A fully connected, or dense, network connects every neuron in one layer to every neuron in the next. This gives it flexibility: in principle, it can approximate almost any function. But dense layers have no built-in assumptions about how inputs are structured, so patterns must be learned from data rather than supported by the architecture. section 3.4.1
That flexibility can be inefficient. Flattening an image obscures spatial proximity; representing a sentence as inputs loses ordering, and a time series loses temporal rhythm. Dense networks can learn such patterns, but may require many parameters and examples and still handle them less naturally than specialized architectures. Convolutional networks encode local spatial structure, while recurrent networks carry information through sequences. section 3.4.1 section 3.4.7


