Friday, February 21, 2025

2/21/25: BH presentation on "Unveiling Bias: Analyzing Federal Sentencing Guidelines with Topological Data Analysis, Explainable AI, and RAG Integration"

 Artificial Intelligence Study Group

 Welcome! We meet from 4:00-4:45 p.m. Central Time. Anyone can join. Feel free to attend any or all sessions, or ask to be removed from the invite list as we have no wish to send unneeded emails of which we all certainly get too many. 
Contacts: jdberleant@ualr.edu and mgmilanova@ualr.edu
Agenda & Minutes  (151st meeting, Feb. 21, 2025)

Table of Contents
* Agenda and minutes
* Appendix 1: Details
* Appendix 2: Transcript (when available)

Agenda and minutes
  • Announcements, updates, questions, presentations, etc. as time allows
    • Wednesday Feb. 26, presentation on AI by Windstream. Pizza 11:30 a.m., presentation 12:15-1:30 in EIT auditorium and perhaps online. See details in the appendix, below.
    • Feb. 28: VK will report on the AI content of a healthcare data analytics conference attended in FL. Very informal.
    • Fri. March 7: CM will informally present. His "prospective [PhD] topic involves researching the perceptions and use of AI in academic publishing."
  • Feb. 21: BH informally presents proposed PhD project on "Unveiling Bias: Analyzing Federal Sentencing Guidelines with Topological Data Analysis, Explainable AI, and RAG Integration"
  • Recall the masters project that some students are doing and need our suggestions about:
    1. Suppose a generative AI like ChatGPT or Claude.ai was used to write a book or content-focused website about a simply stated task, like "how to scramble an egg," "how to plant and care for a persimmon tree," "how to check and change the oil in your car," or any other question like that. Interact with an AI to collaboratively write a book or an informationally near-equivalent website about it!
      • LG: Thinking of changing to "How to plan for retirement." (2/14/25)
        • Looking at CrewAI multi-agent tool, http://crewai.com, but hard to customize, now looking at LangChain platform which federates different AIs. They call it an "orchestration" tool.
        • MM has students who are leveraging agents and LG could consult with them
      • ET: Growing vegetables from seeds. (2/21/25)
The meeting ended here.
  • News: new freshman level AI course! See details in the appendix below.
  • We are up to 19:19 in the Chapter 6 video, https://www.youtube.com/watch?v=eMlx5fFNoYc and can start there.
  • Schedule back burner "when possible" items:
    • If anyone else has a project they would like to help supervise, let me know.
    • (2/14/25) An ad hoc group is forming on campus for people to discuss AI and teaching of diverse subjects by ES. It would be interesting to hear from someone in that group at some point to see what people are thinking and doing regarding AIs and their teaching activities.
    • The campus has assigned a group to participate in the AAC&U AI Institute's activity "AI Pedagogy in the Curriculum." IU is on it and may be able to provide updates now and then.
  • Here is the latest on future readings and viewings

Appendix 1: Details on (i) Windstream presentation, and (ii) new AI course

Autocorrected to graduation
Conversation opened. 1 read message.
(i) Windstream presentation
From Dr. Pierce:
Windstream AI Presentation
On Wednesday, February 26, 2025, representatives from Windstream will be at the EIT Auditorium to talk about how their company is using Artificial Intelligence.  This event is open to all students, faculty and staff who would like to attend.  
   When: Wednesday, February 26 in the EIT Auditorium
·       Pizza and Soda available from 11:30 am to 12:15 pm
·       WindStream Presentation from 12:15 pm to 1:30 pm
Agenda
AI at Windstream
    Team Structure and Collaborations
        Overview of Our AI Team and Roles
        Partnerships with Other IT Teams
        Collaboration with Business Units for Strategic Alignment
Operationalizing GenAI
    Overview and Implementation
        Key Strategies and Operational Framework
        Integration with Business Processes and Goals
    Projects and Innovations
        Key Projects in Progress
        Strategic Vision and Expected Outcomes
High-Level Architecture and Tools
    Technological Framework
        Core Technologies and Platforms
        Innovative Tools and Techniques
Impact and ROI of AI
    Business and Economic Impacts
        Measuring Return on Investment
        Case Studies Illustrating Value Addition
Please share this announcement with your colleagues and students (both undergraduates and graduates).   This is currently an in person event but we will attempt to record the session so that those who cannot attend in person can benefit from the session.
   Note:  Zoom Link is below (just no promises on how well a remote session would go).  
Topic: Windstream AI Workshop
Time: Feb 26, 2025 12:00 PM Central Time (US and Canada)
Join Zoom Meeting
https://ualr-edu.zoom.us/j/86789504204
Meeting ID: 867 8950 4204
 
(ii) New course
 Department of Computer Science planning to offer:

    •  CPSC 1380: Artificial Intelligence Foundations

       Course Description

       Credit Hour(s): 3

      Description: This course introduces key principles and practical applications of Artificial Intelligence. Students will examine central AI challenges and review real-world implementations, while exploring historical milestones and philosophical considerations that shed light on the nature of intelligent behavior. Additionally, the course investigates the diverse types of agents and provides an overview of the societal impact of AI applications.

       Prerequisites: None

       Course Learning Objectives

      Upon successful completion of this course, students will be able to:

      ·         Describe the Turing test and the “Chinese Room” thought experiment.

