Systems for modeling uncertainty in multi-modal retrieval and methods thereof
Multi-modal data representing driving events and corresponding actions related to the driving events can be obtained and used to train a neural network at least in part by using a triplet loss computed for the driving events as a regression loss to determine an embedding of driving event data. In some cases, using the trained neural network, a retrieval request for an input driving event and corresponding action can be processed by determining, from the neural network, one or more similar driving events or corresponding actions in the multi-modal data.
1. A method for performing retrieval operations in neural networks, comprising:
obtaining multi-modal data representing driving events and corresponding actions related to the driving events;
training a neural network based on the multi-modal data at least in part by using a triplet loss computed for the driving events as a regression loss to determine an embedding of driving event data;
processing a retrieval request for an input driving event and corresponding action by determining, from the neural network, one or more similar driving events or corresponding actions in the multi-modal data; and
indicating, in response to the retrieval request, at least a subset of the one or more similar driving events or corresponding actions.
2. The method of claim 1 , wherein training the neural network comprises adding one or more dropout layers after one or more convolutional layers of the neural network to drop one or more connections between the driving event data.
3. The method of claim 1 , wherein training the neural network comprises determining a distance between the driving event data as a Euclidean distance between the embedding of the driving event data.
4. The method of claim 1 , wherein the multi-modal data representing driving events is retrieved from an image sensor positioned on a vehicle and wherein the multi-modal data representing the corresponding actions related to the driving events is retrieved from one or more controller area network (CAN) sensors in a vehicle.
5. The method of claim 1 , wherein training the neural network comprises temporally associating the driving events and corresponding actions related to the driving events.
6. The method of claim 2 , further comprising obtaining multiple embeddings for the driving event data through multiple passes through the one or more dropout layers to determine the embedding of the driving event data.
7. The method of claim 2 , wherein the one or more dropout layers include dropout as a Bayesian approximation.
8. A computing device for performing retrieval operations in neural networks, comprising:
a memory; and
at least one processor coupled to the memory, wherein the at least one processor is configured to:
obtain multi-modal data representing driving events and corresponding actions related to the driving events;
train a neural network based on the multi-modal data at least in part by using a triplet loss computed for the driving events as a regression loss to determine an embedding of driving event data;
process a retrieval request for an input driving event and corresponding action by determining, from the neural network, one or more similar driving events or corresponding actions in the multi-modal data; and
indicate, in response to the retrieval request, at least a subset of the one or more similar driving events or corresponding actions.
9. The computing device of claim 8 , wherein the at least one processor is configured to train the neural network at least in part by adding one or more dropout layers after one or more convolutional layers of the neural network to drop one or more connections between the driving event data.
10. The computing device of claim 8 , wherein the at least one processor is configured to train the neural network at least in part by determining a distance between the driving event data as a Euclidean distance between the embedding of the driving event data.
11. The computing device of claim 8 , wherein the multi-modal data representing driving events is retrieved from an image sensor positioned on a vehicle and wherein the multi-modal data representing the corresponding actions related to the driving events is retrieved from one or more controller area network (CAN) sensors in a vehicle.
12. The computing device of claim 8 , wherein the at least one processor is configured to train the neural network at least in part by temporally associating the driving events and corresponding actions related to the driving events.
13. The computing device of claim 9 , wherein the at least one processor is further configured to obtain multiple embeddings for the driving event data through multiple passes through the one or more dropout layers to determine the embedding of the driving event data.
14. The computing device of claim 9 , wherein the one or more dropout layers include dropout as a Bayesian approximation.
15. A non-transitory computer-readable medium storing computer executable code for performing retrieval operations in neural networks, the code comprising code for:
obtaining multi-modal data representing driving events and corresponding actions related to the driving events;
training a neural network based on the multi-modal data at least in part by using a triplet loss computed for the driving events as a regression loss to determine an embedding of driving event data;
processing a retrieval request for an input driving event and corresponding action by determining, from the neural network, one or more similar driving events or corresponding actions in the multi-modal data; and
indicating, in response to the retrieval request, at least a subset of the one or more similar driving events or corresponding actions.
16. The non-transitory computer-readable medium of claim 15 , wherein the code for training the neural network adds one or more dropout layers after one or more convolutional layers of the neural network to drop one or more connections between the driving event data.
17. The non-transitory computer-readable medium of claim 15 , wherein the code for training the neural network determines a distance between the driving event data as a Euclidean distance between the embedding of the driving event data.
18. The non-transitory computer-readable medium of claim 15 , wherein the multi-modal data representing driving events is retrieved from an image sensor positioned on a vehicle and wherein the multi-modal data representing the corresponding actions related to the driving events is retrieved from one or more controller area network (CAN) sensors in a vehicle.
19. The non-transitory computer-readable medium of claim 16 , further comprising code for obtaining multiple embeddings for the driving event data through multiple passes through the one or more dropout layers to determine the embedding of the driving event data.
20. The non-transitory computer-readable medium of claim 16 , wherein the one or more dropout layers include dropout as a Bayesian approximation.