Groundbreaking Study Uncovers Surprising Connection Between Human and Machine Vision
Recent research conducted by Google DeepMind has revealed a fascinating intersection between human and machine vision, shedding light on their shared susceptibility to adversarial images. Adversarial images are digitally altered images designed to deceive AI models and lead to misclassifications. For instance, an AI might mistake a vase for a cat when presented with these manipulations.
Published in Nature Communications, the study titled Subtle adversarial image manipulations influence both human and machine perception delves into the impact of adversarial images on human perception through a series of experiments. These experiments demonstrate that while adversarial perturbations significantly mislead machines, they can also subtly influence human perception. Remarkably, the effect on human decision-making aligns with the misclassifications made by AI models, albeit to a lesser extent. This discovery emphasizes the intricate relationship between human and machine vision, highlighting how both can be influenced by minor perturbations in an image, even when those perturbations are minor and the viewing times are extended.
DeepMind’s research further explores the properties of artificial neural network (ANN) models that contribute to this susceptibility. The study focuses on two ANN architectures: convolutional networks and self-attention architectures. Convolutional networks, inspired by the primate visual system, apply local filters across the visual field, creating a hierarchical representation. In contrast, self-attention architectures, initially designed for natural language processing, use nonlocal operations to facilitate global communication across the entire image space. These architectures display a stronger preference for shape features over texture features. Interestingly, adversarial images generated by self-attention models exerted a stronger influence on human choices compared to those generated by convolutional models, indicating a closer alignment with human visual perception.
This research underscores the crucial role of subtle, higher-order statistics of natural images in aligning human and machine perception. Humans and machines alike are sensitive to these intricate statistical structures present in images. This alignment suggests a potential avenue for enhancing the robustness of ANN models, making them less susceptible to adversarial attacks. Moreover, it underscores the need for continued research into the shared sensitivities between human and machine vision, which could offer valuable insights into the mechanisms and theories behind the human visual system. The discovery of these shared sensitivities between humans and machines carries significant implications for AI safety and security, suggesting that adversarial perturbations could be leveraged in real-world scenarios to subtly bias human perception and decision-making.
In conclusion, this research represents a significant leap forward in understanding the complex relationship between human and machine perception, shedding light on both the similarities and differences in their responses to adversarial images. It highlights the importance of continued research in the realms of AI safety and security, particularly in comprehending and mitigating the potential impacts of adversarial attacks on AI systems and human perception.
