On logical inference over brains, behaviour, and artificial neural networks

Abstract

In the cognitive, computational, and neuro- sciences, we often reason about what models (viz., formal and/or computational) represent, learn, or "know", as well as what algorithm they instantiate. The putative goal of such reasoning is to generalize claims about the model in question to claims about the mind and brain. This reasoning process typically presents as inference about the representations, processes, or algorithms the human mind and brain instantiate. Such inference is often based on a model's performance on a task, and whether that performance approximates human behaviour or brain activity. The model in question is often an artificial neural network (ANN) model, though the problems we discuss are generalizable to all reasoning over models. Arguments typically take the form "the brain does what the ANN does because the ANN reproduced the pattern seen in brain activity" or "cognition works this way because the ANN learned to approximate task performance." Then, the argument concludes that models achieve this outcome by doing what people do or having the capacities people have. At first blush, this might appear as a form of modus ponens, a valid deductive logical inference rule. However, as we explain in this article, this is not the case, and thus, this form of argument eventually results in affirming the consequent – a logical or inferential fallacy. We discuss what this means broadly for research in cognitive science, neuroscience, and psychology; what it means for models when they lose the ability to mediate between theory and data in a meaningful way; and what this means for the logic, the metatheoretical calculus, our fields deploy in high-level scientific inference.

Obstacles to inferring mechanistic similarity using Representational Similarity Analysis

“…As such, RSA provides a benchmark for rejecting possible models. However, the above simulations show that RSA may be a misleading benchmark for selecting models – two systems may show similar representational geometries and yet work on very different transformations and features of input stimuli (for an in depth discussion about inferring similarity of causal mechanisms from similar outcomes see [38]).…”

Inferring the nature of linguistic computations in the brain

“…While the goal for many ANN methods is to predict new data, in the cognitive and neural sciences, the goal is to explain the capacities of a system and to understand the mechanisms that lead to observed data. The only way to do this is to pose hypotheses about brain mechanisms, constrained through abductive inference based on what we know to an acceptable degree of verisimilitude about language and the brain, and test explicitly for these mechanisms [ 7 , 8 , 15 , 24 ]. Some forms of data science applications simply replace the thing we aim to understand (e.g., the brain, behavior, the capacity of a system for human language) with another black box (e.g., an ANN model; [ 25 , 26 ]).…”

Encoding of speech in convolutional layers and the brain stem based on language experience

“…We do, however, show, that computations and encodings are similar. By focusing on interpretable comparison (rather than on correlations) and by focusing on internal representations rather than only on behavioral data (see discussion in Guest and Martin 2021), we argue that similarities in both representations and computations exist between brains and deep convolutional layers. These similarities open up possibilities for modeling work in order to gain insights both for how humans acquire and process speech as well as for how deep learning models learn internal representations.…”

“…Comparing representations and computations in the human brain and deep learning models is a complex task. The goal of this paper is not to argue that human speech processing operates exactly as in deep convolutional networks (for problems with such an approach, see Guest and Martin 2021). We do, however, argue that computations and encodings are similar in interpretable ways between the two signals and that they result from similar underlying mechanisms (Section 5.2).…”