Neural Predictivity and Representational Alignment in Deep Neural Networks: A Comparison Across Architectures and Learning Regimes

Understanding the degree to which artificial neural networks align with biological visualsystems remains a central question at the intersection of neuroscience and machine learning.Using real neural recordings from macaque inferior-temporal (IT) cortex paired with the BrainScore benchmark, we conduct a systematic comparison of five contemporary architecturesAlexNet, VGG-16, ResNet-50, Vision Transformer (ViT-B/16), and an untrained ResNet-50control to quantify their predictive alignment with primate object recognition circuitry. Weemploy multiple complementary analysis methods: encoding analysis via PLS regression withcross-validation, representational similarity analysis (RSA) via Spearman correlation, and shapebias probing on cue-conflict stimuli. Our results reveal that (1) training provides a statisticallyrobust ~66% boost in neural predictivity compared to architecture alone (0.68 vs. 0.41 of noiseceiling; p < 0.0001) (2) the classic expectation that predictivity should monotonically increasewith network depth does not hold, with ResNet-50's best alignment occurring at intermediatelayers rather than the final layer (3) encoding predictivity and representational geometry metricsyield substantially different model rankings, suggesting they capture complementary rather thanredundant aspects of neural alignment. These findings demonstrate that the CNN-brain analogyis empirically grounded but nuanced, and highlight the importance of combining multiple metricswhen assessing biological realism in artificial systems.