Previous and Current Work (check out my Google Scholar profile)

Discovering a predictive principle for self-organization

2020-10-01 (3).png

What does it mean to find an effective scientific model or theory?


Projects and internships (condensed matter experiments, laser physics, and some theory)



How can we understand the complexity of the everyday world around us? Indeed, starting from fundamental physics, it seems quite surprising that such intricate, precisely coordinated structures can emerge spontaneously – or “self-organize” – from a chaotic origin. In this line of research, we found a theoretical principle, which we termed Low Rattling, that can explain and predict some degree of spontaneous organization in complex systems. It seems that this principle may be quite ubiquitously at work in the world, but it remains to be seen whether it can shed light on the most exciting examples of self-organization: origins of life and society. One exciting phenomenon that Low Rattling does seem to explain is how non-biological complex systems may naturally exhibit something very similar to “adaptation to their environment” familiar to us from biological world.


We validated our theory on a number of examples, such as several toy dynamical systems, random Markov processes, the Vicsek model, and, most prominently, in experiments with a swarm of simple robots. In this last example, we showed that our theory can make quantitative predictions about the self-organizing properties of the swarm, and can further allow to control the swarm behaviors in a regime where traditional control theory tools, and even Machine-Learning techniques, all fail.


Check out the following resources for more information, ranging from popular science texts to full technical exposition

2020-10-01 (1).png

Low rattling: A predictive principle for self-organization in active collectives

The flagship paper of this work - a concise high-level descriptions, detailing our main results for a general science audience.

You can access some videos and full text here 

(see the SI for full technical exposition of the theory)

See also a 12-minute interview with Science mag where I discuss the idea (start at 13:00 here).


Science is concerned with building simple yet effective descriptions of our infinitely-complex real world. "All models are wrong, but some are useful" is a common proverb (attributed to George Box) that captures this sentiment. 

But if we are not really looking for "the truth," then what exactly is it that we are looking for as scientists? If all models are wrong, can one be "less wrong" than another? What metric could we use to decide? And how is it that in our everyday lives we naturally intuit effective descriptions of things around us, without thinking about it nor referring to some "fundamental" atoms? 

A good example here is the intuitive categories of "living" vs. "non-living": these may actually be quite hard to tell apart when looking at an atomic scale - yet they are qualitatively distinct and very useful as effective high-level models. Along similar lines, when should we model a tight-knit bacterial ecosystem as effectively a single multicellular organism? Or how should we best draw the lines among various bacteria to delineate distinct "species"?
All these are questions of how we shall choose among various possible ways to describe the same reality.

See more resources here:


Causal Geometry

This paper builds on the idea that most effective models should clearly show the causes affecting system behavior. Marrying Judea Pearl's definition of causality in terms of counterfactuals, with the math of Information Geometry, we devise a new construct we term "Causal Geometry." In this paradigm, the best models are ones where the allowed interventions on the system most directly influence the observed effects (cf. "Causal Emergence"). In our geometric language, this is elegantly achieved by matching up the geometry of the "intervention manifold" to that of the "effect manifold."

Check out this blog post for a fun high-level discussion of these ideas!


Before starting my doctorate research at MIT, I worked on a number of different projects at different labs around the world. Many of these were experimental studies working with lasers and condensed matter systems, as well as a few theoretical explorations in different areas.




Why do we naturally tend to anthropomorphise some motion (e.g., "a ball wants to roll downhill")? Here we used Bayesian inference to understand why we perceive some motion as "agent-like" and some as "object-like." We hypothesised that the distinction may come from perceived planning-horizon: unlike for objects, agents' motion may be described in terms of a longer-term plan.