This analysis reveals recurrent deep neural networks do not consistently outperform feedforward models in visual recognition tasks, indicating complexity in model design.
Key Points
Increased model size significantly enhanced performance in visual recognition tasks, while architecture did not impact results.
Larger models showed better alignment with human perceptions of task difficulty, regardless of architecture used.
The study identified that recurrent models may not be superior to feedforward counterparts for modeling human behavior in recognition tasks.
Findings suggest reconsideration of recurrent deep neural networks as ideal representations of human visual recognition.