IP Library › Granted Patent US 12688676
Granted Patent B2
US 12688676 · App. 18/325,436 · Granted Jul 21, 2026

Out-of-distribution detection using a neural network

Inventors: Ibrahima Ndiour (Portland, OR); Nilesh Ahuja (Cupertino, CA); Ranganath Krishnan (Hillsboro, OR); Mahesh Subedar (Portland, OR); Omesh Tickoo (Portland, OR); Ergin Genc (Portland, OR)
Assignee: Intel Corporation
G06V10/7715G06V10/80G06V10/82
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12688676
App. No.
18/325,436
Granted
Jul 21, 2026
Kind
B2
Abstract

Features extracted from one or more layers of a trained deep neural network (DNN) are used to detect out-of-distribution (OOD) data, such as anomalies. An OOD detection process includes transforming a feature output from a layer of the DNN from a relatively high-dimensional feature space to a lower-dimensional space, and then performing a reverse transformation back to the higher-dimensional feature space, resulting in a reconstructed feature. A feature reconstruction error is calculated based on a difference between the reconstructed feature and the original feature output from the DNN. The OOD detection process may further include calculating a score based on the feature reconstruction error and generating a visual representation of the feature reconstruction error.

Claims (62)

1 . A computer-implemented method, comprising:

receiving an output feature map output by an intermediate layer of a neural network, the neural network comprising a first layer to receive a representation of an input image and at least one intermediate layer following the first layer;

applying a forward transformation to the output feature map to generate an embedding, the forward transformation reducing a dimensionality of the output feature map, wherein applying the forward transformation to the output feature map to generate the embedding comprises vectorizing the output feature map to generate a vectorized feature, the vectorized feature having a lower rank than the output feature map, and reducing a dimensionality of the vectorized feature to generate the embedding, wherein the embedding has fewer elements than the vectorized feature;

performing a reverse transformation of the embedding to generate a reconstructed feature map, the reconstructed feature map having a same dimensionality as the output feature map;

determining a feature reconstruction error representing a difference between the output feature map and the reconstructed feature map; and

determining a detection score based on the feature reconstruction error, the detection score indicating whether the input image includes out-of-distribution data.

2 . The computer-implemented method of claim 1 , wherein applying the forward transformation to the output feature map to generate the embedding comprises:

performing an orthogonal linear transformation on the output feature map, the orthogonal linear transformation calculated from a training dataset using principal component analysis.

3 . The computer-implemented method of claim 1 , wherein applying the forward transformation to the output feature map to generate the embedding comprises:

applying a transformation learned from a training dataset using a nonlinear manifold learning technique.

4 . The computer-implemented method of claim 1 , wherein performing the reverse transformation of the embedding comprises:

applying a pseudo-inverse of the forward transformation to the embedding, wherein the pseudo-inverse has a same dimensionality as the output feature map.

5 . The computer-implemented method of claim 1 , the method further comprising:

generating a localization map of the feature reconstruction error, the localization map having dimensions corresponding to the input image, the localization map indicating where, in the input image, the out-of-distribution data is located.

6 . The computer-implemented method of claim 5 , wherein the feature reconstruction error is an error vector, and generating the localization map of the feature reconstruction error comprises:

rearranging the error vector to an error tensor, the error tensor having a same dimension as the output feature map;

performing a channel-wise averaging of the error tensor to generate the localization map; and

resizing the localization map to match the dimensions of the input image.

7 . The computer-implemented method of claim 5 , the method further comprising:

determining a second feature reconstruction error based on a second output feature map, the second output feature map obtained from an output of a second intermediate layer of the neural network;

generating a second localization map of the second feature reconstruction error; and

combining the localization map and the second localization map into a combined error localization map.

8 . The computer-implemented method of claim 7 , wherein combining the localization map and the second localization map comprises calculating a pixel-wise geometric average of the localization map and the second localization map.

