Novelty detection is a core function of the nervous system that enables organisms to learn, explore, and behave adaptively in novel environments. Yet, how the brain computes novelty in naturalistic settings -- and how such computations can be formalized beyond simple task paradigms -- remains incompletely understood. In particular, sensory inputs in natural environments are high-dimensional and continuously varying, requiring generalization across stimuli to compute novelty efficiently. At the same time, novelty computation is intrinsically dependent on the temporal scale at which recent and distant experience are integrated. These computations are not static but evolve across cortical hierarchies, as neural representations become more abstract and integration timescales lengthen. How can the effects of stimulus generalization on novelty computation be formalized? What links exist between the structure of neural representation and generalization effects on novelty and exploration in naturalistic environments? And how do these effects interact with finite neural timescales for the integration of past observation?

To address these questions, I propose an integrative computational framework that formalizes novelty computation in naturalistic environments. This framework links neural responses, behavioral signatures and mechanistic neural implementations to shared computational principles of novelty detection in the brain, specifically accounting for: (i) generalization between familiar and novel stimuli based on underlying neural representations, and (ii) the interplay of generalization effects with gradual memory decay on short timescales. I empirically validate my theory across multiple task paradigms, leveraging it to quantify how stimulus similarities affect neural novelty responses and exploration in mice. Using theory-driven experiment design, I develop a new paradigm for systematically studying similarity effects on novelty computation. I show that my computational frameworks provides a unifying perspective on mechanistic implementations of novelty computation, and gives rise to a new mechanistic hypothesis for novelty detection in the brain. By reframing novelty as a computation defined by structured internal representations and processing timescales, this work enables systematic quantification of novelty in naturalistic environments and opens up new research avenues linking sensory perception with intrinsically motivated behavior.