DataTaunew | comments | leaders | submitlogin
Tensorflow Playground (tensorflow.org)
32 points by tfturing 2906 days ago | 2 comments


1 point by Homunculiheaded 2900 days ago | link

My favorite thing about this playground is that it really lets you see the trade-off between feature engineering and letting a deep network learn the features of the data.

For example, you can take all 7 features (which includes the 5 engineered features) and just 1 hidden layer of 7 Neurons [0] and achieve test loss of ~0.008 in 1000 iterations on Spiral.

You can also take just the original 2 features and 4 hidden layers of 8 neurons [1] and get a similar training loss in 1000 iterations.

While both networks perform roughly the same in the same number of iterations the network using feature engineering learns much faster, and in my observations much more consistently.

Worth noting is that all features + 4 layers of 8 neurons [2] achieves the same test loss in less than 400 iterations.

For traditional ML in practice, feature engineering is far more important then the time it is given in texts/courses. This playground is a great demonstration that even though you can learn features with deep networks it is still of practical benefit to put time into engineering features.

[0] http://playground.tensorflow.org/#activation=tanh&batchS...

[1] http://playground.tensorflow.org/#activation=tanh&batchS...

[2] http://playground.tensorflow.org/#activation=tanh&batchS...

-----

1 point by patwater 2906 days ago | link

Oddly hypnotic

-----




RSS | Announcements