      ·         Differentiate between optimal reasoning/behavior and human-like reasoning/behavior.

      ·         Differentiate the terms: AI, machine learning, and deep learning.

      ·         Enumerate the characteristics of a specific problem related to Artificial Intelligence.

      Learning Activities

      ·         Overview of AI Challenges and Applications - Introduces central AI problems and highlights examples of successful, recent AI applications.

      ·         Historical and Philosophical Considerations in AI – Discusses historical milestones in AI and the philosophical issues that underpin our understanding of artificial intelligence.

      ·         Exploring Intelligent Behavior

      o   The Turing Test and Its Limitations

      o   Multimodal Input and Output in AI

      o   Simulation of Intelligent Behavior

      o   Rational Versus Non-Rational Reasoning

      ·         Understanding Problem Characteristics in AI

      o   Observability: Fully Versus Partially Observable Environments

      o   Agent Dynamics: Single versus Multi-Agent Systems

      o   System Dynamics: Deterministic versus Stochastic Processes

      o   Temporal Aspects: Static versus Dynamic Problems

      o   Data Structures: Discrete versus Continuous Domains

      ·         Defining Intelligent Agents - Explores definitions and examples of agents (e.g., reactive vs. deliberative).

      ·         The Nature of Agents

      o   Degrees of Autonomy: Autonomous, Semi-Autonomous, and Mixed-Initiative Agents

      o   Decision-Making Paradigms: Reflexive, Goal-Based, and Utility-Based Approaches

      o   Decision Making Under Uncertainty and Incomplete Information

      o   Perception and Environmental Interactions

      o   Learning-Based Agents

      o   Embodied Agents: Sensors, Dynamics, and Effectors

      ·         AI Applications, Growth, and Societal Impact - Provides an overview of AI applications and discusses their economic, societal, and ethical implications.

      ·         Practical Analysis: Identifying Problem Characteristics - Engages students in exercises to practice identifying key characteristics in example environments.

      Tentative Course Schedule

      Subject to change at the discretion of instructor.

      Week

      Topics

      Learning Activities

      1

      Course Introduction & Overview of AI Problems

      ·         Overview of central AI challenges

      ·         Examples of recent successful applications

      ·         Lecture introducing course objectives and structure

      ·         Reading assignment on current AI trends

      2

      Philosophical Issues and History of AI

      ·         Examination of philosophical issues in AI

      ·         Overview of AI’s historical evolution

      ·         Student presentations summarizing key course takeaways

      ·         Course review session and Q&A in preparation for the final assessment

      3

      What is Intelligent Behavior? I – The Turing Test and Beyond

      ·         The Turing test and its flaws

      ·         Introduction to related philosophical debates (e.g., Chinese Room)

      ·         Lecture with historical context

      ·         Small-group discussion on Turing test limitations

      ·         Reading assignment on classic AI thought experiments

      4

      What is Intelligent Behavior? II – Multimodal I/O & Simulation

      ·         Multimodal input and output in AI

      ·         Simulation of intelligent behavior

      ·         Demonstration of multimodal systems (videos/demos)

      ·         Lab session: Explore a simple simulation environment

      ·         Reflective writing: How does simulation approximate intelligence?

      5

      Intelligent Behavior: Rational vs. Non-Rational Reasoning

      ·         Comparison of optimal (rational) decision-making and human-like (non-rational) behavior

      ·         In-class debate on the merits of optimality vs. human-like reasoning

      ·         Case study analysis

      6

      Problem Characteristics I – Observability and Agent Interactions

      ·         Fully vs. partially observable environments

      ·         Single vs. multi-agent systems

      ·         Group workshop: Analyze example environments for observability and interaction challenges

      7

      Problem Characteristics II – Determinism, Dynamics, and Discreteness

      ·         Deterministic vs. stochastic systems

      ·         Static vs. dynamic and discrete vs. continuous problem spaces

      ·         Hands-on group exercise: Map out characteristics of a provided problem scenario

      ·         Group discussion on design implications

      8

      Defining Agents: Reactive and Deliberative

      ·         What constitutes an agent

      ·         Examples of reactive versus deliberative agents

      ·         Interactive lecture with in-class examples

      ·         Group exercise: Classify agents from provided case studies

      9

      Nature of Agents I – Autonomy and Decision-Making Models

      ·         Autonomous, semi-autonomous, and mixed-initiative agents

      ·         Reflexive, goal-based, and utility-based decision frameworks

      ·         Interactive exercise: Design a decision-making framework for a hypothetical agent

      ·         Group presentations of frameworks

      10

      Nature of Agents II – Decision Making Under Uncertainty & Perception

      ·         Handling uncertainty and incomplete information

      ·         The role of perception and environmental interactions in agent behavior

      ·         Lab: Experiment with a simple decision-making simulation

      ·         Group discussion on sensor integration challenges

      11

      Nature of Agents III – Learning and Embodiment

      ·         Overview of learning-based agents

      ·         Embodied agents: Sensors, dynamics, and effectors

      ·         Group lab: Explore embodied agent models using simulation tools

      ·         Group discussion on design trade-offs

      12

      AI Applications, Growth, and Impact

      ·         Survey of AI applications across industries

      ·         Economic, societal, ethical, and security implications

      ·         Case study analysis: Evaluate the societal impact of an AI application

      ·         Group discussion on ethical dilemmas and future trends

      13

      Deepening Understanding Through Application

      ·         Practice identifying problem characteristics in real/simulated environments

