Yes, convolutional neural networks (CNNs) can have negative weights. During the training process, weights in CNNs are updated using backpropagation and gradient descent, and they can take on positive or negative values depending on how they minimize the loss function.
Negative weights are essential as they allow the network to learn features where suppression is necessary. For instance, a filter with negative weights might detect darker regions in an image or subtract certain features to highlight others.
These weights play a critical role in creating the high-dimensional representations that CNNs use for tasks like object detection and image classification. Without the possibility of negative weights, the network would lose a significant degree of flexibility in learning diverse patterns.