For deep networks, the whole equals the sum of the parts

Abstract

Deep convolutional networks exceed humans in sensitivity to local image properties, but unlike biological vision systems, do not discover and encode abstract relations that capture important properties of objects and events in the world. Coupling network architectures with additional machinery for encoding abstract relations will make deep networks better models of human abilities and more versatile and capable artificial devices.

Bowers et al. raise questions about the validity of methods and types of evidence used to compare deep networks to human vision. Their discussion also draws attention to serious limitations of convolutional neural networks for understanding vision. Here we focus on two ideas. First, compelling evidence is emerging that deep networks do not capture pervasive and powerful aspects of visual perceptual capabilities in humans. These limitations appear to be fundamental and relate to the lack of mechanisms for extracting and encoding abstract relations. Second, the problem of mimicry, both in comparing network and human responses in behavioral tasks and in comparing model unit activations to brain data, highlights a general difficulty in using potentially superficial similarities across systems to draw deep conclusions. We conclude by suggesting that understanding the mechanisms of visual perception will likely require synergies between network processing and processes that accomplish symbolic encoding of abstract relations.

Abstract relations in perception

Human perception derives abstract, symbolic representations from relational information in sensory input (e.g., Baker & Kellman, 2018), enabling visual representations to be widely useful in thinking and learning (Kellman & Massey, 2013). Processes like those found in deep convolutional neural networks (DCNNs) may be an important part of human vision, but their anchor in concrete, pixel-level properties makes them unlikely to be sufficient. DCNNs differ from human perceivers profoundly, for example, in their access to shape information (Baker, Lu, Erlikhman, & Kellman, 2018, 2020; Geirhos et al., 2019; Malhotra, Dujmović, & Bowers, 2022). Whereas shape is the pre-eminent driver of human object recognition, when shape and texture conflict, networks classify by texture. Humans readily see shape in glass figurines, but networks consistently misclassify these (e.g., labeling a robin as a shower cap, a fox as a chain, and a polar bear as a can opener). Silhouettes do better, producing around 40% accuracy (Baker et al., 2018; Kubilius, Bracci, & de Beeck, 2016), but rearranging their parts, which severely impairs human classification, has strikingly little effect on network responses. Conversely, for correctly classified silhouettes, adding small serrations along the boundary reduces network classifications to chance or below, while human perceivers are unaffected. These and other results indicate that networks extract local shape features but have little or no access to global shape (Baker et al., 2018, 2020).

Recent research suggests that these findings regarding global shape reflect broader limitations in DCNNs' abilities to capture abstract relations from visual inputs. Baker, Garrigan, Phillips, and Kellman (2023) attempted to train DCNNs to capture several perceptual relations that human perceivers detect readily and generalize from even a small number of examples. These included the same/different relation (Puebla & Bowers, 2021a, 2021b), judging if a probe was inside or outside of a closed contour, and comparing the number of sides of two polygons. Using restricted and unrestricted transfer learning with networks previously trained for object classification, we found that networks could come to exceed chance performance on training sets. Subsequent testing with novel displays, however, showed that the relations per se were not learned at all. Although human perceivers rapidly acquired and accurately applied these relations to new displays, networks showed chance performance. The limitation of deep networks in representing and generalizing abstract relations appears to be fundamental and general (see also Malhotra, Dujmović, Hummel, & Bowers, 2021).

Methodological issues

The methodological issues raised by Bowers et al. are well-placed. Similarities between model responses and human judgments, and between model activations and brain activations, invite us to think that human processing may resemble deep networks. Yet claims based on both kinds of similarities may be tenuous. In our research, we have often seen deep networks produce somewhat better than chance responding on tests of relational processing, only to find that they were using some obscure, nonrelational property that correlated with the relevant relation.

Parallel concerns apply to similarities between activation patterns in brains and DCNN layers. Although intriguing, it is important to remember that we typically do not know what activations in either layers of a neural network or in the visual brain are signaling. Common representations in two systems might produce similar activation patterns or respectable correlations in representational similarity analyses (RSAs), but it is a case of affirming the consequent when we assume that high representational similarity scores imply common representations (Saxe, McClelland, & Ganguli, 2019). These issues may shed light on puzzling results. For example, Fan, Yamins, and Turk-Browne (2018) interpreted RSA results as suggesting that deep learning systems trained on photographs capture abstract representations such as humans use to see objects from line drawings. RSA was used to correlate similarity matrices obtained for photos and for line drawings; prior work used RSA to argue for quantitative similarities between advanced layers of the model and primate IT. In contrast, we tested classification of outline drawings by deep networks (VGG-19 and AlexNet) trained on ImageNet for object classification and found no evidence of successful classification based on outlines (Baker et al., 2018). For 78% of objects, networks would have done better by choosing an ImageNet category at random, and neither network produced a single correct first-choice classification. Do networks capture an abstract outline representation of objects, as suggested by RSA, yet fail to use it to classify inputs composed solely of outlines? As Bowers et al. suggest, the answer may lie in confounding in the stimulus properties that drive representational similarity (cf. Saxe et al., 2019).

Conclusion

A wealth of evidence suggests that biological vision systems extract and represent abstract relations. DCNNs far exceed humans in sensitivity to local image properties, but for humans, local sensory activations are transient, rapidly discarded, and used to discover and encode relations that capture important properties of objects and events in the world. Peering beyond observed similarities that Bowers et al. suggest may be superficial, these differences between networks and brains may be deep and fundamental. More work is needed to discern the sources of the differences. Combining network architectures with additional machinery for encoding abstract relations might make deep networks better models of human abilities and more versatile and capable artificial devices.