Generate Keras models from a description.https://twitter.com/mattshumer_/status/1287125015528341506 …
-
0:32 -
Replying to @gdb
Good attempt. However this model would blow up immediately as it has neither strided convolutions, nor strided pooling. Also not deep enough for 500x500. With proper striding, it would work quite well on fashion MNIST.
2 replies 0 retweets 34 likes -
Replying to @ChrSzegedy @gdb
I have no doubt that building common models from a description can be automated, but I can't help but think it would be more effective to have a structured intermediate representation, and:
1 reply 0 retweets 19 likes -
1. Use a supervised text model to go from description to structured representation, trained on a specialized dataset 2. Use a hard coded algorithm to generate Keras code from the structured representation (ensures code reliability)
2 replies 0 retweets 11 likes -
Why go through another layer of structured representation?
1 reply 0 retweets 2 likes -
Replying to @ChrSzegedy @gdb
Because you want to achieve generalization via interpolation, and forcing the interpolative model to deal with a huge layer of discrete complexity (code syntax) is a big hurdle for that. The simpler the structure of your target space the better. It's curve fitting after all
1 reply 0 retweets 4 likes -
It seems GPT copes just fine :) Adding layers of transformers seems like the right direction to me. Adding layers of complexity is not.
1 reply 0 retweets 3 likes
It's actually about simplifying the problem -- it's literally feature engineering. But you're right that as long as you have infinite training data and infinite compute resources, it's less work not to do any feature engineering at all
-
-
Computing resources are an issue, but training data is easier to come by for a represenration that is commonly used.
0 replies 0 retweets 2 likesThanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.