Reading the Robot Mind: - How to Avoid the Michigan Unemployment Fiasco



It used to be that only human experts examined data and made decisions. Now Artificial Intelligence (AI) is enabling robotic decision making in an ever-widening variety of applications. As society allows this to happen, there is a greater likelihood that these robot decisions can affect people’s lives.


The MiDAS System for Detecting Unemployment Fraud in the State of Michigan


A poster-child for AI gone bad is the Michigan MiDAS unemployment fraud detection software. On September 17, 2013, the Michigan Unemployment Insurance Agency announced that it would be shutting down for 5 days to install a MiDAS software system that would update the aging automated response system[1]. What ensued was a fiasco of epic proportions that is still costing the state of Michigan millions of dollars. In this case, 34,000 people were falsely accused of unemployment fraud by this robot, and they are still trying to dig out of their misery. A good overview of the problem appeared in this recent IEEE article [2] and you can read more about the possible causes [3], and subsequent legislation and lawsuits [4] in the news and ongoing.


What exactly have these AI robots learned so deeply from all of this big data?


Big data is a term used to describe both the opportunities and the problems associated with so much information now available for decision making. Actionable decisions need to be distilled from big data and AI can only go so far based on linear extrapolation. Because of this, many non-linear deep learning algorithms are being developed. The Michigan MiDAS example underscores the importance of having subject matter experts (not just the AI programmers, but the non-programmers as well) understand the capabilities and societal implications of AI robots.

This question is very reasonable for society to ask. It is not enough to train and create a great AI robot. Many researchers are realizing (or should have realized, in the MiDAS case) that before their systems can be deployed, they must be able to prove to human experts that the robots learned the right things from the right data. This is difficult because human expert decision makers are not necessarily the same people who are good at creating robot AI. These two teams must work together in a user friendly way.

If only we could read the robot mind.


There are several key questions researchers must ask when designing the AI systems, but the answers can only come from human subject matter experts - not the AI researchers! The robot AI must be designed not only to come up with the appropriate conclusions, classifications, and recognition of patterns, but it MUST be able to allow non-programmers to understand the inner workings of the robot "thought process."

As mentioned in [3], faulty data was found to be a root cause of the AI robot failure in the Michigan Unemployment Fraud Detection system. Not just the raw data from the real world, but the individual features extracted from that data and presented to the AI system. In order to train an AI system, example data must be provided. In addition, the example data must be distilled into a small subset of "features" that will be presented to the AI. Unless the AI researcher can demonstrate to the subject matter expert that the "features" are sufficient and accurate enough for an expert to render a decision - then how can we expect the robot to make a good decision? Many years ago, I proposed such a method to prove this to the subject matter experts (the CODEC test, described below), and there are other methods as well - but this step should not be skipped!

What is the CODEC test?


The CODEC test is something that I have proposed to answer the most important question - "Is the robot getting good data?" The CODEC test will help a subject matter expert read the robot's mind, and see if the robot is getting good data or not. A CODEC is a set of algorithms that takes real world data and turns it into codes (the Coder) and then encrypts and transmits those codes to a receiver who turns those codes back into real world data (the Decoder). Together, engineers call this a CODEC, and you are probably familiar with image (.jpg, .gif, etc.) and music (.mp3) and movie (mp4, etc.) CODECS. Here is how the AI Robot CODEC test works:

  • A - The AI Robot designer converts real world data into features. These features are usually a bunch of numbers that are derived from the form data (in the Michigan MiDAS case, the forms that were filled out regarding Unemployment). This is a form of a "Coder." Let's keep it very simple, and say that for every applicant of unemployment insurance, all the information in all of the forms is coded into a single number, from 1 to 10. (usually it will be a long string of numbers).
  • B - The original real world data is not used by the robot. Only the features are used. In the case of [3] it may be that the original data is thrown away! Bad idea. The CODEC test proposes that if we gathered enough features, there should be a way for the AI programmer to work backwards and create an example of real world data that would yield a specific set of feature numbers. So in our simplified example where the features are a single number between 1 and 10, the AI researcher takes each possible feature number and creates a fake set of example unemployment forms (exemplary forms, if you will) that would be coded to that same feature number. This is a form of a "Decoder."In our very simple example, there would only be 10 sets of forms created in this way.
  • C - The subject matter expert examines this exemplary real world data. They will quickly see whether there is sufficient information for them to come up with the appropriate conclusions, classifications, or recognition of patterns. In our simple example, if the fraud examiner looking at these 10 exemplary sets of forms can't tell the difference between them, or says that information to make a decision is missing, garbled, or otherwise not sufficient - then there is a problem.

