Inductive bias is a concept in deep learning theory which specifies how a particular neural network expects data to behave. Importantly, this is not learnt from the training data but rather baked into the neural network architecture. In other words, it is the model's built in assumptions about the world in the form of expectations about the data.

Examples of inductive biases,

CNNs assume ==locality and spatial/translational invariance== ie, pixels closer to each other are more related and if an object occurs in one part of the image then it need NOT have to keep occurring there for the CNN to classify it.

CNNs have low relational inductive bias. This means that while a CNN can identify that there's a dog and a ball in the image, it cannot understand the relationship/iteraction between them such as the dog is chasing the ball.