Back to results
US 12,755,194 B2Granted
Systems and methods for adjustment-based causally robust prediction
AssigneesToyota Research Institute, Inc., Toyota Jidosha Kabushiki Kaisha
InventorsGuy Rosman, Igor Gilitschenski, Xiongyi Cui, Stephen G. Mcgill
Life of this patent
Estimating the term…
Abstract
Prediction training systems that rely on small variation data sets instead of training the model using large passive data sets are disclosed. The smaller variation data sets are used to add loss terms that may mimic intervention. One or more models may be included that mimic the intervention by training with variation datasets. The variation datasets may be collected from such interventions in real world events. The model may mimic an intervention by replacing values in the prediction during a forward model computation.
Claims20
A method of training a prediction model, the method comprising:
- retrieving a first dataset, the first dataset comprising a base set and a first variation dataset;
- encoding, by a first encoder neural network, the first dataset to generate a first encoded dataset;
- encoding, by the first encoder neural network, the base set to generate a complement encoded dataset;
- retrieving a predicate dataset from a second dataset, the second dataset comprising a second variation dataset distinct from the first variation dataset;
- encoding, by a second encoder neural network, the predicate dataset to generate an encoded predicate dataset;
- concatenating the first encoded representation with the encoded predicate dataset to generate a concatenated dataset;
- decoding the concatenated dataset to generate a child prediction result;
- decoding the complement encoded data to generate a child complement prediction result;
- determining a robustness loss as a function of a difference between a combination of the concatenated dataset and the child predication result and a combination of the complement encoded dataset and the child complement prediction result, the robustness loss indicating a robustness of the prediction model to dataset variation;
- and
- updating parameters of the prediction model based on the robustness loss.
8 more elements
A system for training a prediction model, the system comprising:
- one or more processors;
- a memory communicably coupled to the one or more processors and storing a prediction systems, the prediction system including instructions that when executed by the one or more processors train the prediction model by:
- retrieving a first dataset, the first dataset comprising a base set and a first variation dataset;
- encoding, by a first encoder neural network, the first dataset to generate a first encoded dataset;
- encoding, by the first encoder neural network, the base set to generate a complement encoded dataset;
- retrieving a predicate dataset from a second dataset, the second dataset comprising a second variation dataset distinct from the first variation dataset;
- encoding, by a second encoder neural network, the predicate dataset to generate an encoded predicate dataset;
- concatenating the first encoded representation with the encoded predicate-dataset to generate a concatenated dataset;
- decoding the concatenated dataset to generate a child prediction result;
- decoding the complement encoded data to generate a child complement prediction result;
- determining a robustness loss as a function of a difference between a combination of the concatenated dataset and the child predication result and a combination of the complement encoded dataset and the child complement prediction result, the robustness loss indicating a robustness of the prediction model to dataset variation;
- and
- updating parameters of the prediction model based on the robustness loss.
10 more elements
A non-transitory computer-readable medium having program code recorded thereon for training a prediction model, the program code executed by a processor and comprising:
- program code for retrieving a first dataset, the first dataset comprising a base set and a first variation dataset;
- program code for encoding, by a first encoder neural network, the first dataset to generate a first encoded dataset;
- program code for encoding, by the first encoder neural network, the base set to generate a complement encoded dataset;
- program code for retrieving a predicate dataset from a second dataset, the second dataset comprising a second variation dataset distinct from the first variation dataset;
- program code for encoding, by a second encoder neural network, the predicate dataset to generate an encoded predicate dataset;
- program code for concatenating the first encoded representation with the encoded predicate-dataset to generate a concatenated dataset;
- program code for decoding the concatenated dataset to generate a child prediction result;
- program code for decoding the complement encoded data to generate a child complement prediction result;
- program code for determining a robustness loss as a function of a difference between a combination of the concatenated dataset and the child predication result and a combination of the complement encoded dataset and the child complement prediction result, the robustness loss indicating a robustness of the prediction model to dataset variation;
- and
- program code for updating parameters of the prediction model based on the robustness loss.
8 more elements
Dependencies are inferred from the claim text.