[Paper Review] Explanatory models in neuroscience: Part 2 -- constraint-based intelligibility

Rosa Cao, Daniel Yamins|arXiv (Cornell University)|Apr 3, 2021

Neural dynamics and brain function28 references25 citations

TL;DR

The paper argues that intelligibility in neuroscience comes from understanding dependencies between behavior and causally responsible factors, emphasizing top-down constraints (evolutionary/developmental) alongside bottom-up mechanisms, and showing how goal-driven hierarchical CNNs illuminate brain function.

ABSTRACT

Computational modeling plays an increasingly important role in neuroscience, highlighting the philosophical question of how computational models explain. In the context of neural network models for neuroscience, concerns have been raised about model intelligibility, and how they relate (if at all) to what is found in the brain. We claim that what makes a system intelligible is an understanding of the dependencies between its behavior and the factors that are causally responsible for that behavior. In biological systems, many of these dependencies are naturally "top-down": ethological imperatives interact with evolutionary and developmental constraints under natural selection. We describe how the optimization techniques used to construct NN models capture some key aspects of these dependencies, and thus help explain why brain systems are as they are -- because when a challenging ecologically-relevant goal is shared by a NN and the brain, it places tight constraints on the possible mechanisms exhibited in both kinds of systems. By combining two familiar modes of explanation -- one based on bottom-up mechanism (whose relation to neural network models we address in a companion paper) and the other on top-down constraints, these models illuminate brain function.

Motivation & Objective

• Articulate what makes a system intelligible by detailing dependencies between behavior and causal factors, including top-down constraints.
• Demonstrate how optimization in neural network models captures dependencies shaping brain function.
• Show that goal-driven HCNNs provide explanatory power beyond simple curve-fitting.
• Argue that combining bottom-up mechanisms with top-down constraints yields better brain models and predictions.

Proposed method

• Discuss the notion of intelligibility as dependencies between behavior and responsible factors.
• Contrast bottom-up mechanistic explanations with top-down functional constraints from evolution and development.
• Use Olshausen and Field and sparse-autoencoder work to illustrate constraint-based intelligibility.
• Explain how goal-driven HCNNs are optimized for ethologically plausible tasks to model ventral visual pathways.
• Argue that HCNNs predict neural data across V1, V4, and IT without direct neural data in training.

Experimental results

Research questions

• RQ1 How do top-down constraints (evolutionary/developmental) contribute to intelligibility in neural systems?
• RQ2 Can goal-driven HCNNs, optimized for ecologically relevant tasks, account for neural responses in ventral visual areas without neural data during training?
• RQ3 What is the explanatory value of combining bottom-up mechanisms with constraint-based explanations for brain function?

Key findings

• HCNNs optimized for vision tasks predict neural responses in multiple ventral stream areas quite well, even without neural data in training.
• Top-down constraints from functional goals can shape upstream neural representations, aligning with brain data better than simple curve-fit models.
• Intelligibility arises when examining dependencies between mechanism and behavior, including evolutionary and developmental constraints.
• The convergence of brain and HCNN architectures under shared goals highlights why certain features are robust across biological and artificial systems.