Machine learning traceback-enabled decision rationales as models for explainability
Techniques for providing decision rationales for machine-learning guided processes are described herein. In some embodiments, the techniques described herein include processing queries for an explanation of an outcome of a set of one or more decisions guided by one or more machine-learning processes with supervision by at least one human operator. Responsive to receiving the query, a system determines, based on a set of one or more rationale data structures, whether the outcome was caused by human operator error or the one or more machine-learning processes. The system then generates a query response indicating whether the outcome was caused by the human operator error or the one or more machine-learning processes.
1 . A method comprising:
generating, within a database, a rationale data structure associated with at least one artificial-intelligence guided process that is supervised by at least one human operator and triggers automated actions using a set of one or more machine learning models, wherein the rationale data structure includes a set of links tracing a sequence of decisions over time leading to an outcome, wherein the set of links is traversable backwards through time, wherein the set of links identifies: (a) a plurality of outputs of the set of one or more machine learning models, (b) an indication of which of the plurality of outputs resulted in a prompt for human feedback, and (c) at least one action associated with the at least one human operator that lead to the outcome;
receiving a query for an explanation of the outcome of the at least one artificial-intelligence guided process that is supervised by the at least one human operator;
responsive to receiving the query, determining, based on a traversal of the set of links associating the plurality of outputs of the set of one or more machine learning models and the at least one action associated with the at least one human operator, whether the outcome was caused by human operator error or the set of one or more machine-learning models;
generating a query response indicating whether the outcome was caused by human operator error or at least one output of the plurality of outputs of the set of one or more machine-learning models; and
responsive to determining that outcome was caused by at least one output of the plurality of outputs of the set of one or more machine learning models, updating, based on the rationale data structure, one or more machine-learning model parameters to isolate at least one problem associated with the artificial-intelligence guided process causing the outcome.
2 . The method of claim 1 , wherein determining whether the outcome was caused by human operator error comprises: determining that the human operator performed an action that was not induced by guidance of one or more machine-learning processes.
3 . The method of claim 1 , wherein determining whether the outcome was caused by human operator error comprises: determining that an output of a machine-learning model generated an alarm for a set of one or more monitored traces and that the human operator did not perform evasive or mitigating actions.
4 . The method of claim 1 , wherein determining whether the outcome was caused by human operator error comprises: determining that an output of the machine-learning model generated an alarm for a set of one or more monitored traces and that the human operator performed a wrong evasive or mitigating action.
5 . The method of claim 1 , wherein the rationale data structure is generated as a function of a set of one or more logs, metadata, or traces relevant to the output of a machine-learning model.
6 . The method of claim 1 , further comprising: receiving a second query for further details associated with the explanation of the outcome; responsive to the second query, identifying, within a hierarchy of rationale data structures, a second set of one or more rationale data structures that are linked to the rationale data structure; and generating a second response to the second query for the further details associated with the explanation of the outcome based on the second set of one or more rationale data structures.
7 . The method of claim 1 , further comprising: receiving a second query to explain why an alternative was not included in the explanation; responsive to the second query, determining whether the alternative was considered; and generating a second response to the second query based at least in part on whether the alternative was considered.
8 . The method of claim 1 , wherein generating the rationale data structure comprises: receiving a set of raw data from a plurality of sensors; generating, based on the set of raw data, a set of perceptions; and generating, based on the set of perceptions, a set of hypotheses that explain a perceived situation, wherein the set of links include at least one link between the set of perceptions and the set of hypotheses.
9 . The method of claim 1 , wherein the one or more machine-learning models generate an output using a deterministic process.
10 . The method of claim 1 , further comprising: generating, by a training process, a training dataset based at least in part on the explanation of the outcome; training, by the training process as a function of the training dataset, one or more other machine learning models.
