In Systems Neuroscience, single neuron models can take many forms that can answer specific questions or fulfil a targeted purpose. However, model building is an underconstrained process that yields models that are often highly abstracted away from empirical reality. The point neuron assumption, where dendritic spatiotemporal properties and nonlinearities are assumed to be irrelevant, is a common simplification in neuronal and multineuronal models. Though these models can be useful for answering many questions, their use may bias investigations of higher-level function that pertains to how the nervous system computes. Constraining neuron model development with neural data has been historically challenging due to technological limitations and limited knowledge of what a priori assumptions to make. However, investigating how qualitative empirical properties impact a neuron model’s ability to perform a normative goal has the potential to inform and constrain single neuron model selection. Using a supervised learning framework, we investigate which qualitative dendritic properties in single neuron models have the most impact on the normative goals of learning and computation at the single-neuron level. We find that the dendritic properties introduced could either limit and enable performance on a set of computer vision tasks that require complex computation. We observed that a synergy of all dendritic properties enabled performance above that of models with fewer or isolated properties. From this we conclude that single neurons may be enabled by their dendritic properties to learn and compute complex functions. This framework can be further used to identify which empirical properties are most relevant to the normative goal of complex computation. This also suggests that multineuronal models that are used to investigate neural computation should re-evaluate use of the point-neuron abstraction. This dissertation emphasizes the need to connect normative goals to implementable empirical details in neuron modeling.