Dissonance Between Human and Machine Understanding

Abstract

Complex machine learning models are deployed in several critical domains including healthcare and autonomous vehicles nowadays, albeit as functional blackboxes. Consequently, there has been a recent surge in interpreting decisions of such complex models in order to explain their actions to humans. Models which correspond to human interpretation of a task are more desirable in certain contexts and can help attribute liability, build trust, expose biases and in turn build better models. It is therefore crucial to understand how and which models conform to human understanding of tasks. In this paper we present a large-scale crowdsourcing study that reveals and quantifies the dissonance between human and machine understanding, through the lens of an image classification task. In particular, we seek to answer the following questions: Which (well performing) complex ML models are closer to humans in their use of features to make accurate predictions? How does task difficulty affect the feature selection capability of machines in comparison to humans? Are humans consistently better at selecting features that make image recognition more accurate? Our findings have important implications on human-machine collaboration, considering that a long term goal in the field of artificial intelligence is to make machines capable of learning and reasoning like humans.

Laconic Image Classification: Human vs. Machine Performance

“…Another limitation is that for computing human MEPIs in the case of Crop and Combined, the void image begins with a central pixel, which may be far from the relevant region for classification. Refinements of the human MEPI framework would be interesting to explore in the future, for example to replace Crop with a more higher-level component-based analysis [31].…”

“…It would further be interesting to explore how models trained in such a manner perform in more traditional classification metrics -error rates, precision, etc. -on the original input images, as well as whether or not they might help to address the observed lack of robustness of state-of-the-art DNNs in the presence of noisy [2,3,14,22] or incomplete information [12,19,24,27,28,31], or their lack of generalisation [7], or their bias towards texture [6]. It is important to note that the entropy of an image may increase as certain distortions of practical interest are intensified: more suitable general measures of robustness in such settings are left to be explored.…”

“…They showed that fragile recognition images are abundant and can occur at different sizes. Zhang et al [31] compare human and machine performance for image classification over segments of an image based on object boundaries (rather than rectangular regions). Both human and machine classifiers are used to identify key segments for classification, where, interestingly, humans are found to be better at classifying images using segments selected by machine models versus those selected by humans.…”

“…As discussed by Russakovsky et al [22], the 1,000 detailed classes of ILSVRC (e.g., coucal, sealyham terrier) require expert training for humans to perform adequately at classification, not only to visually distinguish the objects, but also to retrieve their label from the 1,000 options. Thus, along similar lines to experiments conducted by Geirhos et al [7] and Zhang et al [31], we frame the task for the following twenty high-level classes: bear, bird, cat, dog, fish, flower, fox, fruit, fungus, hippopotamus, insect, lion, monkey, reptile, shark, spider, tiger, vegetable, vehicle, wolf. We select these classes as they should be generally recognisable to humans without prior training; furthermore, we select mostly plants and animals to provide a more challenging classification task, with visually similar classes, such as lion/tiger, fruit/vegetable, dog/wolf, insect/spider, etc., providing non-trivial cases to visually distinguish.…”

“…Works on adversarial examples [1,20,26], for instance, establish that human and machine perception diverges greatly for specifically constructed images. Other works have presented bespoke experiments comparing human and machine performance beyond classification errors, presenting evidence for a lack of robustness in the presence of noisy [2,3,22] or incomplete information [12,19,24,27,28,31], a sensitivity to spatial [4,5,10,30] or colour [13,14] transformations, a lack of generalisation [7], a bias towards texture [6], etc., in the machine classifiers studied. By transforming test images prior to classification, these works provide insights into the differing types of information that humans and machines rely on for image classification.…”