As in the case of [4] the legislation has already been written in Michigan to make sure the subject matter expert is involved in every future fraud decision, but according to the CODEC test I proposed in the INTEGRAT software - they should have gotten involved much earlier! Let's hope programmers make sure that when they design AI robots in the future, they remember to put in the software needed to read the robot's mind...

Additional details, references, and software examples


With regard to that CODEC test, some old research of mine has been getting increased attention by engineers and scientists working on exactly the issues discussed above. This research yielded several US patents. I never publicly released the software, called INTEGRAT, because I sold rights to the patents to my employer at the time. Now that the patents have expired, I can share my work, as well as the software specifically designed to read the mind of the robot.

To make this research more easily accessible I am releasing the software for Windows, a short YouTube video explaining how to use the software, and the source code on Github.


There is also a wiki for INTEGRAT [6] that goes through each of the questions the subject matter expert needs to have answered, as well as software techniques to answer those questions.

[1] - http://www.wilx.com/home/headlines/Unemployment-Insurance-Agency-Gets-Upgrade-Services-Unavailable-224134801.html

[2] - https://spectrum.ieee.org/riskfactor/computing/software/michigans-midas-unemployment-system-algorithm-alchemy-that-created-lead-not-gold

[3] - https://www.freep.com/story/news/local/michigan/2017/07/30/fraud-charges-unemployment-jobless-claimants/516332001/

[4] - https://www.freep.com/story/news/local/michigan/2017/12/01/after-midas-false-fraud-fiasco-jobless-reform-bills-getting-speedy-passage-michigan-unemployment/908315001/

[5] - https://github.com/prof-nussbaum/INTEGRAT-reading-the-mind-of-the-AI-robot

[6] - https://github.com/prof-nussbaum/INTEGRAT-reading-the-mind-of-the-AI-robot/wiki

Comments

Post a Comment

Popular posts from this blog

Image
Reverse Inference Example Software:  This points to example software on GitHub where you can see some of these concepts in action. Here is the link: https://github.com/prof-nussbaum/Reverse_Inference  Reverse_Inference Working backwards through a deep convolutional network, to recreate the input image - and identify areas for improvement. Please see this article for more details on this technique, that I call "Reading the Robot Mind." https://readingtherobotmind.blogspot.com/2021/03/reading-robot-mind-deep-convolution.html What if we could read the robot's mind? It sounds like a silly statement, but if you could read a person's mind, you could "see" what they were thinking of when they mention a classification. If they say "I am seeing a dog" - reading their mind would give you additional details about the dog, or even perhaps, see what they are seeing. This is the same premise for Artificial Intelligence and Machine Learning. I know that deep conv...
Read more
Image
Auto-Associative How To - INCLUDING SOFTWARE CODE This repository includes "how to" examples of auto-associative neural networks. These networks have the ability to encode and decode (CODEC) sample data in the most efficient and low-loss way. The minimization of the loss function is accomplished through the use of gradient descent training algorithms. Example 1 - Auto-Associative How to (using TGS Salt Columns) The first "how-to" example comes from a competition held on the kaggle.com website. WHAT IS THE OVERARCHING GOAL? The competition, as well as details about the data set used for training, can be found here: https://www.kaggle.com/c/tgs-salt-identification-challenge . The purpose of the competition is to see who best can create an AI system that classifies seismic data. There are human experts who can do this classification from the original data, and indeed have done so to create the training and testing data. WHY USE AUTO-ASSOCIATIVE NETWORKS? In many case...
Read more
Image
Reading the Robot Mind: Deep Convolution Networks Decoded Even as "Explainable AI" is all the rage, coding/transformation fidelity is a critical success factor. Whether you are using frequency bands from Fourier transforms, statistical features of Wavelet decomposition, or various filters in Convolution Networks - researchers must be able to perform the reverse coding/transformation to see if they have retained sufficient information for classification. Without this, they are only guessing via network architecture trial and error. These tools are sorely lacking in Keras on Tensorflow in Python, so I wrote my own. I would like to see these more generalized and made into public libraries. Question: Who can point me to current work in this area, or can give advice on next steps in my effort? EXAMPLE: In a deep convolution network trained to recognize Bengali handwriting, this diagram examines the 64 filters after six layers of 2 dimensional convolution. It performs the reverse t...
Read more