The challenge
The brain must continuously learn new information while keeping previously acquired memories stable. One proposed basis for storing information is the neuronal assembly, a group of strongly connected neurons that becomes active together. Yet it remains unclear how such assemblies can overlap, change with experience, and be combined across brain areas without interfering with one another. This is especially challenging because synaptic plasticity, the process that enables learning by changing connections between neurons, can also alter connections that encode earlier information. We therefore asked whether known properties of dendrites and inhibitory neurons could provide a biological mechanism that allows neural circuits to remain flexible enough to learn while protecting what has already been stored.
Our approach
We built a spiking network model in which excitatory neurons contained several nonlinear dendritic branches. Context-dependent inhibition controlled which dendrites were available for synaptic plasticity, while other branches were protected from change. We tested whether this mechanism could form and maintain neuronal assemblies, support their projection and association across several brain areas, and solve a visual-auditory association task with ambiguous stimuli.
Our findings
Context-dependent inhibition allowed plasticity to be switched on locally at individual dendrites. This enabled neurons to participate in different, overlapping assemblies through different dendritic branches without erasing previously learned connections. The resulting assemblies could be reliably recalled from partial input and combined through projections and associations across brain areas. In a visual-auditory task, context-specific assemblies in a downstream area successfully separated visually ambiguous stimuli that could not be distinguished from visual information alone.
The implications
Our model provides a biologically grounded candidate mechanism for how neural circuits could combine flexible learning with stable memory. Dendrite-specific control of plasticity may therefore be an important principle for complex brain computations and could also inspire new approaches to continual learning in artificial neural networks.
Creating SyNergies
The study was carried out by Sebastian Onasch, Christoph Miehl, M. Maurycy Miękus and our SyNergy member Julijana Gjorgjieva. The work connects computational neuroscience at TUM and the University of Chicago and brings together expertise in synaptic plasticity, dendritic computation and inhibitory circuits to link cellular mechanisms with computations spanning multiple brain areas.
