Lotem Elber-Dorozko

How are new Bayesian hypotheses generated within the framework of predictive processing? This explanatory framework purports to provide a unified, systematic explanation of cognition by appealing to Bayes rule and hierarchical Bayesian machinery alone. Given that the generation of new hypotheses is fundamental to Bayesian inference, the predictive processing framework faces an important challenge in this regard. By examining several cognitive‐level and neurobiological architecture‐inspired models of hypothesis generation, we argue that there is an essential difference between the two types of models. Cognitive‐level models do not specify how they can be implemented in brains and include structures and assumptions that are external to the predictive processing framework. By contrast, neurobiological architecture‐inspired models, which aim to better resemble brain processes, fail to explain important capacities of cognition, such as categorization and few‐shot learning. The “scaling‐up” challenge for proponents of predictive processing is to explain the relationship between these two types of models using only the theoretical and conceptual machinery of Bayesian inference.

How does neuronal activity give rise to cognitive capacities? To address this question, neuroscientists hypothesize about what neurons “represent,” “encode,” or “compute,” and test these hypotheses empirically. This process is similar to the assessment of hypotheses in other fields of science and as such is subject to the same limitations and difficulties that have been discussed at length by philosophers of science. In this paper, we highlight an additional difficulty in the process of empirical assessment of hypotheses that is unique to the cognitive sciences. We argue that, unlike in other scientific fields, comparing hypotheses according to the extent to which they explain or predict empirical data can lead to absurd results. Other considerations, which are perhaps more subjective, must be taken into account. We focus on one such consideration, which is the purposeful function of the neurons as part of a biological system. We believe that progress in neuroscience critically depends on properly addressing this difficulty.

It is generally accepted that, in the cognitive and neural sciences, there are both computational and mechanistic explanations. We ask how computational explanations can integrate into the mechanistic hierarchy. The problem stems from the fact that implementation and mechanistic relations have different forms. The implementation relation, from the states of an abstract computational system to the physical, implementing states is a homomorphism mapping relation. The mechanistic relation, however, is that of part/whole; the explaining features in a mechanistic explanation are the components of the explanandum phenomenon and their causal organization. Moreover, each component in one level of mechanism is constituted and explained by components of an underlying level of mechanism. Hence, it seems, computational variables and functions cannot be mechanistically explained by the medium-dependent states and properties that implement them. How then, do the computational and the implementational integrate to create the mechanistic hierarchy? After explicating the general problem, we further demonstrate it through a concrete example, of reinforcement learning, in the cognitive and neural sciences. We then examine two possible solutions. On one solution, the mechanistic hierarchy embeds at the same levels computational and implementational properties. This picture fits with the view that computational explanations are mechanistic sketches. On the other solution, there are two separate hierarchies, one computational and another implementational, which are related by the implementation relation. This picture fits with the view that computational explanations are functional and autonomous explanations. It is less clear how these solutions fit with the view that computational explanations are full-fledged mechanistic explanations. Finally, we argue that both pictures are consistent with the reinforcement learning example, but that scientific practice does not align with the view that computational models are merely mechanistic sketches.

A popular view presents explanations in the cognitive sciences as causal or mechanistic and argues that an important feature of such explanations is that they allow us to manipulate and control the explanandum phenomena. Nonetheless, whether there can be explanations in the cognitive sciences that are neither causal nor mechanistic is still under debate. Another prominent view suggests that both causal and non-causal relations of counterfactual dependence can be explanatory, but this view is open to the criticism that it is not clear how to distinguish explanatory from non-explanatory relations. In this paper, I draw from both views and suggest that, in the cognitive sciences, relations of counterfactual dependence that allow manipulation and control can be explanatory even when they are neither causal nor mechanistic. Furthermore, the ability to allow manipulation can determine whether non-causal counterfactual dependence relations are explanatory. I present a preliminary framework for manipulation relations that includes some non-causal relations and use two examples from the cognitive sciences to show how this framework distinguishes between explanatory and non-explanatory, non-causal relations. The proposed framework suggests that, in the cognitive sciences, causal and non-causal relations have the same criterion for explanatory value, namely, whether or not they allow manipulation and control.