      ·         Additional examples on the nature of agents

      ·         Extended discussion on AI’s broader impacts

      ·         Interactive workshop: Analyze a complex AI scenario in small groups

      ·         Peer review of group findings

      ·         Hands-on exercises using simulation tools or provided datasets


Appendix 2: Transcript
 
 AI Discussion Group  
Fri, Feb 21, 2025

4:00 - D. B.
All right, everyone. Thanks for being here. Today, we've got a presentation from B. H. It's entitled Unveiling Bias, Analyzing sentencing guidelines with topological data analysis, explainable AI, and RAG integration. And B. is in the PhD program here, the distance ed PhD program. And what else? So just a couple of announcements. On Wednesday, there's a AI presentation by Windstream. And it's in the EIT auditorium. And you can get pizza. And everything if you show up early. And I got a bunch of details about that below in the appendix of these minutes. That's Wednesday. Then V. is on a conference in Florida. He'll be back and is willing to tell us what happened for the AI content of a healthcare data analytics conference he's attending. Since he works for CARTI, which is the cancer company, clinic company in Little Rock and several other branches in other places. And then on the 7th, C. M. will present his prospective PhD topic on the perceptions and use of AI in academic publishing. Okay, so let's, I'm gonna just turn it over to B. and he's gonna tell us about his project.

