Methods and systems for generating an uncertainty score for an output of a gradient boosted decision tree model
A method of generating an uncertainty score for an output of a Gradient Boosted Decision Tree (GBDT) model is disclosed. The output is a prediction of the GBDT model for an in-use dataset. The method comprises acquiring the GBDT model including a sequence of trees beginning with an initial tree and ending with a last tree, a given one of the sequence of trees having been stochastically built during a current training iteration of the GBDT model, and defining a plurality of sub-sequences of trees in the sequence of trees as sub-models of the GBDT model. During a given in-use iteration of the GBDT model executable for the in-use dataset, the method comprises generating a plurality of sub-outputs using the respective sub-models and generating the uncertainty score using the plurality of sub-outputs, the uncertainty score being indicative of how different sub-outputs from the plurality of sub-outputs are amongst each other.
1 . A method of generating an uncertainty score for an output of a Gradient Boosted Decision Tree (GBDT) model for a self-driving vehicle, the output being a prediction of the GBDT model for an in-use dataset, the method executable by a server, the method comprising:
acquiring the GBDT model including a sequence of trees beginning with an initial tree and ending with a last tree,
a given one of the sequence of trees having been stochastically built during a training iteration of the GBDT model based on a sequence of prior trees available at the training iteration;
defining a plurality of sub-sequences of trees in the sequence of trees as sub-models of the GBDT model,
a given one of the plurality of sub-sequences of trees beginning with the initial tree and ending with a respective offset tree in the sequence of trees between the initial tree and the last tree, the respective offset tree being offset from the initial tree by an offset parameter, the offset parameter being different for each of the plurality of sub-sequences of trees;
during a given in-use iteration of the GBDT model executable for the in-use dataset:
generating a plurality of sub-outputs using the respective sub-models by inputting the in-use dataset to the respective sub-models,
wherein the in-use dataset is received from at least one of: an image sensor, a charge-coupled device (CCD) sensor, complementary metal oxide semiconductor sensor (CMOS), a digital camera, a light detection and ranging (LiDAR) sensor, a radar sensor, a time of flight (TOF) 3D camera, a stereo vision sensor or a 3D sonar associated with the self-driving vehicle,
a given one of the plurality of sub-outputs being generated by a respective one of the plurality of sub-sequences of trees;
generating the uncertainty score using the plurality of sub-outputs, the uncertainty score being indicative of how different sub-outputs from the plurality of sub-outputs are amongst each other;
in response to the uncertainty score being below a pre-determined threshold, providing the output of the GBDT model and the uncertainty score to an Object Detection Neural Network (ODNN) algorithm of the self-driving vehicle, the ODNN being configured for classifying objects in a surrounding of the self-driving vehicle based on the output of the GBDT model and the uncertainty score; and
controlling the self-driving vehicle based on the output of the ODNN.
2 . The method of claim 1 , wherein the method further comprises, during the given in-use iteration:
generating the output for the in-use dataset using the sequence of trees, the output being generated using all trees from the sequence of trees.
3 . The method of claim 1 , wherein a first one from the plurality of sub-sequences of trees is nested in a second one from the plurality of sub-sequences of trees.
4 . The method of claim 1 , wherein a first one from the plurality of sub-sequences of trees includes a total number of trees that is different from a total number of trees included in a second one from the plurality of sub-sequences of trees.
5 . The method of claim 1 , wherein a first one from the plurality of sub-sequences of trees and a second one from the plurality of sub-sequences of trees comprise a sub-set of same trees and a sub-set of differing trees.
6 . The method of claim 1 , wherein a first one from the plurality of sub-sequences of trees includes a second one from the plurality of sub-sequences of trees and at least one next tree from the sequence of trees.
7 . The method of claim 1 , wherein the given one from the plurality of sub-sequences of trees excludes the last tree in the sequence of trees.
8 . The method of claim 1 , wherein the GBDT model has been trained to classify the in-use dataset and the output is indicative of a predicted class of the in-use dataset.
9 . The method of claim 1 , wherein the GBDT model has been trained to perform regression on the in-use dataset and the output being a continuous value predicted for the in-use dataset.
10 . The method of claim 1 , wherein generating, the uncertainty score using the plurality of sub-outputs comprises determining a variance of the plurality of sub-outputs.
