Modular adaptation for cross-domain few-shot learning

Inventors

Lin, Xiao • Ye, Meng • Gong, Yunye • BURACHAS, Giedrius T. • Divakaran, Ajay • Yao, Yi • BASIOU, Nikoletta

Assignees

SRI International Inc

Abstract

A method, apparatus and system for adapting a pre-trained network for application to a different dataset includes arranging at least two different types of active adaptation modules in a pipeline configuration, wherein an output of a previous active adaptation module produces an input for a next active adaptation module in the pipeline in the form of adapted network data until a last active adaptation module, and wherein each of the at least two different types of adaptation modules can be switched on or off, determining at least one respective hyperparameter for each of the at least two different types of active adaptation modules, and applying the at least one respective determined hyperparameter to each of the at least two different types of active adaptation modules for processing received data from the pretrained network to determine an adapted network.

Core Innovation

A cross-domain modular adaptation pipeline adapts a pre-trained network for application to a different dataset by arranging at least two different types of active adaptation modules in a pipeline configuration. The output of a previous active adaptation module produces an input for a next active adaptation module in the form of adapted network data until a last active adaptation module. Each of the at least two different types of adaptation modules can be switched on or off, enabling switchable pipeline configuration for classification task adaptation.

The method determines at least one respective hyperparameter for each of the at least two different types of active adaptation modules and applies the at least one respective determined hyperparameter to each of the active adaptation module types for processing received data from the pre-trained network to determine an adapted network. The received data from the pre-trained network includes at least one of classification data or data regarding an embedding space.

The pipeline supports adaptive determination and selection of module types and/or hyperparameters, including selection based on user input and/or an input from a machine learning process. It further supports using stored historically well-functioning collections of adaptation modules or hyperparameters from a storage device, as well as training a machine learning process to determine at least one adaptation module for the pipeline or a hyperparameter for an adaptation module using a target dataset. The process can be iterative so that rearranging, determining, and applying modify at least one active adaptation module type in the pipeline or a corresponding hyperparameter in each subsequent iteration.

Claims Coverage

The document includes three independent claims that cover a method, a non-transitory machine-readable medium, and a system. Across these independent claims, the core inventive features are the switchable pipeline of multiple active adaptation module types with module-to-module adapted network data flow, module-specific hyperparameter determination, and application of those hyperparameters to produce an adapted network.

Claim coverage centers on adapting a pre-trained network to a different dataset by composing at least two different active adaptation module types into a switchable pipeline that passes adapted network data module-to-module, determining respective hyperparameters for each module type, and applying those hyperparameters to received data from the pre-trained network to produce an adapted network.

Stated Advantages

Documented Applications