5:39 - B. H.
Okay, I'll go ahead and take the screen and shares. All right, your screen sharing has started. Okay, are we good to go? Go for it. Okay, I changed the title a little bit just for this. So, you know, the idea is I'm rethinking sentencing disparities. So, and it's at a federal level, just so everyone's aware. Taking a different and approach the analysis. Why is it not changing? Oh, there we go. It's slow. Hold on. Okay. Okay. So the background is I'm looking, you know, we have a lot of lawyers in the family and, you know, we have some very interesting conversations around the dinner table. And I was looking for some ideas for my PhD proposal. And as we were chatting, I have a son-in-law and a daughter who are both doing some federal clerkships and this kind of came up in conversation. So I'm exploring this. So the idea here is that we're going to look at federal sentencing guidelines and kind of challenge some of the research that's been done previously. Most of it's been done in more of a linear or statistical level and very much on two or three or four pieces of information. You know, usually it's, are they, what race are they? What's their economic background? What's their educational background? Maybe you get some into the religion. What I want to do is look at all of that. And as we go through this presentation, you'll get an idea of how much data we're really talking about here. So the challenge is looking at all of that data for you know, tens of thousands, it's closer to about 75,000 sentences that occur every year, and trying to find some of the nuances that occur within that data to group the data a little bit differently. So that's kind of the challenge. So I'm going to look at it using other tools. I'm going to use TDA, which hopefully will help me detect patterns or clusters of data different than a more linear approach will do. Obviously, if I'm taking a different approach to this, I have to do explainable AI so that we can, so that I can show why that's happening. And also, because when we look at from a law perspective, they don't understand a lot of this technology. So it's gonna have to be explained in a way that can also, you know, we can explain it to lawyer and kind of remove the technical from the equation when explaining it to them. The other one is when I start looking at the guidelines and as well as trial transcripts, they're gonna be free text in a lot of cases or somewhat unstructured. So I am gonna have to add some rag to this for when I go out, look at specific trial transcripts and when I compare the year to year sentencing guidelines. Okay, so what is TDA? It's a topological data analysis. It's basically just another way of looking at and shaping the data into different structures. I'm actually leaning towards probably TDA mapper which will help me kind of group statistically larger amounts of data based on larger volumes of data. Variables, you know, using something like Kepler mapper, Guiado, or something along those lines. Not quite sure which ones I'm going to use yet, but I'll probably use several of them. And then part of it is I have 2000 plus attributes. Actually, I'll show you a screen later. It's, I think it's about 18,000 attributes that are collected on a yearly basis. And it varies year to year. Some of that based on public policy, some of that's based on guidelines. Okay, so kind of explain the explainable AI, just giving us some trust in the models and making sure we can explain it to the legal people what we're doing. Again, a couple of libraries there that I'm looking at are SHAP and LIME and others, just making sure everyone understands that the world is my oyster right now. And I could look at a few other things and alternatives as we move forward. And, you know, following Moore's law, new stuff can come along all the time. So I will definitely be looking at that. And then the last thing is reg. So reg is, I'm going to take some, you know, I'm going to have my models that I'm building, there's going to be some pre-trained probably text search models, maybe I'll use, not chat GPT, but something else that I can look at that data, but I'm gonna have to enhance that data with my own library of documents to include the sensing data document. And I have a link for that later. You'll see that. And also into some of the trial transcripts. It's gonna provide some contextual, analysis and linkage so that I can drill down into some of the court cases to validate some of the data that was put into the sentencing data structures that are in SAS and SPSS, which was a goal in itself, just trying to pull that data out and do some data loading there. So my methodology is to compile and pre-process all these attributes, it's over 2,000. Like I said, I have a screen later, I'll show you what that is. I'm gonna have to go through, validate some of the data, normalize it, and then do some data wrangling to make sure that we have valid data throughout, like I said, 25 plus years of data and thousands of attributes. The analysis is going to be doing TDA, explaining it, and then adding the context with the rag on top of it. And then, obviously, there's going to have to be some visual flow charting, some visualizations that come out of this that go into the final dissertation that shows things because, you know, a picture is worth a thousand or a million words. And, you know, I don't think I want to be writing a million words. Try to keep the dissertation small for those that are on that committee. Okay, my approach here, and this is something that I learned from the two previous people that I talked to, Dr. P. and Dr. W., was to look at it and take a smaller set of data, maybe take one or two crimes, take one or two years of data, run through the approach, see if it works before I then do it at scale. And apply it to all 25 years of data and all the attributes and all of the crimes. So that seems like a good approach and it'll help me come up with the models and the approach that I wanna do when I handle the larger data set. Okay, so the goal here is to, let's find out what the true patterns are. Are there biases? Are the biases that we see the linear data, do they really exist in the, when you look at all of the data set together and not just nitpick one or two variables that you want to nitpick on. The implications here, and this is where Dr. Cook is going to have to help me out and make sure I get it right and stay on the straight and narrow, is coming up with a good assessment that can be communicated from a public policy standpoint, making sure that if there is some findings here, that the legal community can understand those findings and either use it to make some changes or use it in court and say, hey, this analysis was done and there actually is some biases. And it will help, hopefully, one way or another. I don't know the results of it. I haven't done the research. So I can't really say what recommendations are going to come out of it yet. But those are some of the recommendations that I could foresee happening out of this. And then the other one is on the technical side, which we're all here for, is, hey, let's figure out how to use TDA, explainable, and RAG all together and do it on legal research and maybe on things outside of legal research. The other thing is, and I think I bring this up later, but right now at a federal level, they are not doing anything with AI or ML. It's not allowed. They have had court cases and court decisions that basically at a federal level, it's not allowed. So several states are doing it. There's about that are doing something like this that have AI to help them with their sensing, but there's some flaws in it. The company that has developed it won't release what the models are or how they've come up with that decision. So there's no explainability. It's a black box and some states are using it, but they're getting some pushback and some approval based on that. Okay, so I have a few links here. We have a federal cookbook. We have some of the data sets and I'll show you that. And then we have the main sentencing data, sentencing guideline from the Sentencing Commission with all of their resources that are out here. And when we're done with the presentation, I'll switch over there and show you what that looks like, Okay. So remember I was telling you earlier, I added in the top pages here. So this is the control file. This is just the header part of the control file. That's a SAS control file. The record size is 58,322. So it's 58,000 bytes wide or a little more than that. Actually it's 59,32. I put a wrong number in there. I got to correct that, but there are 18,000 in the 2023 control set. So yeah, a little bit more than 2000. A lot of them are repeatable. For example, here, some of the things that are on here is there are things that are enhancements on the sentencing. They could be anything was, you know, if you committed murder, was it done with a weapon? Was it done with a gun? Was there multiple guns there? Was there another crime being committed at the same time? Was that on school grounds? You know, there's all these different enhancements and all these different variables that go into the sentencing guidelines. On the other side of things, there are detractors that help reduce the sentence. Things like, does the person have a high school degree? Do they have a college degree? How many years of college do they have? There's other factors that help reduce it as well. So all of those are going to be considered and shown when we do the analysis on the ultimate effect of what the sentencing guidelines are and whether or not the sentences that have occurred actually fall within the guidelines, regardless of any factor. Okay. So what's the significance of the research. The contribution of the field is going to be looking at and challenging some of the research that's been done now on the disparities in sentencing. As I keep mentioning, a lot of it is done, you know, pick an attribute or two, a variable, you know, and then they'll run that data regardless of the other, as you saw, 18, attributes that went into the calculation of why that sensing occurred. And then hopefully provide a template for analyzing complex legal data sets and potentially other data sets of this type. Is typically applied to pictures, you know, images, but it can also be applied to text. And I'm going to prove that in this research. Again, broader impacts, fairness and equity in the justice system, and possibly encouraging at a federal level, the adoption of AI in other things at a federal level, especially in federal courts, where right now it's kind of not allowed to have. As I mentioned earlier, this is Greenfield at the federal level in talking to several people. I know that they would like to explore it, but until something is shown, proven, and understood, it's not gonna be accepted. Okay, so what could be future work coming out of this? Explore other TDA techniques on other datasets and showing how explainability can be used with TDA. Potential applications, like I said, applying it to the state level. State level is different because every state has sentencing guidelines. So putting this on a much larger scale at a state level where most crimes are at a state level, we're talking a much larger data set, and we're talking 50 states, the District of Columbia, all of the, in all of the other territories throughout the United States where this would be, I don't have the hardware and probably the ability to do it at that level. But if I can come up with the models, maybe somebody can tweak each one for each state. And then hopefully it'll help identify biases here and or say there is no biases and maybe it can be applied to other things within society. You know, conclusion, you know, what I'm trying to do here is combine TDA, explainable AI and RAG, show an alternative way to do complex analytics. And you know, who knows, maybe it'll alter the way that we perceive disparities. Final thought here. What did I say here? Using advanced methods to achieve justice for those convicted. The not only just those that are convicted, making sure that they get fair sentencing, but to make sure that the victims of the crime also see that if somebody is found guilty, that they know that the sentence is equal to the crime across the board. And they don't feel that they've received misjustice because somebody gets off when maybe they shouldn't have. And then on society, if society can trust what the courts are coming out with there, it's all the better for society. And again, my acknowledgments. Thank you, Dr. B., and thank everyone else on the team for listening to me. Now, I'll do questions in one minute if I can just...

