Conscious content is revealed by reporting it, often with language. This suggests that high-level variables manipulated consciously are closely related with their verbal forms (like words and phrases).

This yields maybe the most influential inductive bias we want to consider in this paper: that high-level variables (manipulated consciously) are generally verbalizable. To put it in simple terms, we can imagine the high-level semantic variables captured at this top level of a representation to be associated with single words (although we can also use words to identify some lower-level variables).

In practice, the notion of word is not always the same across different languages, and the same semantic concept may be represented by a single word or by a phrase. There may also be more subtlety in the mental representations (such as accounting for uncertainty, concept representation and continuous-valued properties) which is not always or not easily well reflected in their verbal rendering.

Much of what our brains know actually cannot be easily translated in natural language and forms the content of system 1 knowledge. This means that system 2 (verbalizable) knowledge is incomplete: words are mostly pointers to knowledge which belongs to system 1 and thus is in great part not consciously accessible.

The system 2 inductive biases do not need to cover all the aspects of our internal model of the world (they couldn't), only those aspects of our knowledge which we are able to communicate with language. The rest would have to be represented in pure system 1 (non system 2) machinery, such as in an encoder-decoder that could relate low-level actions and low-level perception to semantic variables that can be operated on at the system-2 level.

If there is some set of properties that apply well to some aspects of the world, then it would be advantageous for a learner to have a subsystem that takes advantage of these properties (the inductive priors described here) and a subsystem which models the other aspects. These inductive priors then allow faster learning and potentially other advantages like systematic generalization, at least concerning these aspects of the world which are consistent with these assumptions (system 2 knowledge, in our case).

High-level representations describe verbalizable concepts. There is a simple lossy mapping from semantic representations going through the GWT bottleneck to natural language expressions. This is an inductive bias which could be exploited in grounded language learning scenarios where we couple language data with observations and actions by an agent.

This suggests that natural language understanding systems should be trained in a way that couples natural language with what it refers to. This is the idea of grounded language learning. It would put pressure on the top-level representation so that it captures the kinds of concepts expressed with language.

One can view this as a form of weak supervision, where we don't force the top-level GWT representations to be human-specified labels, only that there is a simple relationship between these representations and utterances which humans would often associate with the corresponding meaning.

Our discussion about causality should also suggest that passive observation may be insufficient: in order to capture the causal structure understood by humans, it may be necessary for learning agents to be embedded in an environment in which they can act and thus discover its causal structure. Studying this kind of setup was the motivation for our work on the Baby AI environment.

Biological phenomena such as bird flocks have inspired the design of several distributed multi-agent systems, for example, swarm robotic systems, sensor networks, and modular robots. Despite this, most machine learning models employ the opposite inductive bias, i.e., with all elements (e.g., artificial neurons) interacting all the time.

The GWT also posits that the brain is composed in a modular way, with a set of expert modules which need to communicate but only do so sparingly and via a bottleneck through which only a few selected bits of information can be squeezed at any time.

If we believe that theory, these selected elements are the concepts present to our mind at any moment, and a few of them are called upon and joined in working memory in order to reconcile the interpretations made by different modular experts across the brain.

The decomposition of knowledge into recomposable pieces, a hallmark of classical AI based on rules also makes sense as a requirement for obtaining systematic generalization: conscious attention would then select which expert and which concepts (which we can think of as variables with different attributes and values) interact with which pieces of knowledge (which could be verbalizable rules or non-verbalizable intuitive knowledge about these variables) stored in the modular experts.

On the other hand, the modules which are not brought to bear in this conscious processing may continue working in the background in a form of default or habitual computation (which would be the form of most of perception).

For example, consider the task of predicting from pixel-level information the motion of balls sometimes colliding against each other as well as the walls. It is interesting to note that all the balls follow their default dynamics, and only when balls collide do we need to intersect information from several bouncing balls in order to make an inference about their future states.

Saying that the brain modularizes knowledge is not sufficient, since there could be a huge number of ways of factorizing knowledge in a modular way. We need to think about the desired properties of modular decompositions of the acquired knowledge, and we propose here to take inspiration from the causal perspective on understanding how the world works, to help us define both the right set of variables and their relationship.

Semantic variables are often also causal variables. We hypothesize that semantic variables are often also causal variables. Words in natural language often refer to agents (subjects, which cause things to happen), objects (which are controlled by agents), actions (often through verbs) and modalities or properties of agents, objects and actions (for example we can talk about future actions, as intentions, or we can talk about time and space where events happen, or properties of objects or of actions).

However, note that we can also name many low-level (like pixels) and intermediate features (like L-shaped edges). It is thus plausible to assume that causal reasoning of the kind we can verbalize involves as variables of interest those semantic variables which we can name, and that they can be at any level of the processing hierarchy in the brain, including at the highest levels of abstraction, where signals from all modalities join, such as pre-frontal cortex, and where concepts can be manipulated in a way that is not specific to a single modality.

The connection between causal representations and modularity is profound: an assumption which is commonly associated with structural causal models is that it should break down knowledge about the causal influences into independent mechanisms.

As explained in Section 4.1, each such mechanism relates direct causes to their direct effect and knowledge of one such mechanism should not tell us anything about another mechanism (otherwise we should restructure our representations and decomposition of knowledge to satisfy this information-theoretic independence property).

This is not about statistical independence of the corresponding random variables but about the algorithmic mutual information between the descriptions of these mechanisms. What it means practically and importantly for out-of-distribution adaptation is that if a mechanism changes (e.g. because of an intervention), the representation of that mechanism (e.g. the parameters used to capture a corresponding conditional distribution) may need to be adapted but that of the others do not need to be tuned to account for that change.

These mechanisms may be organized in the form of a causal graph which scientists attempts to identify. The sparsity of the change in the joint distribution between the semantic variables (discussed more in Section 3.6) is different but related to a property of such high-level structural causal model: the sparsity of the graph capturing the joint distribution itself (discussed in Section 3.8).

In addition, the causal structure, the causal mechanisms and the definition of the high-level causal variables tend to be stable across changes in distribution, as discussed in Section 3.7.