Negative Ontology of True Target for Machine Learning: Towards Recognition, Evaluation and Learning under Democratic Supervision

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Negative Ontology of True Target for Machine Learning: Towards Recognition, Evaluation and Learning under Democratic Supervision".

Jane: The paper was written by Yongquan Yang from Institute of Sciences for AI.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everyone. Today we’re digging into a paper that’s got a title that just grabs you: “Negative Ontology of True Target for Machine Learning: Towards Evaluation and Learning under Democratic Supervision.” Jane, I have to say, that title alone made me stop and read it twice.

Jane: It did the same for me, Tom. And I think the title is actually doing a lot of heavy lifting here. It’s telling us that the paper isn’t just tweaking an algorithm. It’s questioning a core assumption that most of machine learning is built on.

Tom: Right, and that assumption is that there’s a “true target” out there. You know, the ground truth. The correct answer that we’re trying to get our models to predict.

Jane: Exactly. And the paper says, what if that true target doesn’t objectively exist in the real world? Not that it’s hard to find, but that it’s fundamentally not there.

Tom: That’s a wild idea. So if there’s no ground truth, what are we even doing? What are we training against?

Jane: That’s where “Democratic Supervision” comes in. Instead of one authoritative answer, you have a bunch of different perspectives. You have experts, you have non-experts, you have machines. And the supervision signal comes from aggregating, negotiating, or just letting all those views coexist.

Tom: So instead of a single, perfect label, we get a crowd of imperfect ones.

Jane: Precisely. And the paper’s argument is that this isn’t a compromise or a fallback. It’s a more honest and more general way to think about learning, especially for tasks where the “right answer” is genuinely ambiguous.

Tom: I love that. It feels like it’s turning a limitation into a feature. Lu, you’ve been quiet. What do you make of this framing?

Lu: I think it’s a really bold philosophical move. It reminds me of debates in physics about whether a system has a definite state before you measure it. Here, they’re saying the target doesn’t have a definite state. And that forces us to rethink what evaluation even means.

Tom: So it’s not just a technical fix, it’s a whole new lens.

Lu: Absolutely. And I’m curious to see how they make that lens practical. Because saying “there’s no truth” is one thing, but you still have to train a neural network.

Jane: And that’s the bridge they have to build. The title sets up the philosophy, but the real test is in the machinery. I can’t wait to see how they pull that off.

Tom: Same here. Let’s keep going and see what they actually propose.

Summary: Tom: So we’re back, still on “Negative Ontology of True Target for Machine Learning.” And Jane, we just talked about the big idea. But the paper actually gets pretty concrete. It introduces this concept called MIATTs.

Jane: Multiple Inaccurate True Targets. And I love that acronym because it’s exactly what it sounds like. Instead of one true target, you have a set of them, and each one is incomplete or partially wrong.

Tom: Right. So imagine you’re trying to segment a bicycle lane in a photo. One annotator might mark the whole lane, another might only mark the part with clear paint, and a third might include the gutter. None of them is perfectly right, but together they cover the concept.

Jane: And that’s the key insight. No single target is complete, but their union collectively represents the underlying reality. The paper defines this formally with something called semantic facts.

Tom: Semantic facts. So each inaccurate target captures a subset of the facts about the true target. And the union of all those subsets gives you the full picture, or at least a good chunk of it.

Jane: Exactly. And this is what they call the instance-level realization of Democratic Supervision. It’s the concrete data structure that makes the philosophy work.

Tom: So we’ve got the philosophy and we’ve got the data structure. But how do you actually evaluate a model against a set of targets instead of a single one?

Jane: That’s where the Logical Assessment Formula comes in, or LAF. Instead of comparing your prediction to one ground truth, you compare it logically to the set of MIATTs.

Tom: So it’s not just pixel-by-pixel accuracy. It’s about logical consistency with the facts that the MIATTs collectively represent.

Jane: Right. And then there’s the learning side, which they call Undefinable True Target Learning, or UTTL. That’s how you train the model when you don’t have a single target to optimize against.

Meng: Jane, can I jump in here? I’m trying to figure out how this works in practice. If I’m training a segmentation model, I usually have a loss function that compares my output to a ground truth mask. What does that loss look like here?

Jane: Great question, Meng. The paper proposes a two-step process. First, you construct a loss based on the MIATTs. Then, you optimize the model, but you select the best model based on the logical performance metric from LAF.