11 . A non-transitory computer readable medium storing instructions which, when executed by one or more hardware processors, cause:
generating, within a database, a rationale data structure associated with at least one artificial-intelligence guided process that is supervised by at least one human operator and triggers automated actions using a set of one or more machine learning models, wherein the rationale data structure includes a set of links tracing a sequence of decisions over time leading to an outcome, wherein the set of links is traversable backwards through time, wherein the set of links identifies: (a) a plurality of outputs of the set of one or more machine learning models, (b) an indication of which of the plurality of outputs resulted in a prompt for human feedback, and (c) at least one action associated with the at least one human operator that lead to the outcome;
receiving a query for an explanation of the outcome of the at least one artificial-intelligence guided process that is supervised by the at least one human operator;
responsive to receiving the query, determining, based on a traversal of the set of links associating the plurality of outputs of the set of one or more machine learning models and the at least one action associated with the at least one human operator, whether the outcome was caused by human operator error or the set of one or more machine-learning models;
generating a query response indicating whether the outcome was caused by human operator error or at least one output of the plurality of outputs of the set of one or more machine-learning models; and
responsive to determining that outcome was caused by at least one output of the plurality of outputs of the set of one or more machine learning models, updating, based on the rationale data structure, one or more machine-learning model parameters to isolate at least one problem associated with the artificial-intelligence guided process causing the outcome.
12 . The non-transitory computer-readable medium of claim 11 , wherein determining whether the outcome was caused by human operator error comprises: determining that the human operator performed an action that was not induced by guidance of one or more machine-learning processes.
13 . The non-transitory computer-readable medium of claim 11 , wherein determining whether the outcome was caused by human operator error comprises: determining that an output of a machine-learning model generated an alarm for a set of one or more monitored traces and that the human operator did not perform evasive or mitigating actions.
14 . The non-transitory computer-readable medium of claim 11 , wherein determining whether the outcome was caused by human operator error comprises: determining that an output of the machine-learning model generated an alarm for a set of one or more monitored traces and that the human operator performed a wrong evasive or mitigating action.
15 . The non-transitory computer-readable medium of claim 11 , wherein the rationale data structure is generated as a function of a set of one or more logs, metadata, or traces relevant to the output of a machine-learning model.
16 . The non-transitory computer-readable medium of claim 11 , wherein the instructions further cause: receiving a second query for further details associated with the explanation of the outcome; responsive to the second query, identifying, within a hierarchy of rationale data structures, a second set of one or more rationale data structures that are linked to the rationale data structure; and generating a second response to the second query for the further details associated with the explanation of the outcome based on the second set of one or more rationale data structures.
17 . The non-transitory computer-readable medium of claim 11 , wherein the instructions further cause: receiving a second query to explain why an alternative was not included in the explanation; responsive to the second query, determining whether the alternative was considered; and generating a second response to the second query based at least in part on whether the alternative was considered.
18 . The non-transitory computer-readable medium of claim 11 , wherein generating the rationale data structure comprises: receiving a set of raw data from a plurality of sensors; generating, based on the set of raw data, a set of perceptions; and generating, based on the set of perceptions, a set of hypotheses that explain a perceived situation, wherein the set of links include at least one link between the set of perceptions and the set of hypotheses.
19 . The non-transitory computer-readable medium of claim 11 , the one or more machine-learning models generate an output using a deterministic process.
20 . A system comprising:
one or more hardware processors;
one or more non-transitory computer-readable media storing instructions which, when executed by the one or more hardware processors, cause:
generating, within a database, a rationale data structure associated with at least one artificial-intelligence guided process that is supervised by at least one human operator and triggers automated actions using a set of one or more machine learning models, wherein the rationale data structure includes a set of links tracing a sequence of decisions over time leading to an outcome, wherein the set of links is traversable backwards through time, wherein the set of links identifies: (a) a plurality of outputs of the set of one or more machine learning models, (b) an indication of which of the plurality of outputs resulted in a prompt for human feedback, and (c) at least one action associated with the at least one human operator that lead to the outcome;
receiving a query for an explanation of the outcome of the at least one artificial-intelligence guided process that is supervised by the at least one human operator;
responsive to receiving the query, determining, based on a traversal of the set of links associating the plurality of outputs of the set of one or more machine learning models and the at least one action associated with the at least one human operator, whether the outcome was caused by human operator error or the set of one or more machine-learning models;
generating a query response indicating whether the outcome was caused by human operator error or at least one output of the plurality of outputs of the set of one or more machine-learning models; and
responsive to determining that outcome was caused by at least one output of the plurality of outputs of the set of one or more machine learning models, updating, based on the rationale data structure, one or more machine-learning model parameters to isolate at least one problem associated with the artificial-intelligence guided process causing the outcome.