Social Applications of AI Many Current Social . . . So Is There a . . . Examples of Unfair . . . Is There a Contradiction A Simplified Statistical . . . Between Statistics and A Simplified Fuzzy . . . General Description of . . . Fairness: From Intelligent Application to Our . . . Partial Confidence . . . Control to Explainable AI Home Page Title Page Christian Servin 1 and Vladik Kreinovich 2 ◭◭ ◮◮ 1 Computer Science and Information Technology ◭ ◮ Systems Department El Paso Community College (EPCC), 919 Hunter Dr. Page 1 of 39 El Paso, TX 79915-1908, USA cservin1@epcc.edu Go Back 2 University of Texas at El Paso Full Screen 500 W. University, El Paso, TX 79968, USA vladik@utep.edu Close Quit
Social Applications of AI Many Current Social . . . 1. Social Applications of AI So Is There a . . . • Recent AI techniques like deep learning have led to Examples of Unfair . . . many successful applications. A Simplified Statistical . . . A Simplified Fuzzy . . . • For example, we can apply deep learning to decide: General Description of . . . – whose loan applications should be approved and Application to Our . . . whose applications should be rejected, Partial Confidence . . . – and if approved, what interest should we charge. Home Page • We can apply deep learning to decide: Title Page ◭◭ ◮◮ – which candidates for graduate program to accept, – and for those accepted what financial benefits to ◭ ◮ offer as an enticement. Page 2 of 39 • In all such cases, we feed the system with numerous Go Back past examples of successes and failures. Full Screen Close Quit
Social Applications of AI Many Current Social . . . 2. Social Applications of AI (cont-d) So Is There a . . . • Based on these example, the systems predict whether Examples of Unfair . . . a given loan will be a success. A Simplified Statistical . . . A Simplified Fuzzy . . . • Statistically, these systems work well: they predict suc- General Description of . . . cess or failure better than human decision makers. Application to Our . . . • However, the results are often not satisfactory. Let us Partial Confidence . . . explain why. Home Page Title Page ◭◭ ◮◮ ◭ ◮ Page 3 of 39 Go Back Full Screen Close Quit
Social Applications of AI Many Current Social . . . 3. Many Current Social Applications of AI Are So Is There a . . . Unsatisfactory Examples of Unfair . . . • On average, loan applications from poorer geographic A Simplified Statistical . . . areas have a higher default rate. A Simplified Fuzzy . . . General Description of . . . • This is a known fact, and statistical methods underly- Application to Our . . . ing machine learning find this out. Partial Confidence . . . • As a result, the system naturally recommends rejection Home Page of all loans from these areas. Title Page • This is not fair to people with good credit record who ◭◭ ◮◮ happen to live in the not-so-good areas. ◭ ◮ • Moveover, it is also detrimental to the bank. Page 4 of 39 • Indeed, the bank will miss on profiting from such po- Go Back tentially successful loans. Full Screen • Similarly, in many disciplines women has a lower suc- cess rates in getting their PhDs than men. Close Quit
Social Applications of AI Many Current Social . . . 4. Many Current Social Applications of AI Are So Is There a . . . Unsatisfactory (cont-d) Examples of Unfair . . . • Women also, on average, take longer to succeed. A Simplified Statistical . . . A Simplified Fuzzy . . . • One of the main reasons for this is that raising children General Description of . . . requires much more efforts from women than from men. Application to Our . . . • A statistical system, crudely speaking, does not care Partial Confidence . . . about the reasons. Home Page • This system just takes this statistical fact into account Title Page and preferably selects males. ◭◭ ◮◮ • Not only this is not fair, this way the universities miss ◭ ◮ a lot of talent. Page 5 of 39 • And nowadays, with not much need for routine boring Go Back work, talent and creativity are extremely important. Full Screen • Talent and creativity should be nurtured, not rejected. Close Quit
