Deep Problems with Neural Network Models of Human Vision

Abstract

Deep neural networks (DNNs) have had extraordinary successes in classifying photographic images of objects and are often described as the best models of biological vision. This conclusion is largely based on three sets of findings: (1) DNNs are more accurate than any other model in classifying images taken from various datasets, (2) DNNs do the best job in predicting the pattern of human errors in classifying objects taken from various behavioral benchmark datasets, and (3) DNNs do the best job in predicting brain signals in response to images taken from various brain benchmark datasets (e.g., single cell responses or fMRI data). However, most behavioral and brain benchmarks report the outcomes of observational experiments that do not manipulate any independent variables, and we show that the good prediction on these datasets may be mediated by DNNs that share little overlap with biological vision. More problematically, we show that DNNs account for almost no results from psychological research. This contradicts the common claim that DNNs are good, let alone the best, models of human object recognition. We argue that theorists interested in developing biologically plausible models of human vision need to direct their attention to explaining psychological findings. More generally, theorists need to build models that explain the results of experiments that manipulate independent variables designed to test hypotheses rather than compete on predicting observational data. We conclude by briefly summarizing various promising modelling approaches that focus on psychological data.

Obstacles to inferring mechanistic similarity using Representational Similarity Analysis

“…We have argued for this approach in relation to making inferences about mechanistic similarity between DNNs and humans [29]. In fact, research relating DNNs to human vision provides a striking case of a disconnect between RSA and behavioural findings from psychology [29–31]. The findings here may explain contradictory RSA scores between DNNs and human visual processing as pointed out by Xu and Vaziri-Pashkam [20].…”

“…However, another approach is more tractable: conduct controlled experiments to establish whether the two systems are representing information in similar ways. We have argued for this approach in relation to making inferences about mechanistic similarity between DNNs and humans [29]. In fact, research relating DNNs to human vision provides a striking case of a disconnect between RSA and behavioural findings from psychology [29–31].…”

Human shape representations are not an emergent property of learning to classify objects

“…These failures may reflect a range of processes present in humans but absent in CNNs trained to recognise objects through supervised learning, such as figure-ground segregation, completing objects behind occluders, encoding border ownership, and inferring 3D properties about the object [43]. Consistent with this hypothesis, Jacob et al [27] and Bowers et al [7] have recently highlighted a number of these failures in CNNs, including a failure to represent 3D structure, occlusion, and parts of objects. More broadly, these results challenge the the claim that CNNs trained to recognise objects through supervised learning are good models of the ventral visual stream of human vision (see, for example, [52, 8, 37]).…”

Should attention be all we need? The epistemic and ethical implications of unification in machine learning

“…But using ANNs in that way requires careful construction and comparison to ensure meaningful inferences can be drawn precisely because of these (and other) disanalogies baked into the technology. Failure to account for this can lead to misleading conclusions and faulty science [14]. On the other hand, nothing we have said means that ANNs cannot be brought further in line with biology to fruitful ends.…”

“…On a basic level, the tendency to ignore domain experts (Section 4.1) and the issues around path dependency (Section 4.2) may make ML systems less effective tools, which for systems that are so widely used has immediate social welfare implications. 14 And as we discussed in Section 4.4, the trend of increased black-boxing itself has distinct ethical implications related to subjects' rights to explanations.…”