System for visually diagnosing machine learning models

Patent metadata

Field	Value
Publication number	US-11537506-B1
Application number	US-201816172637-A
Country	US
Kind code	B1
Filing date	Oct 26, 2018
Priority date	Oct 26, 2018
Publication date	Dec 27, 2022
Grant date	Dec 27, 2022

Abstract

Computer systems and associated methods are disclosed to implement a model development environment (MDE) that allows a team of users to perform iterative model experiments to develop machine learning (ML) media models. In embodiments, the MDE implements a media data management interface that allows users to annotate and manage training data for models. In embodiments, the MDE implements a model experimentation interface that allows users to configure and run model experiments, which include a training run and a test run of a model. In embodiments, the MDE implements a model diagnosis interface that displays the model's performance metrics and allows users to visually inspect media samples that were used during the model experiment to determine corrective actions to improve model performance for later iterations of experiments. In embodiments, the MDE allows different types of users to collaborate on a series of model experiments to build an optimal media model.

First claim

What is claimed is:

1. A system comprising:
one or more computers that implement a model diagnosis system for machine learning (ML) image models, configured to:
obtain prediction results of a ML image model that classifies a given image to one or more of a plurality of classes, wherein the ML image model was trained using training images from a training set and the prediction results were generated using test images from another set;
generate a model performance interface configured with a zoomable confusion matrix that groups the test images into cells according to their respective truth classes and predicted classes;
responsive to user input via the model performance interface to select a cell of the confusion matrix, cause the model performance interface to zoom in on test images in the cell selected, wherein the test images represent classification errors by the ML image model;
responsive to user input to select one or more of the test images in the cell, generate a model diagnosis interface configured to display the one or more test images with added visual information for a visual diagnosis of the classification errors;
receive user input via the model diagnosis interface indicating feedback regarding the one or more test images selected or one or more training images in the training set;
determine, based at least in part on the feedback, one or more modifications to the training set or the ML image model to improve prediction performance;
generate a diagnosis report interface configured to display the one or more modifications to the training set or the ML image model and to perform the one or more modifications via one or more user control elements; and
responsive to user input receive via the diagnosis report interface, perform at least one of the one or more modifications on the training set or the ML image model.

2. The system of claim 1 , wherein the model diagnosis system is implemented as a multi-tenant service and configured to obtain prediction results of a plurality of ML image models of a plurality of different tenants and determine modifications to training sets or ML image models of the plurality of different tenants.

3. The system of claim 1 , wherein the one or more modifications includes one or more of: adding a test image to the training set, changing annotations of one or more existing images in the training set, and combining two or more classes in the training set.

4. The system of claim 1 , wherein the model diagnosis system is configured to display a saliency map for a test image, wherein the saliency map indicates one or more regions in the test image that were salient in contributing to a prediction result of the ML image model.

5. The system of claim 1 , wherein the model diagnosis system is configured to cause a model diagnosis interface to display, for a test image, a closest image in a most likely predicted class from the training set and another closest image in a second most likely predicted class in the training set.

6. A computer-implemented method, comprising:
performing, by a model diagnosis system implemented on one or more processors and associated memory:
obtaining prediction results of a ML model that classifies a given media sample to one or more of a plurality of classes, wherein the ML model was trained using training samples from a training set and the prediction results were generated using test samples from another set;
generating a model performance interface with a zoomable confusion matrix that groups the test samples into cells according to their respective truth classes and predicted classes;
responsive to user input via the model performance interface to select a cell of the confusion matrix, causing the model performance interface to zoom in on test samples in the cell selected, wherein the test samples represent classification errors by the ML model;
responsive to user input to select one or more of the test samples in the cell, generating a model diagnosis interface configured to display the one or more test samples with added visual information for a visual diagnosis of the classification errors;
receiving user input via the model diagnosis interface indicating feedback regarding the one or more test samples selected or one or more training samples in the training set;
determining, based at least in part on the feedback, one or more modifications to the training set or the ML model to improve prediction performance;
generating a diagnosis report interface configured to display the one or more modifications to the training set or the ML model and to perform the one or more modifications via one or more user control elements; and
responsive to user input receive via the diagnosis report interface, performing at least one of the one or more modifications on the training set or the ML model.

7. The method of claim 6 , wherein the model diagnosis system is implemented as a multi-tenant service, and further comprising performing, by the model diagnosis system:
obtaining prediction results of a plurality of ML models of a plurality of different tenants; and
determining modifications to training sets or ML models of the plurality of different tenants.

8. The method of claim 6 , wherein the one or more modifications includes one or more of: adding a test sample to the training set, changing annotations of one or more existing samples in the training set, and combining two or more classes in the training set.

9. The method of claim 6 , wherein generating the diagnosis report interface comprises indicating on the diagnosis report interface a plurality of modifications to the training set or the ML model, wherein the modifications are prioritized based at least in part on respective impacts of the modifications on prediction performance.

10. The method of claim 6 , wherein generating the model diagnosis interface comprises including on the model diagnosis interface, for a test sample, a list of closest samples from the training set with their respective classes.

11. The method of claim 6 , wherein generating the model diagnosis interface comprises including on the model diagnosis interface, for a test sample, a closest sample in a most likely predicted class from the training set and another closest sample in a second most likely predicted class the training set.

12. The method of claim 11 , wherein determining the closest sample in the most likely predicted class for the test sample comprises:
obtaining a feature vector for the test sample used by the ML model;
computing distances between the feature vector and respective feature vectors of at least some training samples in the most likely predicted class via a distance metric; and
selecting a training sample in the most likely predicted class with a smallest distance as the closest sample.

13. The method of claim 6 , wherein generating the model diagnosis interface comprises including on the model diagnosis interface a saliency map for a test sample, wherein the saliency map indicates one or more regions in the test sample that were salient in contributing to a prediction result of the ML model.

14. The method of claim 13 , wherein generating the model diagnosis interface comprises generating a second saliency map for the test sample, wherein the second saliency map indicates one or more other regions in the test sample that are salient to the ML model to classify the test sample to a different class from the prediction result.

15. The method of claim 13 , wherein generating the model diagnosis interface comprises generating a bounding box surrounding a most salient region of the test sample.

Assignees

• Amazon Tech Inc

Inventors

• Dasgupta Sunny
• Pavani Sri Kaushik
• Venkataramana Shriram
• Mittal Rajul

Classifications

• G06F11/3495

  for systems · CPC title

• G06F11/34

  Recording or statistical evaluation of computer activity, e.g. of down time, of input/output operation {; Recording or statistical evaluation of user activity, e.g. usability assessment} · CPC title

• G06F11/1476

  in neural networks · CPC title

• G06F18/2431

  Multiple classes · CPC title

• G06N20/00

  Machine learning · CPC title