Hello! Life update – I’m got a full ride to grad school with a stipend and I’m taking a class called Introduction to soft computing. This class recently gave me an assignment where I had to read a couple papers on topics like strong AI, explainable AI, and transparency in AI. I had to answer a couple questions regarding these papers and my responses are below. (lol, there are prolly lots of grammatical errors… i didn’t really reread and edit. don’t judge). Please see the papers in the references section

1. What are the opposing arguments in favor and against the need for transparency in AI systems? What is your opinion and state the reasons and examples in support of your argument?

There are many reasons for and against transparency in AI. All valid reasons within context in my opinion. One main reason FOR transparency in AI is the fact that you can understand why your AI worked or did not work. With “black-box” deep learning style AI techniques, all you know is how your model is adjusting its weights. You have no idea why it is doing what it is doing. One great quote from this article is “…you do not know if the fault is in the program, in the method, or because things have changed in the environment.” There are a lot of variables when it comes to constructing an AI and if too much of the process is a black box, then you don’t know what to change if it fails. Transparency in AI fixes all of these problems.

The reasons AGAINST transparency in AI are all around what is feasible right now. Right now, we can throw a whole bunch of data into a black box and get an answer to a question that is correct sometimes. This works. Right now. It has flaws but it works. So, if it works, why fix it? Giving AI more transparency may be too much work for the type of problems we want to solve. That is the biggest problem with this idea, the fact that it doesn’t work yet. If we did have an AI that not only solved our problems but told us why and how it did. There would not be much push back to the idea.

Do we NEED transparency in AI? In my opinion, yes. I believe the next wave of improvements in the realm of AI will be transparent and capable of explaining themselves. Right now, AI is simply not strong enough to tasks that a toddler can do. The approach we are taking to improve AI is inefficient and we don’t really know why. Being able to understand why a child does something wrong can help us help the child understand why what they did was wrong. That is a simple example, but code/algorithms/AI work the same way. When you make a simple math equation and the output is wrong. Right now, we are essentially guessing and throwing things together that gets us an answer – some of the times. If we truly want to see an improvement in the industry. We will need to be able to dissect and comprehend why an AI behaves the way it does. Blind-model AI doesn’t have a bright future in my opinion.

1. Describe the term Strong AI and why do you think it may be important? Why may one need a “mental representation of their environment”?

Strong AI is AI that can perform and reason at a human level. This is the type of AI we usually see in the movie theater. This is ultimately what people think of when they hear the word AI, and it will be very important to our future. If we can create an AI that cannot only think but imagine/simulate, we could solve problems at an enormously fast rate. An AI that can conceptualize/perform at a human level would also be able to write code, create simulations, build conceptual mocks very quickly. We could essentially simulate millions of situations or chemical processes or stress analyzes test in seconds. There are also many repetitive tasks that most jobs do on a daily bases, for example, entering data on an excel sheet.

When it comes to humans having a “mental representation of our environment” is crucial to functioning in society. This simulation or imagination process is, for one, how we empathize with one another. You can not relate to someone if you are unable to imagine what life would be like in their shoes. Imagination is one of the most underrated parts of society today. Having a large imagination enables you to come up with unique ideas and solve problems in an innovative way. Aside from the imagination part of this question having a mental representation of our environment helps us understand where things are without seeing them. If my I am watching or counting cars on the highway and a small car goes behind a large car, I know there are still x amount of cars on the highway. I know this because I understand that even though I can’t see the small car, it’s still there. Lastly, having a mental representation of our environment allows us to start “playing” or changing things our environment just to simulate what would happen. This, again, is very useful in a problem-solving situation.

1. Explain and expand on Pearl’s conclusion that “Human Level AI cannot emerge solely from model-blind learning machines”

Model – Blind learning machines become “intelligent” by understanding trends in datasets. These learning machines aren’t really “learning” like humans do. They are using mathematical equations to try and give a right answer. This type of learning can be great in predictive analyses scenarios, but when it comes to achieving a human level of competence this technique is hopeless. In order to use model-blind learning machines one must first have A LOT of well-organized data. Once the machine is trained on this dataset it can only do one thing with it – what it was trained to do. It can’t manipulate or play with the data to come up with general conclusions. It’s simply not curious enough to do so. Humans are very curious, creative, and always seeking comfort. Model-blind AI’s don’t want anything, therefore cannot ever achieve human level of thought. In order to create a strong AI, we will need to think outside of the box both conceptually and mathematically.

Human level AI should not only be able to look at data and make predictions but understand how what it is analyzing will affect society. It should be curious and “want” something (other than to find the best fitted curve). There are many parts of a human thought that we must consider when developing a strong AI. I won’t get into all of them right now, but I will say, imagination is a big one. So right off the bat model-blind learning machines are ill-equipped to become strong AI’s.

Lastly, model-blind learning machines are not transparent in nature. I hit on this on the questions above, but in order to improve something you must understand it. This is my opinion and the level of understanding is up for question but in general if we want to make advancements in this type of technology, we have to understand the barebones of the system.

2.1 Explain the relevance of the Casual Hierarchy presented in Figure 1 with the development of a Strong AI system?

The casual hierarchy is a fantastic way to determine how intelligent or strong you AI is. The first tear of the hierarchy is asking purely statistical questions. Questions that you must only observe to answer. This is where machine learning models are today, they must have tons of data to make predictions. Once they have made their prediction their capabilities end. The next tear is called Intervention, this is where we actually make changes to the environment. Some AI out there can do this, it’s seen a lot in AB testing of websites or marketing material. The last tear is called counterfactuals, this is the tear we hope Strong AI gets to. Counterfactuals ask the question Why? Or What if? The AI would have to imagine a completely new world in which something changed and then observed what happened to make a conclusion.

