My notes on causal model learning and modular computation line of work which connects to learning independent mechanisms. (very roughly written, just capturing the gist of a couple of papers with excerpts).
GNNs are not suited to extract objects from images, for two main reasons:
Inspiration in cognitive science, production systems. Expressing propositional knowledge is not the strength of current neural networks systems. Merge propositional logic and deep learning.
The paper builds upon the classical idea of production systems that operate in a condition-action way by updating the working memory or performing a certain action on the outside world.
MLP and attention mechanism to determine rule-entity binding.
Production system consits of a set of entities and set of rules.
Production rules are modular, each production rule represents a unit of knowledge.
Production rules are sparse, i.e. dependencies among entities are sparse.
Use straight-through Gumbel-softmax for rule and slot selection.
The topology of the graph induced in NPS is dynamic, while the topology in GNNs is fixed.
Use Ka physics environment with boxes?
Learns latent representations of objects and discovers causal mechanisms between these objects (in representations space). Latent states and transition functions of the latent state have entity-centric inductive bias. The transition functions (called modules) are independently parametrized and shared across entities. The model outputs a set of categorical selection variables to see which mechanisms are applied. Modules are sampled according to their attention importance weights.
A module can only consider a handful of slots as its input argument. In this work they consider only unary modules, n-ary they leave for future work.
Sampling of mechanism to apply via Gumbel-softmax.
Cognitive scientists theorize that humans generalize broadly with “schemas” they discover for regularly occuring structures.
schemas are composable representations over portions of our observations.
In this work, we hypothesize that we can develop a single deep RL architecture that can exhibit multiple types of generalization if it can learn schema-like representations for regularly occurring structures within its experience.
FARM - architecutre to discover task-relevant schemas.
Parallel to word embeddings - better representation for words since it can be used in various in various ways (vector addition) in comparison to standard words.
To have the modules coordinate what they attend to, they share information using transformer-style attention (Vaswani et al., 2017).
One of the claims of the paper is that spatial attention can be detrimental to reinforcement learning.
FARM attends to an observation with feature attention as opposed to spatial attention.
Agent takes in partial observation and task description.
Use Gumbel-based hard selection of appropriate schemata.
“complex generative model, temporal or not, can be thought of as the composition of independent mechanisms or “causal” modules” - basically we are thinking of purely Structural Equation Models that are modular here.
“ne may hypothesize that if a brain is able to solve multiple problems beyond a single i.i.d. (independent and identically distributed) task, they may exploit the existence of this kind of structure by learning independent mechanisms that can flexibly be reused, composed and re-purposed” - trying to write a translation of this statement here, since mechanisms are not affected by distribution shifts which are realized by interventions, they are reusable across different distributions.
“The central question motivating our work is how a gradient-based deep learning approach can discover a representation of high-level variables which favour forming independent but sparsely interacting recurrent mechanisms in order to benefit from the modularity and independent mechanisms assumption.” - core motivation of the work
“The central question motivating our work is how a gradient-based deep learning approach can learn a representation of high level variables which favour learning independent but sparsely interacting recurrent mechanisms in order to benefit from such modularity assumption.”
Each RIM (Recurrent Independent Mechanism) should be activated and updated when the input is relevant to it.
Attention mechanism selects and only activates subsets of independent mechanisms.
RIMs can be used as a drop-in replacement for a GRU/LSTM layer.