inputs, etc -- but it would really cool to just be able to push a button at the end of it and maybe see a really good set of parameters and maybe even some commentary on why they work.
some really bad inputs and why they are bad could be cool too :)
Again, amazing work. thanks for building and sharing!
This is such a rich and helpful illustration! Thanks for building it.
I would love to see a user interface capability where you could just load some parameters that are known to produce a really great outcome in the classification for each dataset. To motivate this, I spent a lot of time experimenting with the "two spirals" example. It took me a while to converge on a set of parameters that did something reasonable. Every step of the way was really valuable in learning, especially seeing the sensitivity of the classification to different activation functions, number of {layers, neurons}, inputs, etc -- but it would really cool to just be able to push a button at the end of it and maybe see a really good set of parameters and maybe even some commentary on why they work.
some really bad inputs and why they are bad could be cool too :)
Again, amazing work. thanks for building and sharing!