Mariel K. Goddu

Torresan & Baltieri propose a framework to analyze causal learning in natural and artificial agents. They are motivated by an important and widely neglected question: How do agents acquire causal models? Most approaches _assume_ a model –– a set of causal variables and relations –– and then use it to assess an agent’s capacities. By contrast, T&B build on philosophy and comparative psychology to develop a formal approach to the problem “how a causal viewpoint can emerge from an agent’s first-person experience... or as the machine learning analogy goes, _from pixels_” (their italics). My purpose will be to dispute this: I think there is little useful analogy between “learning from pixels” and “learning from experience” –– or, more specifically, between the machine learning techniques that T&B’s framework employs and actual causal learning (i.e., the natural phenomenon occurring in animals). Thus, I do not think the account can unify natural and artificial causal cognition in the way that T&B hope.

Why are humans born “helpless”? Cusack et al. propose that human infants’ helplessness has learning benefits analogous to training foundation models in machine learning. The infant’s “limited repertoire of adaptive behavior,” they argue, affords a period of self-supervised learning in which representations are “not yet connected to outputs and are therefore not acted upon” [1]. This “pre-training” stage of sensory data-crunching makes the acquisition of later abilities more efficient. We agree that infants’ ability to learn from statistical regularities is a key driver of early development. However, we challenge Cusack et al.’s stage-like proposal that sensory learning is initially disconnected from motivations and goal-directed actions. This claim is inconsistent with core empirical findings, and it underestimates the many ways in which infants’ behavior is adaptive. We encourage researchers to “think like a baby”: that is, to assume the infant’s agential perspective to see the goals and rewards that drive their early learning. Four insights from developmental science are at odds with Cusack et al.’s proposal.

In their recent Opinion in TiCS, Yildirim and Paul propose that large language models (LLMs) have ‘instrumental knowledge’ and possibly the kind of ‘worldly’ knowledge that humans do. They suggest that the production of appropriate outputs by LLMs is evidence that LLMs infer ‘task structure’ that may reflect ‘causal abstractions of... entities and processes in the real world.' While we agree that LLMs are impressive and potentially interesting for cognitive science, we resist this project on two grounds. First, it casts LLMs as agents rather than as models. Second, it suggests that causal understanding could be acquired from the capacity for mere prediction.

Causal understanding is a defining characteristic of human cognition. Like many animals, human children learn to control their bodily movements and act effectively in the environment. Like a smaller subset of animals, children intervene: they learn to change the environment in targeted ways. Unlike other animals, children grow into adults with the causal reasoning skills to develop abstract theories, invent sophisticated technologies and imagine alternate pasts, distant futures and fictional worlds. In this Review, we explore the development of human-unique causal learning and reasoning from evolutionary and ontogenetic perspectives. We frame our discussion using an ‘interventionist’ approach. First, we situate causal understanding in relation to cognitive abilities shared with non-human animals. We argue that human causal understanding is distinguished by its depersonalized (objective) and decontextualized (general) representations. Using this framework, we next review empirical findings on early human causal learning and reasoning and consider the naturalistic contexts that support its development. Then we explore connections to related abilities. We conclude with suggestions for ongoing collaboration between developmental, cross-cultural, computational, neural and evolutionary approaches to causal understanding.

How should scientists carve up cognition to generate good predictions, explanations, and models of cognition? This chapter argues that cognitive categories should be constructed the same way that biological categories are: in terms of homology. The chapter adapts a developmental account of trait identity from evolutionary-developmental biology to make sense of the notion of “cognitive homology.” The consequence is that both brain structures and the organism’s ongoing interactions with the environment are crucial for individuating cognitive homologies, and thus for cognitive ontology.