Rosenholtz argues that the field of attention is in crisis and calls for a Kuhnian revolution to transform current thinking. She claims that many experimental results, usually thought to address attention, in fact come from limits in peripheral visual processing. In particular, Rosenholtz argues that the TTM (Balas, Nakano & Rosenholtz, Reference Balas, Nakano and Rosenholtz2009), which employs a single-stage computation of summary statistics over local image patches, can explain many phenomena traditionally attributed to attention. She suggests that “one might reasonably call summary statistic encoding a paradigm shift, both for the science of visual attention and for vision more broadly.”

Here, we do not debate whether there is a crisis in attention research. Rather, we review a series of experimental and computational results that directly contradict the TTM as a model of peripheral vision, and argue that the TTM is unlikely to be a good model of vision in general. While the TTM accounts for a range of psychophysical phenomena, it is far from a universal model of peripheral processing. In particular, the TTM cannot reproduce large sets of experimental results in crowding, for which it was specifically developed.

In crowding, the perception of a visual target deteriorates in the presence of flankers. In the TTM, crowding occurs when the target and flankers are combined in the same summary statistics, leading to a degraded representation of the target. A rich body of experimental and theoretical studies showed that this account is fundamentally limited because the global configuration of visual elements determines how local elements are processed (Manassi, Sayim & Herzog, Reference Manassi, Sayim and Herzog2012, Reference Manassi, Sayim and Herzog2013, Manassi et al., Reference Manassi, Lonchampt, Clarke and Herzog2016). These effects of global configuration on visual processing are large, qualitative phenomena, and occur in diverse conditions, across multiple levels of processing (Manassi & Whitney, Reference Manassi and Whitney2018), and even in other sensory modalities (Oberfeld & Stahn, Reference Oberfeld and Stahn2012; Overvliet & Sayim, Reference Overvliet and Sayim2016). Hence, they reflect a general strategy for perceptual computations adopted by the brain. We propose that the TTM is unable to account for these crucial global effects for four main reasons:

1. 1. Locality. The TTM assumes that peripheral vision is fundamentally constrained by local summary statistics computed over small patches of the visual field. Several studies showed that models based solely on local interference mechanisms, including the TTM, cannot explain global effects in crowding (Bornet, Doerig, et al., Reference Bornet, Choung, Doerig, Whitney, Herzog and Manassi2021a; Choung et al., Reference Choung, Bornet, Doerig and Herzog2021; Doerig, Bornet, et al., Reference Doerig, Bornet, Choung and Herzog2020; Doerig et al., Reference Doerig, Bornet, Rosenholtz, Francis, Clarke and Herzog2019).

2. 2. Single stage. Interference between visual elements in TTM occurs as a single stage, that is, summary statistics within local patches. However, large-scale model comparisons including the TTM found that only models incorporating, at least, a two-stage grouping and segmentation process can accurately predict global crowding effects (Bornet, Doerig, et al., Reference Bornet, Doerig, Herzog, Francis and Van der Burg2021b; Doerig et al., Reference Doerig, Bornet, Rosenholtz, Francis, Clarke and Herzog2019; Doerig, Schmittwilken, et al., Reference Doerig, Schmittwilken, Sayim, Manassi and Herzog2020).

3. 3. Feedforward. The TTM is feedforward and static in nature, with a single pass of processing. Experimental and modeling studies showed how recurrent interactions can capture dynamics of global crowding effects (Doerig, Schmittwilken, et al., Reference Doerig, Bornet, Choung and Herzog2020; Morea et al., Reference Morea, Herzog, Francis and Manassi2024). The TTM’s inability to integrate recurrent processing prevents it from capturing the dynamical processes in peripheral vision.

4. 4. Low-level interactions. The TTM only models interference between low-level features via summary statistics. However, Manassi & Whitney, Reference Manassi and Whitney2018, demonstrated that crowding occurs not only between low-level features but also between high-level features such as objects or faces. The TTM does not reproduce these high-level face crowding effects (Bornet, Choung, et al., Reference Choung, Bornet, Doerig and Herzog2021).

In reply to these concerns, Rosenholtz, Yu & Keshvari, (Reference Rosenholtz, Yu and Keshvari2019) argued on theoretical grounds that global effects in crowding might be explained by post-perceptual processes. However, Bornet, Choung, et al. (Reference Choung, Bornet, Doerig and Herzog2021) provided direct and extensive empirical evidence that the post-perceptual processes proposed by Rosenholtz fail to account for global effects in crowding. In a similar vein, Rosenholtz et al. (Reference Rosenholtz, Yu and Keshvari2019) theoretically reasoned that high-level interactions in crowding may emerge from the complexity and high-dimensional nature of summary statistics computed on low-level visual features. However, Bornet, Choung, et al. (Reference Choung, Bornet, Doerig and Herzog2021) showed that the crowding strength predicted by TTM correlates strongly with low-level pixel density rather than high-level interactions.

In addition to being falsified in psychophysical settings, there are also strong reasons to doubt that the TTM could be up to the task of modeling visual processing more generally. As discussed, the TTM relies on a single-stage computational process that extracts summary statistics from local low-level visual features, plus an influence of task readout. However, decades of research in neuroscience suggest that human vision depends on multiple stages of nonlinear processing, including hierarchical feature extraction, recurrent interactions, and top-down feedback. These multi-stage mechanisms are critical for extracting behaviorally relevant representations (DiCarlo, Zoccolan & Rust, Reference DiCarlo, Zoccolan and Rust2012). By providing image-computable models of such multi-stage hierarchical processes, Artificial Neural Networks have revolutionized our ability to accurately model neural activity and visual behavior (review: Doerig et al., Reference Doerig, Sommers, Seeliger, Richards, Ismael, Lindsay, Kording, Konkle, van Gerven, Kriegeskorte and Kietzmann2023).

Computational models should be as simple as possible, but not simpler. Simplicity avoids overfitting and allows for easy interpretation and understanding. But overly simple models miss crucial aspects and lead to poor predictive power. Therefore, rejecting more complex models in favor of the TTM to account for the richness and adaptability of peripheral vision must be supported by exceptionally strong and extensive experimental evidence, which does not exist to date. More complex, multi-stage models with recurrent grouping and segmentation mechanisms are required. The importance of multiple stages in vision is also backed up by decades of research in neuroscience and machine learning. Under this light, the TTM and other models based on simple combinations of low-level features seem unlikely to bring about a Kuhnian revolution to replace attention in our understanding of visual processing. As it stands, both attention and peripheral vision research are unsettled fields. For this reason, replacing one with the other is premature.