Thrilled to announce our new publication titled 'Decoding predicted future states from the brain's physics engine' with @beth_miecz, Cyn X. Fang, @Nancy_Kanwisher, and Josh Tenenbaum. . (1/n).

RT @Mohite_Vaish: Some people are face blind. Others are super recognizers. Curious?. Join us for a fascinating session on the Indian Face….

RT @PramodRT9: Thrilled to announce our new publication titled 'Decoding predicted future states from the brain's physics engine' with @bet….

Thanks to my co-authors and all the people who gave constructive feedback over the course of this project! Special shout out to Kris Brewer for shooting the videos used in Experiment 1 and Georgina Woo for her deep neural network expertise. (12/12).

Our findings show that PN has abstract object contact info & provide the strongest evidence yet that PN is engaged in predicting what will happen next. These results open many new avenues of investigation into how we understand, predict, & plan in the physical world . (11/n).

Our main results are i) not present in the ventral temporal cortex, ii) not present in the primary visual cortex -- i.e, our stimuli were unlikely to have low-level visual confounds and iii) are replicable with different analysis criteria & methods. See paper for details. (10/n).

Short answer: Yes! Using MVPA we found that the PN has information about predicted contact events (i.e., collisions). This was true not only within a scenario (the ‘roll’ scene above), but also generalized across scenarios indicating the abstractness of representation. (9/n)

That is, .(8/n). When we see this: Does the PN predict this?

In our second pre-registered fMRI experiment, we tested the central tenet of the ‘physics engine’ hypothesis – that the PN runs forward simulations to predict what will happen next. If true, PN should contain information about predicted future states before they occur. (7/n).

Given their importance for prediction, we hypothesized that the PN would encode object contact. In our first pre-registered fMRI experiment, we used multi-voxel pattern analysis (MVPA) and found that only PN carried scenario-invariant information about object contact. (6/n)

If a container moves, then so does its containee, but the same is not true of an object that is merely occluded by the container without contacting it! . (5/n)

However, there was no evidence for such predicted future state information in the PN. We realized that object-object contact is an excellent way to test the Physics Engine hypothesis. When two objects are in contact, their fate is intertwined: . (4/n).

These results have led to the hypothesis that the Physics Network (PN) is our brain’s ‘Physics Engine’ – a generative model of the physical world (like those used in video games) capable of running simulations to predict what will happen next. (3/n).

How do we understand, plan and predict in the physical world? Prior research has implicated fronto-parietal regions of the human brain (the ‘Physics Network’, PN) in physical judgement tasks, including in carrying representations of object mass & physical stability. (2/n)

RT @bkhmsi: 🚨New Preprint!!. Thrilled to share with you our latest work: “Mixture of Cognitive Reasoners”, a modular transformer architectu….

RT @RTomMcCoy: 🤖🧠Paper out in Nature Communications! 🧠🤖. Bayesian models can learn rapidly. Neural networks can handle messy, naturalistic….

RT @_coltoncasto: New paper!🧠**The cerebellar components of the human language network**.with: @smallhannahe @MoshePoliak @GretaTuckute @be….

RT @sparuniisc: In a study now out in @eLife, @GeorginJacob @PramodRT9 and I have some exciting results: a novel computation that helps the….

RT @coryshain: New brain/language study w/ @ev_fedorenko! We applied task-agnostic individualized functional connectomics (iFC) to the enti….

RT @aran_nayebi: Are there fundamental barriers to AI alignment once we develop generally-capable AI agents?. We mathematically prove the a….