Yoshua Bengio, one of the world's leading experts in artificial intelligence and a pioneer in deep learning, along with a team of researchers have introduced end-to-end deep learning (DL) model in a new paper that builds object-centric representations of entities in videos and operates on them with differentiable and learnable production rules.

A video or any such composed visual environment contains objects and entities that possess visible as well as latent properties to define how they interact with each other. The conventional way of such interactions has been to use equivariant graph neural nets (GNNs). However, this approach has shortcomings since GNNs fail to predispose sparse interactions or factorize knowledge about interactions in an entity-conditional manner.

The paper proposes a new approach, Neural Production Systems (NPS), to address this issue. NPS is made up of rule templates that bind the placeholder variables in these rules to specific entities, serving to factorize entity-specific and rule-based information in rich visual environments. 

While humans function on propositional knowledge (for eg - even if a child doesn't understand gravity, it understands that a plate that drops from a table will break), for DL propositional knowledge remains a challenge for two reasons - 1> Propositions are discrete and independent from each another, and 2> propositions must be quantified in the manner of first-order logic.

Quintessential AI procedure provides an important viewpoint on how propositional inference on symbolic knowledge representations. For example, the 1980s production systems express information through condition-action rules. This system was the basis for the researchers with a DL approach and proposed a neural production system that automatically integrates perceptual processing and subsequent inference for visual reasoning problems. 

The NPS takes inspiration from traditional systems on four essential properties - modular, abstract, sparse and symmetric. While these properties focus on how knowledge is presented, it doesn't include what knowledge is represented. The system architecture also supports the detection and inference of entity representations and the latent rules which govern their interactions.

During the tests for its effectiveness, researchers made NPS perform an arithmetic task that involved learning addition, subtraction, and multiplication; an MNIST Transformation to test scaling ability to richer visual settings; and an action-conditioned world models simulation of a simple physics world.

In the arithmetic task, NPS had a significantly lower MSE than the baseline. In MNIST transformation, NPS successfully learned to represent each transformation using a separate rule, while the physics environment simulation validated NPS’s ability to extrapolate from simple (few object) environments to more complex environments.

Researchers from Mila, University of Montreal, DeepMind, Waverly and Google Brain helped Bengio write the paper titled, "Neural Production Systems", which is available on arXiv.