22:31 - Unidentified Speaker
Share thing.

22:33 - B. H.
Oh, I stopped sharing. Hold on. Let me share one more set of information. It's here. Okay. Okay. As I mentioned earlier, I would go to those websites and show you.

22:59 - Unidentified Speaker
So you have the United States.

23:04 - B. H.
Can everyone see my screen? Yeah. Okay. So you can see the, sentencing commission data here. And it's hiding my top here.

23:04 - B. H.
Can everyone see my screen? Yeah. Okay. So you can see the, sentencing commission data here. And it's hiding my top here. Hold on. Do I get to it? Okay. Still hiding.

23:27 - E. G.
I can't see my tabs. I can't see my tabs.

23:33 - B. H.
Hold on.

23:41 - B. H.
Oh, there we go. Okay. So there's the sentencing commission. So it's ussc.gov. And come on. So this is what they use now. So they come up with zones of how long a prison case should be. They put the category out there, they put you in a zone. And as you can see right now, it's pretty, there's a long and what the judges do is they basically take out their old slide rule, compare what they think or what they've come up with or noted throughout the trial, and then they come up with a sentencing guideline. Now, in reality, the way this works is the judge tells the jury, here's where you should be. The jury comes back with a recommendation, but the judge gives out the sentence. The final sentence. And then the last thing I wanted to show was, this is the code book. So there is a, it's effectively a data dictionary for all of the data. And as you can see, you know, we start off here, like ACC cap gives a definition, gives you the permitted variables. So this is part of where that reg is gonna come in. Cause I'm gonna have to go through each year read this data dictionary and make sure all the data is valid as I go through the data. So that's gonna be the first part of the data cleansing. And with that, I will open it up for questions and stop sharing.

25:25 - D. B.
Any questions for B.? I have a question.

25:29 - D. D.
So you're looking for hidden features besides the features that they're listing as codes? I'm not looking for features.

25:39 - B. H.
So right now, a lot of analysis on sentencing is done. People, the research that's being done, they nitpick one or two or three variables. What I'm trying to do is take all the variables across the entire spectrum and pull it together and then cluster in a different way because if you read the research, one research that's in my literature review, it says that women are always sentenced for similar crimes, lesser than men, cross the board. Well, I don't know if that's true or not. This one piece of research said it is, but I don't know if they said, is it only for certain crimes Is it only, you know, what were the enhancements on it? You know, men are more likely to use a gun or a weapon in it, which may be why they have a higher sentence amount. Let's compare, let's bring all the variables together and cluster it and see if that's in fact a true statement.

26:53 - D. D.
Say variables, I hear features. I might be confused. Features are variables. Yeah, that's what I'm thinking too. It sounds like that you're doing a feature extraction of the data, and then you're going to try to figure out what features actually contribute to the sentencing. Yeah. And I would look to see if anybody from a machine learning point of view. Not that your way won't be novel and unique, because it sounds like it will be. But you might want to look and see what other people have done just from a machine learning perspective. And you may have in your literature review. I don't know. I do. And there's some at a state level.

27:43 - B. H.
There is absolutely none at a federal level. And at state level, the research that's out there is a proprietary. So they're not releasing the information yet because they're selling it as a product.

27:56 - V. W.
I missed the first part of your talk and I regret that because this is really an interesting area. Catching the last part, it made me wonder if you're including the feature data for the defendants, for example, their ethnicity, how well they're doing from an income point of view, what their net worth sort of thing. I'd be interested if you could detect any biases in sentencing according to that criteria, which would make it really newsworthy and in the interest of justice.

28:26 - B. H.
So the answer to that is a lot of that research has been done. And that's that's the point is they they a lot of that research has been done and they will nitpick. They will they will do is, you know, what is the race of the person? What is the sex of the person? What is the income? What is the what is their educational background, and they might combine one or two of them, but it's a very linear or very statistical approach to it. Whereas how about if you take other things that go along with it, not only those things together, which rarely are combined and analyzed, but also how about the enhancements that go along with that? How about, you know, some, you know, some of it is family dynamics. Do they have children? What are the ages of the children? Some of that does go into the sentencing as well, but none of that is ever talked about. None of that, you know, that is rarely researched. And so what this is going to do is do a comprehensive research on all 18,000 variables at once. What about their occupation?

29:30 - E. G.
Just real quick note here, E.

29:32 - V. W.
What about their occupation? You know, they have a stated occupation and then they have their actual occupation, maybe career criminal, but their stated occupation may have within a certain level of academic training and how does that figure in the mix? Like I noticed you had a career criminal or like career politician or what their category of occupation is.

29:54 - Multiple Speakers
Aren't career criminals and career politicians the same?

29:57 - B. H.
Well, your study would show definitively once and for all.

