Unsupervised neighbor-preserving embedding for image stream visualization and anomaly detection
Methods and systems for detecting and correcting anomalous inputs include training a neural network to embed high-dimensional input data into a low-dimensional space with an embedding that preserves neighbor relationships. Input data items are embedded into the low-dimensional space to form respective low-dimensional codes. An anomaly is determined among the high-dimensional input data based on the low-dimensional codes. The anomaly is corrected.
1. A method for detecting and correcting anomalous inputs, comprising:
training a neural network to embed high-dimensional input data into a low-dimensional space with an embedding that preserves neighbor relationships;
embedding a plurality of input data items into the low-dimensional space to form respective low-dimensional codes;
determining an anomaly among the high-dimensional input data based on the low-dimensional codes; and
correcting the anomaly.
2. The method of claim 1 , wherein the high-dimensional input data comprises images of manufactured products.
3. The method of claim 2 , wherein correcting the anomaly comprises automatically discarding anomalous products.
4. The method of claim 1 , further comprising visualizing the high-dimensional input data based on neighbor relationships among the low-dimensional codes.
5. The method of claim 1 , wherein determining the anomaly comprises estimating a density of embeddings using a Gaussian mixture model and fitting a one-class support vector machine.
6. The method of claim 1 , wherein training the neural network comprises determining pairwise neighboring probabilities in a batch of high-dimensional input training data.
7. The method of claim 6 , wherein training the neural network further comprises embedding each high-dimensional input training data in the batch into the low-dimensional space.
8. The method of claim 7 , wherein training the neural network further comprises reconstructing high-dimensional data from the high-dimensional training data that has been embedded into the low-dimensional space.
9. The method of claim 8 , wherein training the neural network further comprises minimizing a Kullback-Leiber loss function, reconstruction errors, and hinge losses.
10. A system for detecting and correcting anomalous inputs, comprising:
a neural network configured to embed high-dimensional input data into a low-dimensional space;
a training module configured to train the neural network to embed the high-dimensional input data with an embedding that preserves neighbor relationships;
an anomaly detector configured to use the neural network to embed a plurality of input data items into the low-dimensional space to form respective low-dimensional codes and to determine anomaly among the high-dimensional input data based on the low-dimensional codes; and
an anomaly correction module configured to correct the anomaly.
11. The system of claim 10 , wherein the high-dimensional input data comprises images of manufactured products.
12. The system of claim 11 , wherein the anomaly correction module is further configured to automatically discard anomalous products.
13. The system of claim 10 , further comprising a user interface configured to visualize the high-dimensional input data based on neighbor relationships among the low-dimensional codes.
14. The system of claim 10 , wherein the anomaly detector is further configured to estimate a density of embeddings using a Gaussian mixture model and fitting a one-class support vector machine.
15. The system of claim 10 , wherein the training module is further configured to determine pairwise neighboring probabilities in a batch of high-dimensional input training data.
16. The system of claim 15 , wherein the training module is further configured to determine embed each high-dimensional input training data in the batch into the low-dimensional space.
17. The system of claim 16 , wherein the training module is further configured to determine reconstruct high-dimensional data from the high-dimensional training data that has been embedded into the low dimensional space.
18. The system of claim 17 , wherein the training module is further configured to determine minimize a loss function, reconstruction errors, and hinge losses.