Statistical prediction alone cannot identify good models of behavior | JoVE Visualize

Area of Science:

Behavioral science
Cognitive science
Computational neuroscience

Background:

Deep neural networks (DNNs) offer powerful statistical prediction capabilities.
However, their use in scientific explanation, particularly in vision research, has been questioned.
A dissociation between prediction and explanation is a growing concern in computational modeling.

Purpose of the Study:

To investigate the extent to which the dissociation between statistical prediction and scientific explanation, observed in vision research using DNNs, applies to other domains of behavior research.
To determine if this dissociation is an inherent limitation when fitting large, weakly-theorized models to stochastic behavioral data.

Main Methods:

Analysis of existing literature and theoretical frameworks concerning supervised learning models and behavioral data.
Examination of the fitting process of large models, such as deep neural networks and other supervised learners.
Consideration of the impact of restricted samples and highly stochastic behavioral phenomena on model interpretability.

Main Results:

The dissociation between statistical prediction and scientific explanation is not limited to vision research but is prevalent across various behavior research domains.
This phenomenon is an unavoidable consequence of fitting large supervised learning models with minimal theoretical grounding to limited, noisy behavioral datasets.
The inherent stochasticity of behavioral data exacerbates the challenge of extracting robust scientific explanations from predictive models.

Conclusions:

The findings highlight a fundamental limitation in using complex, data-driven models for scientific explanation in behavior research.
Researchers must acknowledge and address the inherent dissociation between prediction and explanation when employing large models like deep neural networks.
Future research should focus on developing methods that bridge the gap between statistical prediction and meaningful scientific understanding in the analysis of behavioral phenomena.