Back to results
US 2025/0053391 A1Published application
Method for converting models to programs
AssigneeMartian Learning, Inc.
InventorsShriyash K. Upadhyay, Etan J. Ginsberg
Life of this patent
Estimating the term…
Abstract
In variants, the method can include generating mapping model training data, determining the mapping model, and predicting a program based on a transformer. The method can optionally include evaluating the mapping model, running analyses on the program, and/or utilizing the program and/or generated program analyses. The method functions to convert transformer models into programs that can be characterized and/or analyzed using program analysis techniques.
Claims20
A method, comprising:
- determining a transformer representation of a test transformer using a transformer representation module;
- predicting a program based on the transformer representation using a mapping model, wherein the program comprises a set of explicit instructions and is configured to mimic logical processes performed by the test transformer, wherein the mapping model is trained by:
- determining a set of target programs;
- generating a set of training transformers from the set of target programs;
- and
- training the mapping model to predict target programs from the set of target programs given training transformers from the set of training transformers as training inputs;
- and
- receiving an input for the test transformer;
- and
- determining an output based on the input using the program in lieu of the test transformer.
7 more elements
A method, comprising:
- extracting a set of network weights from a deep neural network (DNN);
- using a representation model, determining a DNN representation based on the set of network weights;
- and
- using a mapping model, non-deterministically determining a program based on the DNN representation, wherein the program is configured to replicate a set of input-output functionalities of the DNN.
1 more element
Dependencies are inferred from the claim text.