Methods, systems, and media for bi-modal generation of natural languages and neural architectures
Methods, systems, and computer-readable media for bi-modal generation of natural language (NL) and artificial neural network architectures (NA), with reference to an example implementation framework entitled “ArchGenBERT”. A model and method of training the model for bi-modal generation of NL and NA are described. The model trained for bi-modal generation of NL and NA can be deployed to perform a number of useful tasks to assist with designing, describing, translating, and modifying neural network architectures.
1. A method comprising:
obtaining a generative bi-modal model trained with a bi-modal understanding of natural language in relation to neural network architectures;
providing input information to the model, the input information comprising at least one of the following:
natural language information; and
neural network architecture information; and
using the model to:
encode the input information to generate encoded representations of the input information; and
decode the encoded representations of the input information to generate output information comprising at least one of:
natural language information; and
neural network architecture information,
wherein the input information comprises:
natural language information comprising a question; and
neural network architecture information corresponding to a first neural network architecture; and
wherein the output information comprises natural language information comprising an answer responsive to the question with respect to the first neural network architecture.
2. The method of claim 1 , wherein the model comprises:
a text encoder to process natural language information to generate word embeddings;
a neural network architecture encoder to process neural network architecture information to generate graph encodings;
a cross transformer encoder to process word embeddings and graph encodings to generate joint embeddings;
a cross transformer decoder to process joint embeddings to generate word embeddings and graph encodings;
a neural network architecture decoder to process graph encodings to generate neural network architecture information; and
a text decoder to process word embeddings to generate natural language information.
3. The method of claim 2 , wherein:
the text encoder comprises:
a tokenizer to process natural language information to generate a sequence of tokens; and
a word embedder to process the sequence of tokens to generate word embeddings.
4. The method of claim 2 , wherein:
the neural network architecture encoder comprises:
a graph generator to process neural network architecture information to generate a graph comprising a plurality of nodes, a plurality of edges, and a plurality of shapes;
a shape embedder to process the plurality of shapes to generate shape embeddings;
a node embedder to process the plurality of nodes to generate node embeddings;
a summation module to sum the shape embeddings and node embeddings to generate a shape-node summation; and
a graph attention network (GAT) for processing the summation and the plurality of edges to generate a graph encoding.
5. The method of claim 2 , wherein:
the text decoder comprises:
a word decoder to process word embeddings to generate a sequence of tokens; and
a beam search module to process the sequence of tokens, using a beam search algorithm, to generate natural language information.
6. The method of claim 2 , wherein:
the neural network architecture decoder comprises:
a graph attention network (GAT) decoder for processing graph encodings to generate a plurality of edges, a plurality of node embeddings, and a plurality of shape embeddings;
a shape decoder to process the plurality of shape embeddings to generate a plurality of shapes;
a node decoder to process the plurality of node embeddings to generate a plurality of nodes; and
a network architecture generator to process the plurality of nodes, the plurality of edges, and the plurality of shapes to generate neural network architecture information.
7. The method of claim 1 , wherein obtaining the model comprises:
providing a training dataset comprising:
a plurality of positive training samples, each positive training data sample comprising neural network architecture information associated with natural language information descriptive of the neural network architecture information; and
training the model, using supervised learning, to:
minimize a difference between the neural network architecture information and the natural language information of the positive training samples.
8. The method of claim 2 , wherein:
the model further comprises:
a pooling module to pool the joint embeddings to generate encoded representations comprising fixed-size one-dimensional (1D) representations; and
a similarity evaluator for processing encoded representations to determine a similarity measure using a cosine similarity metric; and
obtaining the model comprises:
providing a training dataset comprising:
a plurality of positive training samples, each positive training data sample comprising neural network architecture information associated with natural language information descriptive of the neural network architecture information; and
a plurality of negative training samples, each negative training data sample comprising neural network architecture information associated with natural language information not descriptive of the neural network architecture information; and
pre-training the text encoder and the neural network architecture encoder of the model, using supervised learning, to:
maximize a similarity measure generated between the neural network architecture information and the natural language information of the positive training samples; and
minimize the similarity measure generated between the neural network architecture information and the natural language information of the negative training samples.
9. The method of claim 1 , wherein obtaining the model comprises:
providing an additional training dataset comprising a plurality of additional training samples, each additional training data sample comprising:
neural network architecture information corresponding to a neural network architecture;
a first natural language data sample corresponding to a question; and
a second natural language data sample corresponding to an answer to the question with respect to the neural network architecture of the neural network architecture information; and
fine-tuning the model, using supervised learning, to associate, for each training data sample, the second natural language data sample with the neural network architecture information and the first natural language data sample.
10. The method of claim 1 , wherein:
the input information comprises natural language information comprising a textual description descriptive of a first neural network architecture; and
the output information comprises neural network architecture information corresponding to the first neural network architecture.
11. The method of claim 1 , wherein:
the input information comprises neural network architecture information corresponding to a first neural network architecture; and
the output information comprises natural language information comprising a textual description descriptive of the first neural network architecture.
12. The method of claim 1 , wherein:
the input information comprises neural network architecture information corresponding to a first neural network architecture in a first domain; and
the output information comprises neural network architecture information corresponding to the first neural network architecture in a second domain.
13. The method of claim 12 , wherein obtaining the model comprises:
providing an additional training dataset comprising a plurality of additional training samples, each additional training data sample comprising:
a first neural network architecture information data sample corresponding to a neural network architecture in the first domain; and
a second neural network architecture information data sample corresponding to the neural network architecture in the second domain; and
fine-tuning the model, using supervised learning, to associate, for each training data sample, the second neural network architecture information data sample with the first neural network architecture information data sample.
14. The method of claim 12 , wherein:
the input information further comprises natural language information comprising a textual description;
using the model to process the input information to generate the output information further comprises:
processing the natural language information, using the model, to generate an encoded representation of the natural language information; and
the output information comprises neural network architecture information that:
corresponds to the first neural network architecture in a second domain; and
is described by the textual description.
15. The method of claim 1 , wherein:
the input information comprises neural network architecture information corresponding to an incomplete version of a first neural network architecture; and
the output information comprises neural network architecture information corresponding to a complete version of the first neural network architecture.
16. The method of claim 15 , wherein obtaining the model comprises:
providing an additional training dataset comprising a plurality of additional training samples, each additional training data sample comprising:
a first neural network architecture information data sample corresponding to an incomplete version of a neural network architecture; and
a second neural network architecture information data sample corresponding to a complete version of the neural network architecture; and
fine-tuning the model, using supervised learning, to associate, for each training data sample, the second neural network architecture information data sample with the first neural network architecture information data sample.
17. The method of claim 15 , wherein:
the input information further comprises natural language information comprising a textual description;
using the model to process the input information to generate the output information further comprises:
processing the natural language information, using the model, to generate an encoded representation of the natural language information; and
the output information comprises neural network architecture information that:
corresponds to a complete version of the first neural network architecture; and
is described by the textual description.
18. A non-transitory computer-readable medium having instructions tangibly stored thereon that, when executed by a processing system, cause the processing system to perform the method of claim 1 .