9 . One or more non-transitory computer-readable media storing instructions executable to perform operations, the operations comprising:

receiving an output feature map output by an intermediate layer of a neural network, the neural network comprising a first layer to receive a representation of an input image and at least one intermediate layer following the first layer;

applying a forward transformation to the output feature map to generate an embedding, the forward transformation reducing a dimensionality of the output feature map, wherein applying the forward transformation to the output feature map to generate the embedding comprises vectorizing the output feature map to generate a vectorized feature, the vectorized feature having a lower rank than the output feature map, and reducing a dimensionality of the vectorized feature to generate the embedding, wherein the embedding has fewer elements than the vectorized feature;

performing a reverse transformation of the embedding to generate a reconstructed feature map, the reconstructed feature map having a same dimensionality as the output feature map;

determining a feature reconstruction error representing a difference between the output feature map and the reconstructed feature map; and

determining a detection score based on the feature reconstruction error, the detection score indicating whether the input image includes out-of-distribution data.

10 . The one or more non-transitory computer-readable media of claim 9 , wherein applying the forward transformation to the output feature map to generate the embedding comprises:

performing an orthogonal linear transformation on the output feature map, the orthogonal linear transformation calculated from a training dataset using principal component analysis.

11 . The one or more non-transitory computer-readable media of claim 9 , wherein applying the forward transformation to the output feature map to generate the embedding comprises:

applying a transformation learned from a training dataset using a nonlinear manifold learning technique.

12 . The one or more non-transitory computer-readable media of claim 9 , wherein performing the reverse transformation of the embedding comprises:

applying a pseudo-inverse of the forward transformation to the embedding, wherein the pseudo-inverse has a same dimensionality as the output feature map.

13 . The one or more non-transitory computer-readable media of claim 9 , the operations further comprising:

generating a localization map of the feature reconstruction error, the localization map having dimensions corresponding to the input image, the localization map indicating where, in the input image, the out-of-distribution data is located.

14 . The one or more non-transitory computer-readable media of claim 13 , wherein the feature reconstruction error is an error vector, and generating the localization map of the feature reconstruction error comprises:

rearranging the error vector to an error tensor, the error tensor having a same dimension as the output feature map;

performing a channel-wise averaging of the error tensor to generate the localization map; and

resizing the localization map to match the dimensions of the input image.

15 . The one or more non-transitory computer-readable media of claim 13 , the operations further comprising:

determining a second feature reconstruction error based on a second output feature map, the second output feature map obtained from an output of a second intermediate layer of the neural network;

generating a second localization map of the second feature reconstruction error; and

combining the localization map and the second localization map into a combined error localization map.

16 . The one or more non-transitory computer-readable media of claim 15 , wherein combining the localization map and the second localization map comprises calculating a pixel-wise geometric average of the localization map and the second localization map.

17 . An apparatus, comprising:

a computer processor for executing computer program instructions; and

a computer-readable memory storing computer program instructions executable by the computer processor to perform operations comprising:

receiving an output feature map output by an intermediate layer of a neural network, the neural network comprising a first layer to receive a representation of an input image and at least one intermediate layer following the first layer;

applying a forward transformation to the output feature map to generate an embedding, the forward transformation reducing a dimensionality of the output feature map, wherein applying the forward transformation to the output feature map to generate the embedding comprises vectorizing the output feature map to generate a vectorized feature, the vectorized feature having a lower rank than the output feature map, and reducing a dimensionality of the vectorized feature to generate the embedding, wherein the embedding has fewer elements than the vectorized feature;

performing a reverse transformation of the embedding to generate a reconstructed feature map, the reconstructed feature map having a same dimensionality as the output feature map;

determining a feature reconstruction error representing a difference between the output feature map and the reconstructed feature map; and

determining a detection score based on the feature reconstruction error, the detection score indicating whether the input image includes out-of-distribution data.

18 . The apparatus of claim 17 , the operations further comprising:

generating a localization map of the feature reconstruction error, the localization map having dimensions corresponding to the input image, the localization map indicating where, in the input image, the out-of-distribution data is located.

19 . The apparatus of claim 17 , the operations further comprising:

generating a localization map of the feature reconstruction error, the localization map having dimensions corresponding to the input image, the localization map indicating where, in the input image, the out-of-distribution data is located.

20 . The apparatus of claim 19 , wherein the feature reconstruction error is an error vector, and generating the localization map of the feature reconstruction error comprises:

rearranging the error vector to an error tensor, the error tensor having a same dimension as the output feature map;

performing a channel-wise averaging of the error tensor to generate the localization map; and

resizing the localization map to match the dimensions of the input image.