CNCLgithub / mot

Model implementation for "Adaptive computation as a new mechanism of human attention"
0 stars 0 forks source link

Scenes where hierarchical objects dynamically form and dissolve #15

Closed iyildirim closed 3 years ago

iyildirim commented 5 years ago

Just registering here so that we think about it. This is when you have a bird of flocks but then half of the birds separate to their own way. Or when you have two rectangles moving together part ways halfway. etc. Fig. 5 in Balaban et al. 2019 is an example.

iyildirim commented 4 years ago

Initial Proposal:

What is the most efficient way of going about implementing this? I am happy to give it a try if people are busy. Let me know.

iyildirim commented 4 years ago

We would like to model such clusters of beads using Chinese Restaurant Process. We will utilize Pakman & Paninski approach here too using their Neural Clustering Process. This will enable us with a data-driven smart proposal distribution to sample cluster assignments based on the velocities and positions.

belledon commented 4 years ago

I think the data generative procedure is rather elegant

  1. initialize k objects
  2. each object moves using the dynamics from Vul
  3. As objects approach one another the probability increases as a sigmoid peaking to some low value (say 0.1) where distance between two dots is measured as distance from the closest points along the circumference (distance between centers + 2*radius)
  4. given that probability, at each time step, for each dot, the probability of being attached is a multinomial with those weights
  5. If true, then the unordered set of attached objects now moves with the same dynamics applied to its centroid

We can play with clusters have N > 2.. I think it would be cool

iyildirim commented 4 years ago

Let me know if you guys can take a stab at this. I would think it shouldn't be too complicated.