While reviewing a convolutional detector, an engineer explains that a neuron in a deep layer reacts to a wide area of the original image. What is this area called?
Show answer and explanation
Answer: B
A neuron's receptive field is the set of input positions that can change its activation. A single 3x3 convolution sees only a 3x3 patch, but stacked convolutions compose: each additional layer widens the footprint, and strides or pooling widen it faster. That is why deep stacks capture global context while early layers respond to edges and textures. Designs such as dilated convolutions or skip connections deliberately enlarge the receptive field without shrinking the feature map, which matters for segmentation and detection.
Why the other options are wrong
- A. The learning rate is a step-size hyperparameter for the optimiser and is unrelated to which input pixels influence a unit.
- C. Filter count sets how many feature maps a layer produces, that is channel depth, not the spatial extent of the input each unit sees.
- D. Batch size controls how many samples are averaged per update and does not change a unit's spatial footprint.