Can an Autoencoder neural network be used to create a simple code completion plugin for a developer using a certain programming language ?
The idea is that the training data will be generated from keyboard input, which will be fed into the network while the user is coding.
For example: Trainindata = tokens generated while coding = ['for(', 'let', 'if(', 'i++)'] etc.
The input and output space will be an 8D vector (longest word this network can recreate is 8). The hidden layer will have smaller dimension: 3D for ex, just to force a recreation loss so the recreation is more generalized later on.
If the above can be done and work efficiently, then can this architecture be later improved to also learn based on the word's position in the file ? As this is a code-completion, then tokens may be predicted as a sequence: for example a "(" will surely come after "for", etc. (first assumption is using one-hot encoding of words instead of word bagging)
1- Is that possible and If yes can you please provide a lightweight link or document that already implemented this as a reference ? If no why not ?
2- This is not considered a Denoising autoencoders, as no noise is intentionally introduced in the input ?
3- Like mentioned above, can this architecture be later improved to also learn based on the word's position in the file ?