Abstract
The primate inferior temporal (IT) cortex, at the apex of the ventral visual stream, encodes information that supports diverse representational goals—from recognizing objects to determining which images are likely to be remembered. Specific artificial neural networks (ANNs), that currently serve as the leading computational hypotheses of ventral stream processing, are typically trained exclusively for object recognition. We asked whether incorporating image memorability as an additional optimization objective could improve ANN–brain alignment. Models optimized for memorability explained additional, non-overlapping variance in IT responses beyond that captured by recognition-optimized networks, indicating that memorability and recognition rely on partly independent dimensions of IT representation. Notably, these models also exhibited fewer non–brain-like units, bringing their representational geometry closer to that of IT. Furthermore, networks jointly optimized for both objectives were more predictive of human memorability than memorability-only models, while maintaining their alignment with human object recognition performance patterns. Together, these findings suggest that IT encodes multiple representational goals and that models trained solely for recognition provide an incomplete account of ventral stream computation.
Significance: Brain regions often serve multiple representational goals, and identifying those goals is critical because they provide the key to building better encoding models of the system. The primate ventral visual stream has traditionally been understood as a pathway for object recognition, with the inferior temporal (IT) cortex regarded as its core substrate. However, IT responses also predict image memorability—a robust phenomenon whereby some images are consistently remembered better than others. Here we show that memorability constitutes a separable representational goal of IT. ANNs optimized for memorability explained neural variance not captured by recognition models, and the two objectives produced distinct representational geometries. Critically, models jointly optimized for both recognition and memorability provided the best match to IT responses, improved prediction of human memorability, and preserved recognition performance. These findings highlight memorability as an organizing principle of IT and demonstrate that multi-goal optimization yields more brain-like computational models of vision.
Introduction
For over a decade, deep neural networks trained on ImageNet for object recognition have been the "gold standard" for modeling the primate visual system. While these models effectively predict neural responses in the Inferior Temporal (IT) cortex—the brain's object recognition hub—they fail to capture a significant amount of neural variance. This gap suggests that the ventral stream may not be a single-purpose machine. Beyond identifying objects ("What is this?"), the brain also evaluates the intrinsic property of an image that makes it likely to be remembered ("Is this important enough to remember?"). This study investigates whether image memorability is a distinct representational goal of the IT cortex and whether incorporating it into computational models improves their alignment with biological vision.
Methods
The researchers employed a multi-pronged approach combining human behavior, macaque neurophysiology, and machine learning.
Behavioral and Neural Benchmarking
Human performance was measured on two distinct tasks: a binary object discrimination task (recognition) and a continuous recognition task (memorability). Simultaneously, large-scale neural activity was recorded from the IT cortex of macaques using chronically implanted Utah arrays while they viewed hundreds of naturalistic images from the ImageNet and LaMem datasets.
ANN Training Regimes
Twenty-six different architectures (including CNNs and Transformers) were trained under three conditions:
- Recognition-only: Standard ImageNet training.
- Memorability-only: Trained from scratch to predict memorability scores using the LaMem dataset (~60,000 images).
- Joint (Recognition + Memorability): Utilizing Learning-without-Forgetting (LwF) to distill knowledge from recognition models while simultaneously learning to predict memorability, preventing "catastrophic forgetting" of object features.
Results
Memorability as a Separable Goal
A key finding was that image memorability is only weakly correlated with recognition ease (r = 0.18). This behavioral dissociation suggests that memorability and recognition rely on partly independent dimensions of representation. Crucially, IT cortex responses were found to carry unique information about memorability that recognition-only models missed.
Improved Brain-Model Alignment
Jointly optimized models significantly outperformed recognition-only models in explaining IT neural variance across all subjects. Variance partitioning revealed that memorability-trained models explained unique neural residuals that recognition models could not account for. Furthermore, joint optimization improved "reverse predictivity" (how well neurons can reconstruct model features), indicating a more reciprocal, structural alignment between the ANNs and the biological brain.
Representational Geometry
Using Centered Kernel Alignment (CKA), the authors demonstrated that memorability training drives a systematic shift in the deeper layers of the networks. These models exhibited fewer "non-brain-like" units, bringing their internal representational geometry closer to that of the macaque IT cortex.
Discussion
This research highlights memorability as an organizing principle of the IT cortex. While the contribution of memorability to neural variance is smaller than that of object recognition, it represents a distinct and significant axis of computation. The success of joint optimization suggests that models trained solely for recognition provide an incomplete account of ventral stream processing.
Future work should address the temporal dynamics of memory formation and explore other potential representational goals, such as image typicality or aesthetic value. Ultimately, building AI that truly replicates human vision requires moving beyond simple classification toward multi-goal optimization that reflects the priorities of biological observers.
Summary updated