Investigating the potential for insect place learning using multi-scale CNNs

<p dir="ltr">Many insects, such as ants, are highly adept navigators, using learned visual information for route navigation and homing. They do this despite low resolution vision and small brains. Whether insects are also capable of more complex spatial cognition, such as pose independent place recognition, is an open question. In this study we first explored whether Convolutional Neural Networks (CNNs) of varying size are capable of a real world ‘ant’s eye’ place recognition task. We collected panoramic images from a set of 11 distinct places with variations in pose, time of day, weather and season. The CNNs were trained to categorise the places for input images of varying resolution. Whilst VGG16 performed best for all image resolutions, there was a general trend relating model size and image resolution to performance. Of the custom models, smaller models learn lower resolutions better than higher resolutions, and visa versa. We also found that resolutions 113 x 36 and 57 x 18 elicit the first or second best performance of the custom models, suggesting optimal performance for low computational processing lies between these two highlighted resolutions.</p>