Social Applications of AI Many Current Social . . . 5. So Is There a Contradiction Between Statistics So Is There a . . . And Fairness? Examples of Unfair . . . • It seems that if we want the systems to be fair: A Simplified Statistical . . . A Simplified Fuzzy . . . – we cannot rely on statistics only, General Description of . . . – we need to supplement statistics with additional Application to Our . . . fairness constraints. Partial Confidence . . . • The need for such constraints is usually formulated as Home Page the need for explainable AI . Title Page • The main idea behind explainable AI is that: ◭◭ ◮◮ – instead of relying on a machine learning system as ◭ ◮ a black box, Page 6 of 39 – we extract some rules from this system, Go Back – and if these rules are not fair, we replace them with Full Screen fairer rules. Close Quit
Social Applications of AI Many Current Social . . . 6. What We Show in This Talk So Is There a . . . • We show that the seeming inconsistency comes from Examples of Unfair . . . the fact that we use simplified statistical models. A Simplified Statistical . . . A Simplified Fuzzy . . . • We show that: General Description of . . . – a more detailed description of the corresponding Application to Our . . . uncertainty – probabilistic or fuzzy, Partial Confidence . . . – eliminates this seeming contradiction, and Home Page – enables the system to come up with fair decisions Title Page without any need for additional constraints. ◭◭ ◮◮ ◭ ◮ Page 7 of 39 Go Back Full Screen Close Quit
Social Applications of AI Many Current Social . . . 7. Examples of Unfair Decisions So Is There a . . . • We want to understand why the existing techniques Examples of Unfair . . . can lead to unfair solutions. A Simplified Statistical . . . A Simplified Fuzzy . . . • So let us trace some detailed simplified examples. General Description of . . . • We will start with statistical examples. Application to Our . . . • Then, we will show that: Partial Confidence . . . Home Page – mathematically similar examples – this time not Title Page related to fairness, ◭◭ ◮◮ – can be found in applications of fuzzy techniques as well, ◭ ◮ – namely, when we apply the usual intelligent control Page 8 of 39 techniques. Go Back Full Screen Close Quit
Social Applications of AI Many Current Social . . . 8. A Simplified Statistical Example So Is There a . . . • Let us consider a statistical version of a classical AI Examples of Unfair . . . example: A Simplified Statistical . . . A Simplified Fuzzy . . . – birds normally fly, General Description of . . . – penguins are birds, Application to Our . . . – penguins normally do not fly, and Partial Confidence . . . – Sam is a penguin. Home Page • The question is: does Sam fly? Title Page ◭◭ ◮◮ • To make it into a statistical example, let us add some probabilities. ◭ ◮ • Let us assume: Page 9 of 39 – that 90% of the birds fly, and Go Back – that 99% of the penguins do not fly. Full Screen Close Quit
Social Applications of AI Many Current Social . . . 9. A Simplified Example (cont-d) So Is There a . . . • Of course, in reality, 100% of the penguins do not fly. Examples of Unfair . . . A Simplified Statistical . . . • However, let us keep it under 100% since in most real- A Simplified Fuzzy . . . life situations, we are never 100% sure about anything. General Description of . . . • From the viewpoint of common sense, the information Application to Our . . . about birds flying in general is rather irrelevant. Partial Confidence . . . Home Page • Indeed, we know that Sam is not just any bird, it is a penguin. Title Page • Penguins are very specific type of bird for which we ◭◭ ◮◮ know the probability of flying. ◭ ◮ • So, to find the probability of Sam flying, we should Page 10 of 39 only take into account information about penguins. Go Back • Thus, we should conclude that the probability of Sam Full Screen flying is 100 − 99 = 1%. Close Quit
Social Applications of AI Many Current Social . . . 10. A Simplified Example (cont-d) So Is There a . . . • However, this is not what we would get if we use the Examples of Unfair . . . standard statistical techniques. A Simplified Statistical . . . A Simplified Fuzzy . . . • Indeed, from the purely statistical viewpoint, here we General Description of . . . have two rules that lead us to two different conclusions: Application to Our . . . – since Sam is a bird, we can make a conclusion A Partial Confidence . . . that Sam flies, with probability a = 90%; and Home Page – since Sam is a penguin, we can make a conclusion Title Page B that Sam does not fly, with probability b = 99%. ◭◭ ◮◮ • These two conclusions cannot be both right. ◭ ◮ • Indeed, the probabilities of Sam flying and not flying Page 11 of 39 should add up to 1, and here we have Go Back 0 . 9 + 0 . 99 = 1 . 89 > 1 . Full Screen • This means that these conclusions are inconsistent. Close Quit
Recommend
More recommend