30:01 - V. W.
But the answer is, I don't know.

30:04 - B. H.
I have not done enough of the reading of the guideline to know if that is one of the attribute, one of the elements that I'm, the features that I'm looking at. So yeah, I'm going to have to, I'm hoping with TDA, it helps me with some of that analysis, but yeah, I'm going to have to do that reading and understand exactly what. It's a real discovery opportunity, which is what's really exciting about this new generation of AI tools.

30:32 - V. W.
Right. Someone else had a question?

30:34 - E. G.
One of the easy ones that I had done some reading of on way back when was public defender versus private defender. The sentencing guidelines tend to be drastically different. Those types of things lean into what V. was alluding to that I wanted to get into was what are the socioeconomic indicators that dictate what sentencing, if they have the money to push it out, are they able to negotiate a more lenient sentence? Do dream teams really help the defendant?

31:18 - B. H.
Yeah, I hadn't thought about that, but I know socioeconomic is part of it that goes into the data, but I hadn't thought about the public defender versus private defender. And so I wrote a note, I definitely need to find some resources on that as well. That's actually a good one that I need to add to my literature review. So I appreciate that.

31:43 - D. B.
I have a question about your data source or data preparation, I guess you could say. Do you have access to a database of all these thousands and thousands of features, or do you have to curate the data yourself, or what? I do.

31:59 - B. H.
It is a pain in the butt, because so SAS only works with 32,000 bytes wide. So I'm actually running SAS. There's an academic version of SAS that I can get access to. It's online. So I have to actually break the data set down. As you saw, that one was 58,000 bytes wide or 59,000 bytes wide. I have to take the first 32,000, chunk it into one file, take the second half of it, chunk it into another file, and then I have to, and then I'm going to use some Python or something to pull them back together.

32:37 - Multiple Speakers
So you considered using Unix tools, text processing tools to help you out with that task?

32:43 - B. H.
So Haven actually does it for me, and there's Haven in both R and Python.

32:49 - D. B.
So do you, I mean, how do you, who takes the data, you know, you're talking about trial transcripts and so on, who takes that text data and creates database entry out of it?

33:02 - B. H.
The the court clerks do. Okay. That's why I said some of it, I'm going to have that's where the rag comes in. I'm going to, I'm not going to be able to obviously do 75,000 cases, court cases every year for 25 years, look at the trials, transcripts and verify everything, but I am going to have to have to do a test, grab a small subset and have it do read the trial cases and make sure that the data that was entered for that sentence is in fact accurate.

33:38 - V. W.
You can also summarize trial cases so that you don't have to deal with all the minutiae of the case, but rather just the salient features that you believe to be relevant. And LLMs with long context length like Claude's sonnet 3.5 200k are an excellent shortcut to doing a lot of cases. And I've seen people who have risen to real prestige, when they did a job nobody else wanted to do. And they went through a whole bunch of backstory or research or literature, eking out the details that nobody else bothered to eke out. And in the midst of that, there's discovery and you know, benefit to the research that's very tangible.

34:19 - B. H.
Maybe that's another benefit that I hadn't thought about that will come out of this like like I said, I wasn't planning on doing a deep dive on the trial cases, but that may, you know, if I can come up with that aspect of it to just even research a few of them for validation, maybe that model can be applied on future research.

34:42 - E. G.
There'll be treasure there. Yep. One of the things that I like to do is I'll run OLAMA locally.

34:49 - B. H.
Yeah.

34:50 - E. G.
And, um, or, or E., Dr. M. with, uh, with Rapids.ai I run a lot of stuff locally to start going through I would not want to go through each one line by line but the computer doesn't give a crap about what it's doing for the next hour so I'll let it sit there and just chug across everything.

35:12 - B. H.
Yeah so yeah that's actually so I do a lot of work with DoD and that's actually how we set up AI for the military was for classified space, we actually set off a limited access and set up OLAMA, and then we put Mistral on it.

35:35 - M. M.
Yeah, this was actually my question. Thank you. Excellent proposal, B. Excellent work. Interesting, too. Like E. mentioned, you can use RAPIDS or any the rating libraries for manipulating with the data. Is fine, but you need to know it's LLAMA integrated with RAC or any of the large language models, which one you're using, so you have to check. This TDA is topic modeling, I think.

36:16 - Multiple Speakers
Yeah. Yeah.

36:18 - M. M.
So also you need to figure out what kind of embeddings you're using or whatever, because topic modeling. Yeah. This is truly a problem that lends itself to concurrency as well.

36:29 - V. W.
So while you might prototype your method locally in an atmosphere that you have very tight control over it, when you get ready to spin up the big job of 25 years or 75,000 cases, well, 75,000 cases sounds a lot if you're running on one GPU, but if you're harnessing a stable of GPUs, which is possible through Colab and other facilities, you can make quick work of that.

36:55 - Multiple Speakers
Yeah, I'm trying to figure out where I'm going to put the infrastructure for this and try not to pay too much out of pocket.

37:03 - B. H.
I do a lot of cloud build outs for government, so I know how to do it there, but trying to find other places. Do you think Colab has enough that I I could be able to do it there.