11 . A method of generating an uncertainty score for an output of a Gradient Boosted Decision Tree (GBDT) model for a self-driving vehicle, the output being a prediction of the GBDT model for an in-use dataset, the method executable by a server, the method comprising:
acquiring the GBDT model including a sequence of trees, a given one from the sequence of trees having been stochastically built during a training iteration of the GBDT model based on a sequence of trees available at the training iteration;
during a given in-use iteration of the GBDT model executable for the in-use dataset:
generating using the sequence of trees, a plurality of tree-outputs for the in-use dataset, a given one from the plurality of tree-outputs generated by a respective one from the sequence of trees, wherein the in-use dataset is received from at least one of: an image sensor, a charge-coupled device (CCD) sensor, complementary metal oxide semiconductor sensor (CMOS), a digital camera, a light detection and ranging (LiDAR) sensor, a radar sensor, a time of flight (TOF) 3D camera, a stereo vision sensor or a 3D sonar associated with the self-driving vehicle;
for a first sub-sequence of trees in the sequence of trees, generating in response to receiving the in-use dataset, a first sub-output by combining tree-outputs from the first sub-sequence of trees beginning with an initial tree, the first sub-sequence of trees including M number of sequential trees from the sequence of trees;
for a second sub-sequence of trees in the sequence of trees,
generating in response to receiving the in-use dataset, a second sub-output by combining tree-outputs from the second sub-sequence of trees beginning with the initial tree, the second sub-sequence of trees including N number of sequential trees from the sequence of trees, the N number being different from the M number; and
generating the uncertainty score using the first and the second sub-outputs, the uncertainty score being indicative of how different the first sub-output is from the second sub-output;
in response to the uncertainty score being below a pre-determined threshold, providing the output of the GBDT model and the uncertainty score to an Object Detection Neural Network (ODNN) algorithm of the self-driving vehicle, the ODNN being configured for classifying objects in a surrounding of the self-driving vehicle based on the output of the GBDT model and the uncertainty score; and
controlling the self-driving vehicle based on the output of the ODNN.
12 . The method of claim 11 , wherein the method further comprises:
generating the output for the GBDT model by combining all tree-outputs in the plurality of tree-outputs.
13 . The method of claim 11 , wherein the first sub-sequence of trees and the second sub-sequence of trees form a virtual ensemble in the sequence of trees.
14 . A server for generating an uncertainty score for an output of a Gradient Boosted Decision Tree (GBDT) model for a self-driving vehicle, the output being a prediction of the GBDT model for an in-use dataset, the server comprising at least one processor and a memory configured to store instructions which, upon being executed by the at least one processor, cause the server to:
acquire the GBDT model including a sequence of trees beginning with an initial tree and ending with a last tree,
a given one of the sequence of trees having been stochastically built during a training iteration of the GBDT model based on a sequence of prior trees available at the training iteration;
define a plurality of sub-sequences of trees in the sequence of trees as sub-models of the GBDT model,
a given one of the plurality of sub-sequences of trees beginning with the initial tree and ending with a respective offset tree in the sequence of trees between the initial tree and the last tree, the respective offset tree being offset from the initial tree by an offset parameter, the offset parameter being different for each of the plurality of sub-sequences of trees;
during a given in-use iteration of the GBDT model executable for the in-use dataset:
generate a plurality of sub-outputs using the respective sub-models by inputting the in-use dataset to the respective sub-models,
wherein the in-use dataset is received from at least one of: an image sensor, a charge-coupled device (CCD) sensor, complementary metal oxide semiconductor sensor (CMOS), a digital camera, a light detection and ranging (LiDAR) sensor, a radar sensor, a time of flight (TOF) 3D camera, a stereo vision sensor or a 3D sonar associated with the self-driving vehicle,
a given one of the plurality of sub-outputs being generated by a respective one of the plurality of sub-sequences of trees;
generate the uncertainty score using the plurality of sub-outputs, the uncertainty score being indicative of how different sub-outputs from the plurality of sub-outputs are amongst each other;
in response to the uncertainty score being below a pre-determined threshold, provide the output of the GBDT model and the uncertainty score to an Object Detection Neural Network (ODNN) algorithm of the self-driving vehicle, the ODNN being configured for classifying objects in a surrounding of the self-driving vehicle based on the output of the GBDT model and the uncertainty score; and
control the self-driving vehicle based on the output of the ODNN.
15 . The server of claim 14 , wherein the server is further configured to, during the given in-use iteration:
generate the output for the in-use dataset using the sequence of trees, the output being generated using all trees from the sequence of trees.
16 . The server of claim 14 , wherein a first one from the plurality of sub-sequences of trees is nested in a second one from the plurality of sub-sequences of trees.
17 . The server of claim 14 , wherein a first one from the plurality of sub-sequences of trees includes a total number of trees that is different from a total number of trees included in a second one from the plurality of sub-sequences of trees.
18 . The server of claim 14 , wherein a first one from the plurality of sub-sequences of trees and a second one from the plurality of sub-sequences of trees comprise a sub-set of same trees and a sub-set of differing trees.
19 . The server of claim 14 , wherein a first one from the plurality of sub-sequences of trees includes a second one from the plurality of sub-sequences of trees and at least one next tree from the sequence of trees.
20 . The server of claim 14 , wherein the given one from the plurality of sub-sequences of trees excludes the last tree in the sequence of trees.
21 . The server of claim 14 , wherein the GBDT model has been trained to classify the in-use dataset and the output is indicative of a predicted class of the in-use dataset.
22 . The server of claim 14 , wherein the GBDT model has been trained to perform regression on the in-use dataset and the output being a continuous value predicted for the in-use dataset.
23 . The server of claim 14 , further configured to determine a variance of the plurality of sub-outputs to generate the uncertainty score using the plurality of sub-outputs.