Abstract Vision deep neural networks (VDNNs) only simulate the attention‐based significance selection function in human visual perception, rather than the full spectrum of visual cognition, reflecting the divide between cognitive science (CS) and artificial intelligence (AI). To address this problem, this work proposes a cognitive modeling framework (CMF) comprising three stages: functional abstraction, operator structuring, and program agent. Then, this work defines the prior information of basic image features as the long‐term memory content in VDNNs. Meanwhile, this work introduces a memory modeling method for VDNNs based on the fast Fourier transform (FFT) and statistical methods—the unbiased mapping algorithm (UMA). Finally, this work develops visual cognitive neural units (VCNUs) and a baseline model (VCogM) based on CMF and UMA, and conduct performance testing experiments on different datasets such as natural scene recognition and agricultural image classification. The results show that VCogM and VCNU achieved state‐of‐the‐art (SOTA) performance across various recognition tasks. The model's learning process is independent of data distribution and scale, fully demonstrating the rationality of cognitive‐inspired modeling principles. The research findings provide new insights into the deep integration of CS and AI.
Li et al. (Tue,) studied this question.