37:16 - V. W.
I sure got a benefit from it because I went from the free tier to the $10 a month tier. And then when my dissertation was in the heat of it's needing a lot of processing power, I paid the 60 bucks a month. And all of a sudden I have a lot more power at my disposal. Plus people are building out so rapidly now they need test cases to test the way their ensembles working. And so you can always say, Hey, I have this test case I need to run on your new, uh, multiprocessor array, could I just camp here for a few weeks?

37:49 - Y. P.
B., you can get some credits from Google. I can share some guidelines on that.

37:57 - B. H.
OK. Appreciate it. All right.

38:00 - D. B.
Any other questions for B.? I have a question around explainability.

38:06 - Y. P.
OK. Because one of the key aspects of this is going to be the data explainability and model explainability, especially if we are using open source machine learning models or we are thinking of using generative AI. You did mention, B., about explainability and importance of it, but are you thinking of using cross models or any other tools or are you just going to check a few outputs based on the rules engine, how are you going to make sure that whatever is coming out, either with manual oversight or traditional technology oversight or some oversight can really explain all the components of processing the data and the models?

38:57 - B. H.
Yeah, so I think the only place I'm going to use a large language model is with the RAG enhancements where I'm doing the text I think I'm actually going to be building my own model with Python or I'll figure out what other libraries and tools I need, but I think it's going to be more of a purpose built model to be able to, between using Haven and then adding in the TDA mappers and then adding in like Limer for the explainability. I think I'm going to have to build a lot of it from scratch.

39:35 - V. W.
Sometimes it's fun when you're doing that to run a competing approach. So for example, you could take the approach you're currently taking, and then you could say, well, I want to build a robotic judge that will fulfill some justice department need. And then you could use reinforcement learning. And we know that the reinforcement learning people and the deep layer neural network people don't even eat lunch with each other. So that makes a great counterpoint to your approach because you can kind of begin to compete in the GAN tradition of generative adversarial networks. You can compete not just two instances, but two different approaches to see if you get a windfall of any new information.

40:20 - D. D.
One more quick question.

40:22 - B. H.
So the RAG is going to be like a advisor or a Two things, so it's gonna be a validator of the data as I look at the trial data to validate the data that I have that is collected by the clerks within the data sets that they post on the sentencing guideline website. One, I wanna just do, pick and choose and validate a few per year just to make sure that data is right. The other thing is the sentencing guideline document itself is textual. And so it's got the data dictionary in there, but I'm gonna have to read that to pull out a dictionary to be able to validate and add knowledge or, you know, define what each of the elements are that I'm

41:32 - D. D.
looking at. All right.

41:34 - D. B.
Well, thank you, B. And I'll get with you later about the approval I need to provide you. By the way, I will give it, of course. OK, thank you. And let's go on to the next item. Couple of masters students who are using AIs to try to write books. And I think one person is here. E. is here. And here's your entry. So E., can you give us an update for a moment and ask us any questions you have?

42:18 - E. T.
So this week, last week, I said that I was gonna try to find better prompts for being able to actually reach word count. So what I found this week, I'm actually pretty excited about it, I found a course on cognitive class for prompt engineering and I am not done with the course yet but it's giving very valuable information about how to use the prompt to get the word count goal. So I tried some of those prompts, and I actually got a lot better results than previous time. So as your suggestions, I tried those as well, tried to get the word count to a lesser amount, like 5,000 words. It still didn't reach the 5,000 words or above, but it got really close, like the highest word count I got was 4,100, something like that. So I'm excited that it will help me get my reach at my goal using those prompts. And I'm also working on my proposal for the graduate project. I also used Gemini this week, yes. So I got some prompts from ChatGPT and used them on Gemini. And I actually got, I really liked the writing on Gemini compared to ChatGPT. So I'm assuming I will get better results from Gemini. I don't use the paid version of Gemini, but still I like the tone and how it makes it engaging and how the titles were a lot more cheerful compared to ChatGPT. Okay. Any comments, anyone?

44:27 - D. B.
So you're still doing growing vegetables from seeds, is that right? Yes. Okay.

44:33 - V. W.
Yeah, I put a quick comment in the chat, which I will complete later, but basically one trick I've used in prompting lately, as I say, Now that I've given you what I feel is the ultimate, most comprehensive possible prompt, tell me if I've missed anything. And if I have missed anything, please rewrite the prompt with that additional content and use that prompt to run this query. And I found I usually save two or three steps because invariably I'll think I didn't miss anything and it'll find some huge glaring hole in my reasoning. And then when I run it, when it runs it with that, I get a work product that's much more and much fewer shots through the loop of the design model. So, there's that.

45:18 - D. B.
So, you're saying that you can actually ask the AI to improve your prompt? Oh, yeah.

45:24 - Multiple Speakers
I do it all the time now, and I can't believe I lived without it. It's a significant development in my prompting style.

45:33 - D. D.
Haitian takes the prompt engineering class.

45:35 - V. W.
It's a very meta thing to do, and AIs tend to like it when you go meta or outside your current scope of reasoning.

45:43 - D. B.
I mean, how's the AI supposed to know what an improvement to your prompt is? I mean, maybe, how does it know how to improve your prompts?

45:52 - Multiple Speakers
Because it has that vector space of meanings that it can then reevaluate according to your request. It's the best thing since sliced bread.

45:59 - V. W.
It's amazing. Yeah.

