Using ant biology and natural environments to enhance models of vision and robot navigation

点击翻译按钮获取中文摘要

Over the past decade, Deep neural networks (DNNs) have revolutionised AI, achieving highly impressive perceptual feats especially in the domain of vision. In particular, convolutional neural networks have surpassed humans in their ability to classify naturalistic images. These findings, together with their similarities in architecture and activity patterns, have sparked major renewed interest in DNNs as models of the brain. However, these models diverge from the perceptual abilities of biological vision in important and unexpected ways, e.g., through their weakness to noise, susceptibility to adversarial attacks and inability to recognise stylised representations. This has been shown to be due to their over-reliance on diagnostic but brittle features such as texture [1] or even a single pixel correlated with the image class [2]. In part, these limitations arise from a lack of inductive biases in the models to regularise their learning, coupled with unrealistic training environments. Some of these shortcomings have recently been addressed through incorporating more of the properties of the primate visual system into these models [3] or by using more extensive and naturalistic training regimes [4]. The aim of this project is to leverage our knowledge of the brain and foraging behaviour of ants to see if both incorporating biological constraints and naturalistic training data can generate DNN's whose perceptual abilities have the robustness and generalisability of biological vision.