Predictive processing (PP) has become an influential theoretical framework in cognitive science and artificial intelligence (AI), suggesting that the capability of accurately predicting incoming sensory information is key to embodied interactions with the environment. PP suggests that perception is not solely bottom-up but shaped by top-down predictions. Recent evidence indicates that foundational predictive capabilities also exist in early sensory processing. The hypothesis is that low-order visual areas, like the retina, have context-independent, bottom-up predictions (“constraints”) that shape visual perception. This commentary bridges these ideas, which draws on Coombs and Trestman’s multi-trait embodied evolution of cognition and Teufel and Fletcher’s (Reference Teufel and Fletcher2020) distinct forms of prediction. Synthesizing these perspectives enhances understanding of foundational predictive mechanisms in perception and cognition, potentially contributing to the development of advanced AI.

Coombs and Trestman’s embodied framework traces brain and cognition evolution through key interconnected traits in active, visually guided animals. Their analysis of 35 traits across 17 lineages reveals that advanced cognition in vertebrates, cephalopods, and euarthropods arose from synergistic upgrades in body, sensory, brain, and motor systems – not isolated changes. Central to their model are “pivotal traits,” like high-resolution vision and laminated visual brain regions, which enabled the shift from slow, reactive behaviors to fast, proactive 3D movement. These traits demanded greater neural computation, suggesting prediction as a core function. Even simpler lineages may share fundamental sensory-neural linkages, hinting at primitive predictive capacities. Thus, the evolution of complex cognition likely relied on integrated trait networks, with vision playing a decisive role in advancing PP.

Teufel and Fletcher (Reference Teufel and Fletcher2020) offer a critical framework for understanding predictive capabilities in early visual systems by distinguishing between two fundamental forms of neural prediction: bottom-up “constraints” and top-down “expectations.” This distinction challenges the traditional view of prediction as exclusively a top-down process originating from higher cortical areas. “Constraints,” according to their model, represent long-term predictions developed from learning stable, nonhierarchical environmental patterns over extended periods. These experiential priors become embedded within sensory systems’ tuning properties, fundamentally influencing how incoming sensory information is processed. These constraints shape initial context-independent sensory predictions based on environmental statistics encountered throughout an organism’s lifetime. In contrast, “expectations” operate as context-dependent, top-down predictions relying on hierarchical processing and memory, having a more temporary influence on sensory processing.

Teufel and Fletcher (Reference Teufel and Fletcher2020) argue that PP frameworks should consider both mechanisms as to how predictive information flows through top-down expectations and bottom-up constraints, suggesting that early sensory areas play a more fundamental role in anticipating and interpreting sensory input than traditionally recognized. The vertebrate retina’s complex layered organization enables extensive initial processing that extracts key features while adapting to light statistics, with retinal ganglion cells implementing predictive coding by signaling deviations from expectations based on spatial and temporal patterns (Johnston et al., Reference Johnston, Seibel, Darnet, Renninger, Orger and Lagnado2019). Similarly, invertebrate optic lobes employ predictive mechanisms such as specialized neurons that anticipate prey movement trajectories and matched filters tuned to ecologically relevant stimuli (Wiederman et al., Reference Wiederman, Fabian, Dunbier and O’Carroll2017). In both systems, lateral inhibition enhances contrast by filtering expected correlations (Huang & Protti, Reference Huang and Protti2016; Wiederman et al., Reference Wiederman, Fabian, Dunbier and O’Carroll2017), supporting the hypothesis that early visual systems across diverse animal lineages implement inherent bottom-up “constraints” based on stable environmental regularities – precisely the type of long-term predictions that Teufel and Fletcher describe.

By considering Teufel and Fletcher’s concept of “constraints,” Coombs and Trestman’s notion of “pivotal traits,” particularly high-resolution eyes, can be reconceptualized. Rather than viewing high-resolution eyes solely as advanced sensory input devices, they should be understood as systems fundamentally embedded with context-independent predictive mechanisms (“constraints”). The spatial and temporal information provided by high-resolution vision allows for the learning and encoding through the evolution of highly stable environmental regularities as inherent predictive biases. Over evolutionary timescales, natural selection would favor organisms with sensory systems that efficiently process and predict ecologically relevant information. This pressure could lead to encoding basic predictive mechanisms within the structure of early visual areas. Furthermore, during an individual’s lifetime, learning processes can refine these inherent predictive mechanisms, allowing for adaptation to the unique statistical environmental properties. The increased processing demands and computational power associated with “pivotal traits” like high-resolution eyes, as highlighted by Coombs and Trestman, are likely directly linked to encoding inherent PP mechanisms.

Integrating bottom-up predictive constraints, inspired by biological sensory systems, is key for advanced cognitive AI beyond current top-down models. While today’s AI often lacks these mechanisms for efficient environmental processing, new architectures incorporating features such as lateral inhibition (e.g., SpiLiFormer enhancing attention; Zheng et al., Reference Zheng, Huang, Yu, Zhu, Tang, Yu and Jin2025), dynamic retina-like coding (improving robustness; Zhang et al., Reference Zhang, Wang, Yao, Xu, Zhou, Xu and Li2024), and neuromorphic hardware (providing richer data; Zhu et al., Reference Zhu, Xie, Lv, Leng, Zhang, Xu and Han2024) show potential. These approaches promise breakthroughs in embodied AI and bio-inspired robotic vision (Choi et al., Reference Choi, Lee, Chang, Song and Kim2024).

The evolution of complex cognitive functions is strongly linked to the computational demands of high-resolution eyes. These demands are met through embedded, context-independent, bottom-up predictive mechanisms (“constraints,” per Teufel & Fletcher, Reference Teufel and Fletcher2020) that anticipate environmental regularities, creating the basis for higher cognition. To achieve advanced AI, embedding analogous predictive mechanisms in AI sensory systems is not merely potential but necessary. This integration allows AI models to effectively navigate novel situations and predict environmental statistics. Ultimately, incorporating bottom-up “constraints” is the pathway to building genuinely adaptable embodied systems exhibiting robust generalization and higher cognitive functions.