This hierarchy also proves that today’s AI is to very large steps behind where humans are. We simple are not taking in enough variables. The last tear of the hierarchy requires imagination, or some sort of simulation environment to recreate actions and models. One of the main questions in my opinion is – are we able to create an AI that starts at counterfactuals conceptually and mathematically. Or do we need to first allow an AI to see/observe data, then interact and intervene to have it be capable of Imagining new scenarios and developing new ideas.

The last great relevancy of this Casual hierarchy is the act that the author maps out each tear with a mathematical formula. This is a great start to actually creating an AI that can become counterfactual and achieve the level of intelligence that humans have.

2.2 In your own words, summarize the seven Pillars of SCM (“Structural Casual Models”) and why you feel they may be the key to Strong AI.

The first pillar is called encoding causal assumptions, it’s all about creating a mathematical representation of the assumptions used to obtain an estimate to a challenge.

The second pillar is called Do-calculus and the control of confounding, this explains how to properly deal with unobserved causes that effect variables we are using to determine the outcome of a situation. To deal with these missing variables we can use a various set of techniques including “back-door” or do-calculous. Do-calculus “imagines” different scenarios and runs experiments in these scenarios which can determine what variables or assumptions we need to accurately predict the effect of unobserved causes.

The third pillar is called algorithmization of counterfactuals which is what it sounds like. Creating a mathematical or graphical representation of questions that involve the use of experimental or observational studies.

The fourth pillar is called mediation analysis and the assessment of direct and indirect effects. This pillar is all about determining what other causes might produce an effect, AND OR, what causes change other variables that immediately impact the effect.

The fifth pillar is called external validity and sample selection bias. This pillar is all about how to make our AI adaptable to different environments that it may not have been trained in. The solution to this problem can be solved by creating an AI that has do-calculous capabilities.

The sixth pillar is called missing data. Missing data has been a problem since the start of AI and machine learning. To deal with missing data graphical models can be created to help make predictions on the values of the missing data.

The seventh pillar is called casual discovery. This is all about using the above techniques to create a couple accurate casual models of a situation.

These pillars hit on a lot of what will make strong AI so powerful. They explain how important it is to mathematically represent assumptions and create graphical models of scenarios that exist and that are “imagined”. This, although brief, in my opinion is the key to strong AI.

3.1 Why would we as engineers and scientists be interested in reflecting on the aspect of “AI in the Legal Domain”?

AI in any domain is going to be a good and well received technology. If AI can outperform human capabilities companies will save money and most customers will be happier. AI in the legal domain is an interesting and complicated task given the need for explain ability. With that being said, most legal situations are and should be unbiased and algorithmic to a standpoint. Right now, in some legal cases there may be some discrimination or bias towards something or someone. With an AI tuned and created well the bias legal system seizes to exist. Along side those reasons, having an AI in the legal domain is good because we will know exactly what’s happening and why decisions were made the way they were. explainability in Legal AI may be more important than any other field. To convict someone of a crime or give someone freedom is a serious decision to make. If an AI makes these decisions it MUST be able to explain why it made the decision it did.

As with any field, processes can be improved and optimized to a standpoint. AI, specifically in the legal domain, can help guide workers to make better legal decisions as well as customers understand the ins and outs of a legal process. The word “Legal” comes with a lot of subdefinitions and subtopics. There is a ton of information in this field to be sift through, to understand, and to update. These are all tasks that an AI could be capable of.

AI in the legal domain could also be used to predict different legal trends around the world and use those predictions to benefit either our country as a whole or an individual. Once we make a capable strong AI it could be used to imagine and prepare our society for changes in how we operate our country and facilities.

3.2 Why may there be a need for more algorithmic transparency from a legal standpoint?

Why wouldn’t there be? I hit on this in the last question, but if you can not explain why you are convicting someone then we have a problem. When it comes to almost everything legal, there is a reason or logical process that should be explained and unbiased. If algorithms are used to in legal processes explainability is important. Now, when it comes to creating or training an AI algorithmic transparency is key to making sure that it is “thinking” the right way. We don’t want an AI that makes the right conclusions for the wrong reasons. This will ultimately lead to a bad conclusion down the road.

We MUST be careful when designing and testing these legal algorithms. The data set MUST be diverse, and the AI must also understand what it means to discriminate. Just as we humans check ourselves often (or at least we should) on if what we are doing/saying is discriminatory in anyway AI should do the same. Once we get to the creation of strong AI there should always be a simulation where the result is discriminatory, just so it knows what to stay away from.

3.3 Why may Explainability be regarded as an Intrinsic Property of Machine Learning Algorithms?

From an optimization standpoint explainability allows the creator of an algorithm to optimize and debug their algorithms. The more explainable an algorithm is the more likely it will actually be used (assume it is accurate enough). The idea of black box conclusive AI’s is dying away because they have shown discrimination in the past. AI’s can make the wrong assumptions when they solely look at data, this is why explainability is becoming more and more intrinsic. Giving people reasons for why something is happening the way that it is happening is better than just giving them some prediction. As this field grows, explainability will be just as important of a metric as accuracy.

REFERENCES

[REF 1] Pearl, J., 2019, “The Limitations of Opaque Learning Machines”, Chapter 2 in Book:
Possible Minds: 25 Ways of Looking at AI, Edited by John Brockman, Penguin Press, New York
LINK

[REF 2] Pearl, J., 2018, “Theoretical Impediments to Machine Learning with Seven Sparks
from the Casual Revolution”, January 15, 2018, arXiv:1801.04016v1 [cs.LG]
LINK

[REF 3] Waltl, B., and Vogfl, R., 2018, “Explainable Artificial Intelligence – The New Frontier
in Legal Informatics”, Stanford Law, in: Jusletter IT 22. February 2018
LINK