Back to all patents
Expiring soonaiAssigned to DNNresearch Inc / Google LLC

System and method for addressing overfitting in a neural network

Dropout-style feature-detector disabling that reduces overfitting and unlocks more reliable neural network generalization.

What it is

The method trains a neural network by using a probabilistic switch to randomly disable feature detectors (units) in selected layers for each training example, then normalizes the contributions when applying the trained network to test data. Technically, it creates an ensemble effect from many thinned sub-networks, typically driven by fixed drop probabilities, rather than relying on a single deterministic model. Previously, using this regularization strategy at scale often required licensing or access to proprietary training techniques from organizations like the assignee.

What you could build

Build generalization-focused image and text classifiers that train with dropout regularization without licensing fees.

Patent number
7412306B2
Expiration
Invalid Da
Assignee
DNNresearch Inc / Google LLC
Inventors
Geoffrey E. Hinton, Alexander Krizhevsky, Ilya Sutskever, Nitish Srivastava

More in ai

Building on this patent?

Submit a problem you're solving with this technology and connect with founders, investors, and researchers.

System and method for addressing overfitting in a neural network | Expiring Patents | Questd | Questd