We can try to implement an interactive service that aids users in creation of neural networks. It can suggest architectures, improvements in architectures given a task (an example suggestion:
User: I'm trying to create an autoencoder on the MNIST dataset, but it is giving low performance.
Service: [Analyzes the project info available, such as architecture, dataset used, training testing method, etc.] Sure! Your autoencoder seems to have an encoding vector of dimensions 2000,1. However, images in MNIST are of dimensions 28,28. That gives us 784 pixels. This means the entire image is being placed in the embedding vector, along with some extra dimensions left. Try reducing the dimension of the vector to a smaller number, say 200,1.
)
It should also be able to cite any sources used, and apply suggested changes to the user's graph (rather than just suggesting them).
We can try to implement an interactive service that aids users in creation of neural networks. It can suggest architectures, improvements in architectures given a task (an example suggestion:
)
It should also be able to cite any sources used, and apply suggested changes to the user's graph (rather than just suggesting them).