Dendritic inhibition may protect overlapping memories from forgetting

By Marcus Bennett

A Neuron study published in August shows in simulation that context-dependent inhibition of individual dendritic compartments can let one neuron join multiple memory assemblies without erasing earlier associations.

The long-standing question is how the human brain retains older information while learning new information. The evidence here is computational rather than a patient recording or animal experiment, and experimental confirmation remains necessary.

For electrophysiology laboratories, the immediate value is experimental: recordings can test whether inhibitory subgroups, targeted dendritic branches and population activation change together with context. The paper predicts a coordinated circuit signature, not merely a sharper single-neuron response.

How does dendritic inhibition preserve overlapping memories?

The model implements several specific mechanisms:

• Every neuron in the simulated network connected to every other neuron.
• Each neuron contained a soma and several semi-independent dendritic compartments that could independently generate NMDA-receptor-driven spikes and strengthen their connections.
• Inhibitory neurons selectively silenced individual dendritic compartments, preventing plasticity there while leaving other compartments unchanged.
• Several context signals silenced subsets of inhibitory neurons, mimicking a new environment or internal state and reducing inhibition on selected compartments so learning could proceed.

Julijana Gjorgjieva, professor of computational neuroscience at the Technical University of Munich, said: “We have somas, dendrites, neurons that are inhibitory and excitatory.” She added that this makes the model “biologically plausible in the way that [artificial neural networks] are not.”

Her mechanism shifts the unit of gating from the whole neuron to the dendritic compartment. One compartment learns while others remain unchanged, allowing a neuron to join several assemblies without erasing what it learned previously.

How does partial reactivation behave across contexts?

Groups of excitatory neurons activated the network in sequence, mimicking an animal viewing images. Each group formed a different assembly, analogous to forming a memory; the same group in another context formed a separate assembly.

Reactivating only part of an excitatory group, as after seeing part of an image, recruited the full assembly only when the context matched the original learning. In a different context, partial activation did not restore that assembly.

This context gate is the model’s proposed answer to catastrophic forgetting, the loss of older information when artificial networks learn new information. It preserves overlap by separating assemblies according to context rather than forcing one representation to replace another.

How do separate brain areas combine an assembly?

The researchers represented multiple brain areas as recurrently connected networks linked by feedforward connections. The model could copy an assembly from one area to another. When two upstream areas projected to a shared downstream area under the same context, that area combined their patterns into one joint representation.

That process resembles binding information across modalities, such as combining an image and sound into a single concept. The team found two routes for forming the association:

• Simultaneous activation: both areas fired at once and formed a new assembly downstream.
• Sequential activation: a new association attached to an assembly that already existed.

Christoph Miehl, a postdoctoral fellow in Tim Vogels’s lab at the Institute of Science and Technology Austria, said single neurons matter for larger computations such as cross-area information combination. He described the work as a bridge between those two scales.

Can the model separate visually similar characters?

In a simulated task, the model distinguished visually similar but phonetically distinct characters, including “0” and “O.” Pairs that looked alike activated largely overlapping assemblies in the simulated visual area.

In the downstream “concept” area, which received visual and auditory inputs, those stimuli occupied nonoverlapping assemblies. The concept area therefore separated them even when it saw only the image, because the context supplied the relevant information.

What can experiments test next?

Sebastian Onasch, a former graduate student in Gjorgjieva’s lab, identified a simplifying assumption: in a specific context, the inhibitory population connects to only one dendrite of each neuron. Real inhibitory connectivity is likely more complicated.

Gjorgjieva said that assumption need not hold for the model to remain informative. Without context-dependent inhibition, however, the model loses some ability to segregate stimuli.

Attila Losonczy, a neuroscience professor at UT Southwestern Medical Center who was not involved in the work, said the model makes three predictions:

• Distinct subgroups of an inhibitory cell class should respond to distinct contextual signals.
• Those subgroups should preferentially target specific dendritic branches.
• Dendritic activation patterns should shift systematically with context.

His recent findings support the premise that different dendritic regions can compute differently. Distal and proximal dendrites can decouple, while distal dendrites encode contextual information.

A Science Advances paper published the month before described tiny constrictions in hippocampal and cortical neurons that partition dendrites into distinct electrical compartments. Gjorgjieva said these constrictions could provide the physical basis for dendritic compartmentalization.

The constrictions cannot, by themselves, explain context-dependent assembly formation, which requires inhibition and disinhibition of specific compartments. She described them as anatomical support for the broader idea that dendrites contain local computational and plasticity compartments.

What remains unknown about context?

Christos Papadimitriou, professor of computer science at Columbia University and not involved in the study, called the work a major advance toward linking individual-neuron spiking to complex behavior. He also noted that assemblies, rather than single neurons, increasingly serve as the brain’s basic computational units.

Papadimitriou had previously proposed that assemblies could, in principle, perform any computation by copying themselves or merging with others. He judged this paper biologically realized that earlier inspiration, while praising its thoroughness, imagination and his own hope that someone would do the work.

What “context” means biologically, and where it comes from, remains the largest unresolved question, Miehl said. The authors propose moving beyond the simple visual-auditory task toward a reward-based behavioral task that a mouse might solve.

Losonczy said the study assigns dendritic inhibition a novel function. “It provides a really strong testable hypothesis of what comes next,” he said.

The proposed next step is a reward-based mouse task in which context-dependent dendritic inhibition can be tested against the model’s assembly-segregation predictions.