An activation is the output value a neuron in a neural network produces by applying an activation function to the weighted sum of its inputs. The set of activations across a layer represents how that layer encodes the input and serves as the input to the next layer.
Activation functions such as ReLU, sigmoid, and GELU apply a nonlinear transformation that allows neural networks to approximate complex functions. In interpretability research, the magnitude and pattern of a neuron's activations on particular inputs are analyzed to infer which features the neuron detects.