Dendritic nonlinearities and synapse type-specific input clustering enable the development of input selectivity in a wide range of settings.

The importance of dendritic structures for neural computation has been recognized since the early days of neuroscience. However, it is only recently, with the increasing availability of data and advances in modelling techniques, that researchers have begun to map out the precise roles of different branching structures and synaptic distributions [1, 2].
Here, we investigate how the nonlinear integration of excitatory and inhibitory currents in dendritic branches can enhance the ability of a neuron to develop input selectivity and E/I balance via Hebbian synaptic plasticity. We model a single rate neuron that receives input from a population of E and I cells, divided into sub-populations, each tuned to a specific orientation (Fig 1.a). A simple Hebbian learning rule with synapse-type specific normalisation develops tight E/I balance and strong input selectivity in the postsynaptic neuron [3].
We demonstrate that a neuron whose inputs are randomly grouped into dendritic branches that are nonlinearly summed (Fig 1.b) significantly outperforms an equivalent neuron without dendritic nonlinearities in terms of the development of input selectivity in a wide range of settings, including increased noise levels, changes in the input statistics and different plasticity parameters (Fig 1.c, d). Using simulation-based inference [4], we identify a range of plausible dendritic nonlinearities that can consistently improve the performance of the synaptic plasticity (Fig 1.e).
We further investigate the impact that the distribution of excitatory and inhibitory synapses can have on the learning dynamics. We define a metric of dendritic specificity that controls how localised the input of each population is to a specific dendrite. We find that asymmetric specificity of E and I inputs, and in particular higher inhibitory than excitatory specificity (Fig 1.f), significantly enhances the ability of synaptic plasticity to produce input selectivity.
Our findings suggest that simple dendritic structures with nonlinearities can significantly enhance the ability of neurons to learn via synaptic plasticity. Moreover, the impact of different levels of E/I dendritic specificity on the learning dynamics indicates that phenomena such as synaptic clustering [5] and different connectivity patterns of various neuron types [6] may play an important role in driving local learning in cortical neurons.