Training artificial neural networks using context-dependent gating with weight stabilization
Inventors
Masse, Nicolas Y. • Grant, Gregory D. • Freedman, David J.
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12579440-B2
Publication Date
2026-03-17
Expiration Date
Abstract
A computing device may receive a first set of training data for training an ANN to predict output data for a first task, and may train the ANN with the first set of training data by only adjusting values of weights associated with a first subset of neurons, the first subset selected based on an identity of the first task. The computing device may receive a second, different set of training data for training the ANN to predict output data for a second task, and may train the ANN with the second set of training data by only adjusting values of weights associated with a second subset of neurons, the second subset selected based on an identity of the second task. During training, adjusting of the value of any weight may entail weight stabilization that depends on whether there has been any training for one or more previous tasks.
Core Innovation
The invention relates to a continual learning approach for an artificial neural network (ANN) that reduces catastrophic forgetting when applying the ANN to sequential tasks. A context-dependent gating mechanism activates only a task-specific subset of neurons while gating other neurons, so only corresponding weights are updated for a given task. The neuron subsets for different tasks partially overlap, with nonzero numbers of neurons belonging exclusively to a first subset, exclusively to a second subset, and to both subsets.
During training, the ANN is trained to predict ground truth data for different task identities by adjusting values of only the weights associated with the neuron subset selected based on the identity of the task type. In addition to gating, the invention uses weight stabilization that biases training updates depending on previously learned task types. The bias is applied so that adjusting a given weight is dependent on whether the ANN has been previously trained for one or more task types different from the given type, corresponding to synaptic stabilization such as synaptic intelligence (SI) and/or elastic weight consolidation (EWC).
The documented approach combines task-dependent neuron gating with importance-based weight stabilization to mitigate catastrophic forgetting more effectively than using either mechanism alone. It includes empirical motivation and results using sequential permuted-MNIST and ImageNet task sequences, where gating alone or stabilization alone partially mitigates forgetting, while together substantially improve accuracy and slow degradation over many sequential tasks. The description further includes examples of gating implementation via multiplication that zeroes activation pathways, and represents gating via gating table or matrix structures, including two-dimensional and three-dimensional representations linked to task identities, hidden layers, and neurons.
Claims Coverage
The independent claims provide coverage for a computer-implemented method and a computing device that use context-dependent gating to activate task-specific neuron subsets with partial overlap, and perform training where weight updates for any given task are biased depending on prior training for other task types. Together, the claims specify three main inventive features: task identity-based neuron subset activation, weight-update bias based on prior task exposure, and structured gating-table representations that correlate task types to neuron gating decisions.
Task identity-based activation of partially overlapping neuron subsets
Applying an artificial neural network (ANN) such that only a first subset of neurons of the ANN are activated when applying the ANN to data associated with a first type of task, and only a second subset of neurons of the ANN are activated when applying the ANN to data associated with a second type of task, wherein the first subset and second subset partially overlap such that nonzero numbers of neurons belong exclusively to each subset and a nonzero number of neurons belong to both subsets.
Training with weight update bias depending on prior task types
During training of the ANN for any given type of task, adjusting of the value of any given weight associated with neurons of the ANN is subject to a bias in dependence of whether or not the ANN has been previously trained for one or more task types different from the given type.
Training by adjusting weights only for task-selected neuron subsets
Training the ANN to predict ground truth data for a first type of task using a first set of training data, wherein during training values of only those weights associated with the first subset of neurons are adjusted and the first subset of neurons is selected based on an identity of the first type of task, and training the ANN for a second type of task using a second set of training data, wherein during training values of only those weights associated with the second subset of neurons are adjusted and the second subset of neurons is selected based on an identity of the second type of task.
Task-correlated gating table for neuron activation
Applying a gating table that maps task types to binary activation/gating decisions for neurons in the hidden layers, wherein a neuron associated with an entry in the gating table is either activated or gated during training depending on the task type.
Gating controlled by a predefined optimal percentage
Assigning whether a neuron is activated or gated for a task based in part on a predefined optimal percentage of neurons to gate.
Two-dimensional gating table mapping task types to neuron columns
Structuring the gating table as a two-dimensional table whose rows map to different task types and whose columns map to neurons from the intermediate hidden layers of the ANN.
Three-dimensional gating table as stacked two-dimensional gating matrices
Defining the gating table as a three-dimensional structure formed by stacking like-sized two-dimensional gating matrices, where tasks correspond to the third dimension and each two-dimensional matrix maps columns to intermediate hidden layers of an ANN and rows to neuron positions within those layers.
The independent claims cover continual learning for an ANN where task identities select neuron subsets for activation, those subsets partially overlap, and training adjusts only weights tied to the selected subsets while applying a bias to weight adjustment depending on whether the ANN previously trained on different task types. Dependent claim coverage further specifies gating-table mechanisms and representations, including task-correlated binary gating, optional gating governed by a predefined optimal percentage, and two-dimensional and three-dimensional gating table structures.
Stated Advantages
Reduces catastrophic forgetting when applying the ANN to sequential tasks.
Gating alone or stabilization alone partially mitigates forgetting, while together yields substantially higher accuracy and slower degradation over many sequential tasks.
Improves accuracy and slows degradation over many sequential tasks.
Documented Applications
Continual learning across sequential permuted-MNIST task sequences.
Continual learning across ImageNet task sequences.
Interested in licensing this patent?