17933473. CERTIFICATION-BASED ROBUST TRAINING BY REFINING DECISION BOUNDARY simplified abstract (Massachusetts Institute of Technology)
Contents
- 1 CERTIFICATION-BASED ROBUST TRAINING BY REFINING DECISION BOUNDARY
- 1.1 Organization Name
- 1.2 Inventor(s)
- 1.3 CERTIFICATION-BASED ROBUST TRAINING BY REFINING DECISION BOUNDARY - A simplified explanation of the abstract
- 1.4 Simplified Explanation
- 1.5 Potential Applications
- 1.6 Problems Solved
- 1.7 Benefits
- 1.8 Potential Commercial Applications
- 1.9 Possible Prior Art
- 1.10 Original Abstract Submitted
CERTIFICATION-BASED ROBUST TRAINING BY REFINING DECISION BOUNDARY
Organization Name
Massachusetts Institute of Technology
Inventor(s)
Lam Minh Nguyen of Ossining NY (US)
Wang Zhang of Cambridge MA (US)
Subhro Das of Cambridge MA (US)
Pin-Yu Chen of White Plains NY (US)
Alexandre Megretski of Acton MA (US)
Luca Daniel of Cambridge MA (US)
CERTIFICATION-BASED ROBUST TRAINING BY REFINING DECISION BOUNDARY - A simplified explanation of the abstract
This abstract first appeared for US patent application 17933473 titled 'CERTIFICATION-BASED ROBUST TRAINING BY REFINING DECISION BOUNDARY
Simplified Explanation
The abstract describes a computer-implemented method for certifying the robustness of image classification in a neural network. The method involves initializing a neural network model, receiving a data set of images, image labels, and a perturbation schedule, determining the distance from the decision boundary for the images, applying re-weighting values and modified perturbation magnitudes to the images, calculating a total loss function, and evaluating the confidence level of the classification for certifiable robustness.
- Initializing a neural network model
- Receiving a data set of images, image labels, and a perturbation schedule
- Determining the distance from the decision boundary for the images
- Applying re-weighting values and modified perturbation magnitudes to the images
- Calculating a total loss function
- Evaluating the confidence level of the classification for certifiable robustness
Potential Applications
This technology can be applied in various fields such as image recognition, autonomous vehicles, medical imaging, and security systems.
Problems Solved
This technology addresses the issue of ensuring the robustness of image classification in neural networks, which is crucial for applications where accuracy and reliability are paramount.
Benefits
The benefits of this technology include improved accuracy and reliability of image classification, enhanced security in systems relying on image recognition, and increased confidence in the performance of neural networks.
Potential Commercial Applications
Potential commercial applications of this technology include image recognition software, security systems, medical imaging devices, and autonomous vehicles.
Possible Prior Art
One possible prior art in this field is the work on adversarial attacks and defenses in neural networks, where researchers have explored methods to improve the robustness of image classification systems.
What are the specific perturbation schedules used in this method?
The specific perturbation schedules used in this method are not detailed in the abstract. Further information on the types of perturbations and their scheduling could provide insights into the effectiveness of the method.
How does the re-weighting value affect the classification results?
The abstract mentions applying a re-weighting value to the images, but it does not elaborate on how this affects the classification results. Understanding the impact of re-weighting on the classification process could help in optimizing the method for better performance.
Original Abstract Submitted
A computer implemented method for certifying robustness of image classification in a neural network is provided. The method includes initializing a neural network model. The neural network model includes a problem space and a decision boundary. A processor receives a data set of images, image labels, and a perturbation schedule. Images are drawn from the data set in the problem space. A distance from the decision boundary is determined for the images in the problem space. A re-weighting value is applied to the images. A modified perturbation magnitude is applied to the images. A total loss function for the images in the problem space is determined using the re-weighting value. A confidence level of the classification of the images in the data set is evaluated for certifiable robustness.