Meng: So you separate the training from the selection. You use the loss to explore, and the logical metric to decide which exploration was best.

Jane: Exactly. It’s a way to make the non-differentiable logical evaluation usable in training.

Tom: And that’s a really clever workaround. It keeps the machinery running even when the target is fuzzy.

Meng: It does. And I’m curious to see if it actually holds up in a real experiment. Let’s see what they did.

Improvements: Tom: We’re back on “Negative Ontology of True Target for Machine Learning,” and we’ve covered the philosophy and the framework. Now let’s talk about what this actually improves. Jane, what’s the big win here?

Jane: The big win is that it lets non-experts participate in building and evaluating models. In the traditional setup, you need a domain expert to create the ground truth. That’s expensive, slow, and sometimes impossible.

Tom: And the paper shows this with a real example. They used this framework for bicycle lane segmentation, and they treated themselves as complete non-experts.

Jane: Right. They didn’t know what a bicycle lane was in the images. They just generated multiple inaccurate targets using tools like SAM and PixLab. And then they trained a U-Net model using their MIATTs framework.

Meng: And how did it perform? I’m always skeptical when you remove the ground truth from the loop.

Jane: That’s the surprising part. They defined a stopping criterion: Logical IoU greater than zero point nine nine nine and logical errors less than one hundred. And their model hit that at epoch six hundred twenty.

Tom: So the model converged to a point where it was logically consistent with the MIATTs. And visually, the predictions aligned really well with the logical true target derived from the set.

Meng: So it’s not just noise. The model actually learned something coherent.

Jane: Exactly. And the qualitative results show that the model’s prediction integrates the complementary information from the MIATTs while filtering out the individual biases and errors.

Lu: That’s a really nice result. It shows that the democratic process isn’t just about averaging noise. It’s about finding the structure that multiple imperfect views agree on.

Tom: So what does this improve in the real world? What’s the practical impact?

Jane: The biggest one is education and professional development. Because the framework gives you feedback. As a non-expert, you see the model’s prediction compared to the MIATTs. You see where it agrees and where it doesn’t. And that teaches you something about the underlying concept.

Tom: So the act of building the model becomes a learning process for the human.

Jane: Exactly. It’s not just a tool that gives you an answer. It’s a system that helps you understand the problem better. And that’s a huge shift from how we usually think about eye.

Meng: So it’s like the model is a tutor, and the MIATTs are the practice problems.

Jane: That’s a great way to put it, Meng. And it means that people without formal training can develop expertise just by interacting with the system.

Tom: That’s a powerful idea. It could democratize expertise itself.

Jane: And that’s the real improvement here. It’s not just a better algorithm. It’s a more inclusive way to do machine learning.

Conclusion: Tom: Alright, we’ve reached the end of our time with “Negative Ontology of True Target for Machine Learning.” Jane, let’s wrap this up. What’s the one thing you want our listeners to remember?

Jane: I want them to remember that the paper challenges the very foundation of supervised learning. It says the true target might not exist, and that’s not a problem. It’s an opportunity to build more inclusive systems.

Tom: And they didn’t just leave it as philosophy. They built a whole framework around it. MIATTs for the data, LAF for evaluation, UTTL for learning.

Jane: And they showed it works on a real task. They trained a segmentation model without any expert labels and got results that were logically coherent and visually plausible.

Lu: And the implications go beyond just getting a model to work. This framework turns the modeling process into a learning experience for the human. It supports self-education and professional growth.

Meng: From an engineering standpoint, I was impressed that they handled the non-differentiability of the logical metric by splitting optimization from selection. That’s a practical solution to a real problem.

Tom: And Lalam, you’ve been listening to all of this. What’s your take on the bigger picture?

Lalam: I think the most impactful vision here is cultural. We’re moving from a world where eye is an oracle that gives you the answer, to a world where eye is a collaborator that helps you form your own understanding. That’s a profound shift in how humans and machines relate.

Jane: And that’s what makes this paper exciting. It’s not just about better predictions. It’s about better people.

Tom: Well said, Jane. So we’re saying goodbye to this paper, but I have a feeling we’ll be talking about these ideas for a long time.

Jane: Absolutely. And we’ve got another fascinating paper coming up next. So stay tuned, everyone.

Tom: Thanks for listening, and see you on the next episode.