Explainable artificial intelligence mechanism
View Patent ↗A method of determining and explaining an artificial intelligence, AI, system employing an opaque model from a local or global point of view, the method comprising the steps of providing an input and a corresponding output of the opaque model; sampling the opaque model around the input to generate training data samples; performing feature selection to determine dominant features generating a Type-2 Fuzzy Logic Model, FLM; training the Type-2 FLM with the training data samples; and inputting the input into the Type-2 FLM to provide an explanation of the output from the opaque model.
1. A method of determining and explaining an existing artificial intelligence (AI) system, the existing AI system employing an opaque model from a local or global point of view, the method comprising the steps of:
providing an input and a corresponding output of the opaque model;
generating a Type-2 Fuzzy Logic Model (FLM), by training the Type-2 FLM with training data samples, wherein the data samples are synthetic or from the opaque model's evaluation; wherein generating the Type-2 FLM comprises:
generating a Restricted Universe of Rules (UoR);
pruning the UoR;
creating an initial set of rules; and
applying a pseudo-random search algorithm, to obtain a subset of rules from the initial set of rules;
and
inputting the input into the Type-2 FLM to provide an explanation of the output from the opaque model from a local or global point of view.
2. The method according to claim 1 , wherein the method further comprises, after the steps of providing an input and before the step of generating a Type-2 Fuzzy Logic Module (FLM), sampling the opaque model around the input to generate training data samples; and
performing feature selection to determine dominant features;
wherein the explanation of the output is from a local point of view.
3. The method according to claim 2 , further comprising the step of, after sampling the opaque model and before performing feature selection, inputting the training data samples into the opaque model to determine a score and/or classification output.
4. The method according to claim 2 , further comprising the step of computing a distance between the input and a data point comprised in the generated training data samples.
5. The method according to claim 4 , further comprising the step of generating, using the computed distance, a weight of the data point.
6. The method according to claim 2 , wherein sampling the opaque model comprises, for each sample:
S1: Setting a sample value z equal to the input x;
S2: Providing a randomly generated number N, wherein N is between 1 and Nf, the number of features to be changed;
S3: Shuffling each chosen feature randomly; and
repeating steps S1 to S3 for each sample.
7. The method according to claim 2 , wherein performing feature selection comprises using step-wise linear regression feature selection, to select dominant features.
8. The method according to claim 2 , further comprising the step of resampling by repeating the step of sampling the opaque model around the input, using the determined dominant features, to generate secondary training data samples.
9. The method according to claim 2 , wherein the Type-2 FLM is local.
10. The method according to claim 1 , wherein the Type-2 FLM is global and wherein the explanation of the output is from a global point of view.
11. A system for determining and explaining an existing artificial intelligence(AI) system, the existing AI system employing an opaque model from a local or global point of view, the system comprising a processor adapted to perform the steps of:
providing an input and a corresponding output of the opaque model;
generating a Type-2 Fuzzy Logic Model (FLM), by training the Type-2 FLM with training data samples wherein generating the Type-2 FLM comprises:
generating a Restricted Universe of Rules (UoR);
pruning the UoR;
creating an initial set of rules; and
applying a pseudo-random search algorithm, to obtain a subset of rules from the initial set of rules; and
inputting the input into the Type-2 FLM to provide an explanation of the output from the opaque model.
12. A method of determining and explaining an existing artificial intelligence (AI) system, the existing AI system employing an opaque model from a local or global point of view, the method comprising the steps of:
providing an input and a corresponding output of the opaque model;
generating a Type-2 Fuzzy Logic Model (FLM), by training the Type-2 FLM with training data samples, wherein the data samples are synthetic or from the opaque model's evaluation; and
inputting the input into the Type-2 FLM to provide an explanation of the output from the opaque model from a local or global point of view.