46:00 - D. B.
I mean, when you make a prompt, how does it know that you didn't make the prompt that you wanted to make? You know, it's like second guessing when you're designing an IT system. They say they want some stuff, but you don't believe them.

46:14 - Multiple Speakers
Well, how does the AI know not to believe you?

46:16 - V. W.
It's the act of surfing the emergent properties that we've all been oohing and aahing about for the past year. You should try it.

46:23 - D. D.
Dr. B., I can just imagine this being an endless loop.

46:26 - E. G.
Please update the prompt with the prompt that I really want. Well, the prompt that I really want wasn't set, but it's going to keep updating it. It'll just keep going to and updating the prompt and never actually provide an answer.

46:39 - D. B.
Yeah, I think there would be a master's degree project or thesis in sort of asking AIs to iteratively keep going until they reach a fixed point. So you're saying, well, ask it to improve the prompt. Well, ask it to improve it again and again and again. And what happens in the end?

46:56 - V. W.
I think there's an 80-20 rule in play that you get the most bang for the first buck. OK, but where's the fixed point?

47:04 - Multiple Speakers
point. I mean, if you kept going forever, or do it, you know, 20 times or 100 times, what, where would you get them?

47:15 - D. D.
You know, the prompts test a different model, see, see how they work. E., can you share a link to your to that prompt engineering class you were talking about?

47:28 - E. T.
Sure. Sure. I'm actually thinking I'm Sorry.

47:32 - D. D.
Maybe you could post it in the chat so we could visit it and check it out.

47:39 - E. T.
Sure. Sure. I'll try to get it. So I'm actually taking a data science technologies course as well this semester. So Dr. P. always sharing some courses from cognitive class AI. So those are self-paced courses. And I found that course on cognitive class AI. Let me find a link.

48:03 - D. D.
You have several in NVIDIA, if you want.

48:07 - M. M.
Yeah.

48:08 - D. B.
Yeah, prompt engineering is a, you know, it's like people are getting, at least at one point, humanities majors were getting high paying jobs by being good prompt engineers.

48:24 - M. M.
Replacing a lot jobs. All right. Anything else anyone wants to talk about?

48:33 - D. B.
E., if you have it now, I can type it in now. If you want to get it to me later, I can add it later. Either way.

48:53 - E. T.
I think I put it in the chat, but I'm not sure if I put it in the correct place. Are you able to see it in the chat?

49:03 - D. B.
I don't see it in my chat, at least to everyone. Maybe you sent it to me.

49:09 - E. G.
You sent it to the AI, I bet.

49:12 - Multiple Speakers
It's probably the AI, yeah.

49:13 - E. T.
I've typed in and sent stuff to the AI. Sorry.

49:17 - D. D.
Down there where it says two, maybe it's in blue, you hit that little drop down menu and select everyone. Got it.

49:27 - E. T.
Thank you. OK, I see that link.

49:31 - D. B.
Say whatever is your message, please send to me.

49:37 - M. M.
Yeah, you gotta be enrolled to see it.

49:42 - E. G.
Yes, it asks you to enroll, but it's all free.

49:51 - Unidentified Speaker
Good.

49:52 - E. G.
OK, well.

49:53 - Multiple Speakers
OK, I stuck that stuff in the prompt.

49:57 - V. W.
A couple of tricks that might help you, including asking for an HTML, CSS, JavaScript product as a single file so you don't have to worry about putting the pieces together. But then as your website becomes more sophisticated, you can ask it to provide it in installments. First give me the body, then the CSS, then the head and all the script. And if you do it in installments like that, you can have it write much bigger programs for you without running it into any token length limitations on output, which it has. And another trick is if you want to do local file IO on your machine, because you want to read a configuration file, you want to write the results of what you simulated, you can run a little Python start a shell, run a Python server. And then instead of putting the name of your HTML file in, you put the like 192.168.0.0 colon port number of the server. And then it'll just allow you to step past all the safeguards that are in place to prevent file IO in an unsecured fashion. So by running a local server, you can increase the file system capability of your web application enormously for relatively little hassle. I mean, you literally just start it. And if you're not sure on how to implement any of these things, you can just ask the LLM, how might I use a server on my own local machine to run a HTML file that's doing file IO, and it'll give you the instructions on how to run the server. So it's real made in cool.

51:41 - D. B.
All right. I think we're at the good stopping point. Thanks for joining everyone. And we'll, I guess next time we'll sort of go back to our normal routine and watch some video segments and talk about those and go on from there. Thanks B. Well, thanks everyone. I appreciate that.

51:59 - B. H.
I've got some notes here, so I'm going to have some updating to do here, but I appreciate it. It'll be great to see what you come up with.

52:10 - V. W.
Outstanding. Good job, B.

52:13 - D. D.
Maybe you can do your proposal in the group, too.

52:19 - Multiple Speakers
That would be nice.

52:22 - D. B.
All right, guys.

52:24 - B. H.
Thanks.

52:25 - M. M.
Take care, everyone.

52:27 - Unidentified Speaker
Thanks. Bye. Bye. Thanks a lot, H.

52:34 - D. B.
All right, take care.



Displaying jan10FinishedTranscript.txt.
at

No comments:

Post a Comment

Subscribe to: Post Comments (Atom)