The Costs and Benefits of Goal-Directed Attention in Deep Convolutional Neural Networks

摘要: People deploy top-down, goal-directed attention to accomplish tasks, such as finding lost keys. By tuning the visual system relevant information sources, object recognition can become more efficient (a benefit) and biased toward target potential cost). Motivated by selective in categorisation models, we developed a mechanism that process naturalistic (photographic) stimuli. Our be incorporated into any existing deep convolutional neural networks (DCNNs). The processing stages DCNNs have been related ventral stream. In light, our attentional incorporates top-down influences from prefrontal cortex (PFC) support behaviour. Akin how weights models warp representational spaces, introduce layer of mid-level DCNN amplify or attenuate activity further goal. We evaluated using photographic stimuli, varying target. found increasing has benefits (increasing hit rates) costs false alarm rates). At moderate level, improves sensitivity (i.e. increases $d^{\prime }$ ) at only increase bias for tasks involving standard images, blended images natural adversarial chosen fool DCNNs. These results suggest reconfigure general-purpose better suit current task goal, much like PFC modulates along addition being parsimonious brain consistent, approach performed than machine learning transfer learning, namely retraining final network accommodate new task.