How the brain understands that there is something to learn

A paper by neurobiologists from Carnegie Mellon was published in Cell Reports, which explains one of the most mundane, yet mysterious facts about learning: why the brain “prints out” plasticity when a stimulus actually predicts something (a reward), and does not do so when there is no connection. The authors showed that during whisker learning in mice, somatostatin interneurons (SST) in the somatosensory cortex steadily weaken their inhibitory effect on pyramidal neurons in the superficial layers - and only if the stimulus is associated with a reward. If the stimulus and reward are separated in time (there is no contingency), inhibition does not change. Thus, the brain “understands” that there is something to learn, and locally transfers the network to a state of facilitated plasticity.

Background of the study

The brain does not learn continuously, but in “chunks”: windows of plasticity open when a new sensory signal actually predicts something—an outcome, a reward, an important consequence. In the cortex, this learning “faucet” is largely turned by the inhibitory network of interneurons. Its different classes perform different functions: PV cells quickly “squeeze” the discharge of pyramids, VIP cells often inhibit other inhibitory neurons, and SST interneurons target the distal dendrites of pyramids and thereby regulate which inputs (sensory, top-down, associative) even get a chance to get through and take hold. If the SSTs hold the “steering wheel” too tightly, the cortical maps are stable; if they let go, the network becomes more susceptible to restructuring.

Classical models of learning predict that contingency (a rigid stimulus→reward link) is the key to whether plasticity will kick in. Neuromodulators (acetylcholine, norepinephrine, dopamine) carry a “salience score” and a prediction error signal to the cortex, but they still need a local switch at the microcircuit level: who exactly and where in the cortex “releases the brake” so that the dendrites of pyramidal neurons can integrate useful combinations of inputs? Evidence from recent years has hinted that SST cells often take on this role, because they regulate the activity of branching dendrites - the place where context, attention, and the sensory trace itself are formed.

The mouse whisker sensorimotor system is a convenient platform for testing this: it is well mapped in layers, it is easy to associate with reinforcement, and plastic shifts in it are reliably detected by electrophysiology. It is known that when assimilating associations, the cortex switches from the “strict filtering” mode to the “selective depressurization” mode - dendritic excitability increases, synapses are strengthened, and recognition of subtle differences improves. But a critical question remained: why does this happen only when the stimulus actually predicts a reward, and which node in the microcircuit gives permission for such a switch.

The answer is important not only for basic neuroscience. In rehabilitation after a stroke, in auditory and visual training, in teaching skills, we intuitively build lessons around timely feedback and the "meaning" of actions. Understanding how exactly the SST circuit along the cortex layers opens (or does not open) a window of plasticity in the presence (or absence) of contingency brings us closer to targeted protocols: when it is worth strengthening disinhibition, and when, on the contrary, to maintain the stability of maps so as not to "shake up" the network.

How was this tested?

The researchers trained mice to make a sensory association of whisker touch → reward, and then recorded synaptic inhibition from SST interneurons to pyramidal cells in different layers in brain slices. This “bridge” between the behavioral task and cellular physiology allows us to separate the fact of learning from the background activity of the network. Key control groups received an “undocked” protocol (stimuli and rewards without connection): no weakening of SST inhibition occurred there, i.e. SST neurons are sensitive precisely to the stimulus-reward contingency. Additionally, the authors used chemogenetic suppression of SST outside the context of training and phenocopied the observed depression of outgoing SST contacts, a direct hint at the causal role of these cells in triggering the “plasticity window.”

Main results

• Spot "unblocking" from above: a long-term decrease in SST inhibition was detected in pyramidal neurons of the superficial layers, while no such effect was observed in the deep layers. This indicates layer- and target-specificity of disinhibition in the cortex.
• Contingency is decisive: when the stimulus and reward are “unlinked”, there are no plastic shifts - the network is not transferred to the “in vain” learning mode.
• Cause, not correlate: artificial reduction of SST activity outside of training reproduces the weakening of inhibitory outputs to the pyramids (phenocopy of the effect), indicating that SST neurons are sufficient to trigger disinhibition.

Why is this important?

In recent years, much has suggested that cortical plasticity often begins with a brief “depressurization” of inhibition – particularly via parvalbumin and somatostatin cells. The new work goes a step further: it shows a rule for triggering this depressurization. Not just any stimulus “releases the brakes,” but only those that make sense (predict a reward). This is economical: the brain does not rewrite synapses without reason, and preserves details where they are useful for behavior. For learning theories, this means that the SST circuit acts as a causal detector and a “gateway” for plasticity in the superficial layers where sensory and associative inputs converge.

What this tells practitioners (and what it doesn't)

- Education and rehabilitation:

• The "windows" of plasticity in sensory cortical maps appear to depend on the meaningfulness of the content - there needs to be an explicit stimulus→result connection, not just repetition.
• Trainings where the reward (or feedback) is time-linked to the stimulus/action are likely to be more effective in triggering changes.

- Neuromodulation and pharmacology:

• Targeting the SST circuit is a potential target for enhancing learning after stroke or in perceptual disorders; however, this is still a preclinical hypothesis.
• Importantly, the layer-specificity of the effect suggests that “broad” interventions (general stimulation/sedation) may blur beneficial changes.

How does this data fit into the field?

The work continues the team's line of research, where they previously described layer- and type-specific shifts in inhibition during learning and emphasized the special role of SST interneurons in tuning the inputs to pyramidal neurons. Here, a critical variable is added - contingency: the network "releases the brakes" only in the presence of a causal stimulus→reward connection. This helps to reconcile previous contradictions in the literature, where disinhibition was sometimes seen and sometimes not: the issue may not be in the method, but in whether there was something to learn.

Restrictions

This is mouse sensory cortex and sharp-slice electrophysiology; transfer to long-term declarative learning in humans requires caution. We see long-term (but not lifelong) depression of SST outputs; how long this persists in the living network and how exactly it relates to behavior beyond the whisker task is an open question. Finally, there are multiple classes of inhibitory neurons in the cortex; current work highlights SST, but the balance between classes (PV, VIP, etc.) under different types of learning remains to be described.

Where to go next (what is logical to check)

• Temporal "windows": width and dynamics of the SST-dependent "window of plasticity" at different learning rates and reinforcement types.
• Generalization to other modalities: visual/auditory cortex, motor learning, prefrontal decision-making circuits.
• Neuromarkers in humans: noninvasive indicators of disinhibition (e.g. TMS paradigms, MEG signatures) in tasks with overt and absent contingency.

Study source: Park E., Kuljis DA, Swindell RA, Ray A., Zhu M., Christian JA, Barth AL Somatostatin neurons detect stimulus-reward contingencies to reduce neocortical inhibition during learning. Cell Reports 44(5):115606. DOI: 10.1016/j.celrep.2025.115606