The Learn-Experiment-Interact Loop - A Philosophy for Navigating Complexity

This belief system I learned from Mu Sigma—I’m ever grateful for that. It comes from Dhiraj’s learnings and his strong understanding and convictions about complexity science, which drives the world and its dynamics.

The Three Core Elements

The framework has three things: learning over knowing, extreme experimentation, and interaction property.

I want to put it like a loop—it’s a cycle, right? It’s learn, experiment, interact, and repeat. Learn, experiment, interact, and repeat.

Why I’m saying this is because, for example, most of the time people feel like, “Okay, there is nothing to learn in this subject,” or “there is nothing to learn in this role,” or maybe we don’t need to learn one further thing. But ultimately, we need to find what could drive better learning, right? That’s why I put it as a loop—because interactions is what makes us learn something new.

Once we learn things, we end up doing or we need to do experiments. The reason is, such experiments help us to test the learning, validate them, and improvise it. I mean, now you’ll be able to understand the gaps. And then further we interact from the learnings or maybe from whatever we got from the experiments.

So it’s like you learn, you experiment, you interact. Interaction could be anything. You interact with the subject. You interact with the concepts and questions, try to reflect on what we understand best or what is not going well.

Such interactions could be with yourself, with the subject, with your friends, with the industry itself, right? At the end of the day, everything is interactions over there. Then that should guide us—okay, what extra we need to learn. Then, okay, we’ll learn something. Then, are we learning in the right way? Can we apply it somewhere? Can we test whether this will work out or not? Then we are interacting, which means, oh, I have done something—can we see whether it will work or not.

A Practical Example

For example, let’s say we are doing MS in computer science or data science or anything related to this field. So we learn something called machine learning. But how do we test whether we have satisfaction in it, or how do we move ourselves one step extra, right? Academic setups may not be enough, or they might be very conservative.

So we might be doing our own experiments, which means our own projects or our own applied research or research work. The term “experiments” is actually very good and very feasible and democratized to everyone, right? Because research may not be possible for everyone. But experiments are a way to answer your questions, a way to address some of the common problems, a way to relate ourselves with the concepts and the domain.

Okay, we have done an experiment. So next step is interaction, which means we have to interact with this domain or we have to interact with this ecosystem to present, “Oh, this is what I’ve learned. This is what I’ve observed, which is going right and wrong. And this is what I feel I need to focus more on,” right? Then you’ll again learn, then you’ll experiment, again interact.

This, I would say, it need not be in a continuous loop—maybe it can be on a repeat mode, but you learn, then interact, then you’ll try to experiment. I guess these are the three important elements that I could see that are working out, right?

My Personal Application

For example, since I have understanding in machine learning, I’m not just keeping myself like, “Okay, I know ML. I’ll just go and prepare for interviews.” But I interact and take problems from different departments or maybe I collaborate with a couple of folks or labs, trying to understand their problem statement and also how we can approach it from the data science perspective. So that’s something I feel good and also motivating, right? Because I could see I’m able to empathize with the problem statement. I could focus on how we can build the solution, able to do literature survey and things.

And then, so before doing this itself, the learnings and the prior experiments help me to naturally align or self-organize things. “Okay, this is how we might need to look into the problem.” Because over the years I’ve interacted with a very large number of people, and they always think of things as input-to-output mapping. Like if someone comes up with a problem statement, people just end up throwing, “Oh, can we build this? You can use this algorithm. You can use random forest or you can adjust some hyperparameter values.”

But ultimately, even any problem statement, even if it is a research problem statement, we cannot just throw some judgments. So that interaction helps us to understand.

A Real Research Example

For example, last two weeks I’ve been working with a PhD scholar at an engineering lab at Columbia. So we started with the problem statement with very vague aspiration that, “Oh, okay, we need to build or we need to build a large model for this particular material science or industrial engineering problem statement.” And just by throwing jargons—way of conversation was, “Oh, let’s build a neural network, or let’s build a graph neural network, or let’s build another LoRA-based or agents.” So that’s what the conversation was looking like.

And then over the conversation and reading and doing some workout, we were able to come up with, “Oh, okay, we have to think from the first principles, we have to think from overall—this is how we need to look at the problem components,” because one problem might be dissected into multiple components. How are we going to achieve it?

Sometimes, some of the problem statements—for example, this particular problem statement, if you take—which is like, we want to understand and simulate black swan events, which are very rare and surprise events which might break the systems, right? For example, you consider it for fraudulent tasks, or you consider it for cybersecurity attacks, risk-related things. Almost all these things are very rare events, which you cannot just fit with very conventional machine learning or maybe probabilistic distributions. And the data is less over there.

And another thing is, it’s not just like one small project, which is like “just build the model.” No, it is like kind of a research problem statement where we might need to craft a very new or maybe very novel modeling technique, which may be the combination of Newtonian mechanics plus data-driven methods.

So in that case, this arrival of “okay, we need to combine Newtonian mechanics plus data-driven way” itself happened through multiple interactions. I’m not saying just conversation, because conversation is one way of interaction, but the other way of interaction is working on or reading through or trying to conceptualize it—how can we combine two different elements together?

For example, in a system, you take two subsystems or two components. If we want to combine them or couple them, I don’t have any Newtonian mechanics way of representation for that. I just have a representation or equation for a subcomponent only. But how can I combine these two Newtonian mechanics-based subcomponents into a coupled one, which means for which we don’t have any equations?

So likewise, this level of understanding and problem definition itself happened through this loop only, I would say. And I was quite happy—I was able to drive that conversation. “Okay, this is how it has to go. We shouldn’t just throw darts randomly, but we have to learn from the prior trajectories.” Okay, we started with conversation, but it’s not going well, right?

And yeah, I seriously value that. I really liked it. And I value such projects, ideas, rather than just someone coming to me, “Hey, how can I memorize type one error and type two errors?” But when you start conversing in a way that I have been doing, you will easily relate and you will easily organize your knowledge in a way that you grasp that concept. That’s what I’m trying to do. I’ve been doing it and I see that more and more things make me have a better understanding, right?

The Bottom Line

So yeah, anytime when you have a problem or if you are struggling with a subject or maybe if you are not sure about what project to be built, I would really say: start interacting, start experimenting. And then it will lead to better learning, and it’s a loop, I would say. Once you are comfortable, it’s all going to be different loops.

The beauty of this approach is that it doesn’t have to be a rigid sequence. Sometimes you might learn first, then experiment, then interact. Other times you might start with interaction, then learn, then experiment. The key is recognizing that these three elements feed into each other, creating a dynamic process that keeps evolving your understanding and capabilities.

This framework has fundamentally changed how I approach complex problems, and I hope it might be useful for others navigating similar challenges in their academic and professional journeys.