In classical models of sensory processing, the brain is often viewed as a static pipeline: sensors collect data, relays pass it forward, and executive centers decide on an action. However, biological systems do not treat the world as a constant stream of objective facts. Instead, they employ state-dependent sensory gating, a mechanism where internal physiological needs—such as hunger, thirst, or arousal—actively tune the sensitivity and logic of neural circuits.
By integrating internal states directly into sensory and decision-making pathways, organisms can dynamically reconfigure their behavior without rewiring their brains.
Internal States as Sensory Rheostats
Rather than acting as simple on-off switches, internal states function as "rheostats" or gain control mechanisms. Beer et al. and Chiel proposed that internal states function as regulatory gates that determine the gain of sensory-to-motor pathways. This allows an agent to ignore stimuli that are irrelevant to its current needs while remaining hyper-sensitive to them when internal conditions change.
This gating occurs even at the earliest stages of perception. For instance, Li et al. found that physiological hunger enhances both the initial detection and sustained cognitive processing of food-related visual stimuli. Similarly, Fujii demonstrated that the sensitivity of a cricket's antennae is regulated by internal neurochemistry, adjusting behavioral activation thresholds based on social density.
Even motor states act as gates. Dombeck et al. showed that the cerebellum suppresses sensory-evoked responses during locomotion, shifting the neural state from environmental monitoring to active motor execution. Ando et al. observed a similar phenomenon in insects, where the visual threshold for triggering collision avoidance is significantly higher when the animal is already moving compared to when it is stationary.
Reconfiguring Circuit Logic
State-dependency goes beyond simple volume control; it can fundamentally change the "valence" or meaning of a stimulus. Vogt et al. argue that early sensory processing stages are not static relays but dynamic decision-making centers. In Drosophila larvae, hunger can switch the valence of a stimulus from aversion to attraction, allowing the organism to reconfigure its behavioral response to the same environmental cue depending on metabolic needs.
This reconfiguration often happens downstream of the sensors. de Tredern et al. found that hunger biases sensorimotor decisions by modulating central processing circuits via neuropeptide signaling, rather than changing the primary sensory neurons themselves. This allows the animal to maintain a consistent perception of the environment while dynamically altering its motor strategy—such as prioritizing foraging over predator avoidance.
The Mathematics of Motivated Learning
Standard Reinforcement Learning (RL) often struggles to explain why an animal's preference for a reward can change instantly without new learning. To address this, researchers are redefining the fundamental equations of motivation.
- Utility-Based Prediction Errors: van Swieten proposed redefining the reward prediction error (RPE) as the difference between utility and expected utility, where utility is a function of both objective reward and the agent's current physiological state.
- Homeostatic Drive: Yoshida et al. and Gutkin introduced Homeostatically Regulated Reinforcement Learning (HRRL), where "reward" is defined as the reduction of internal "drive"—the deviation of a physiological state from an optimal setpoint.
- Multi-Attribute Salience: Smith et al. and Read developed the Multi-Attribute Incentive Salience (MAIS) model. They argue that internal states act as gain factors that scale specific sensory attributes of a stimulus. If you are salt-deprived, the "saltiness" attribute of a cue is multiplied by a high gain, immediately increasing the cue's motivational value.
These models explain instant revaluation: the ability of an organism to update the incentive salience of a cue the moment its internal state changes, a feat that traditional "model-free" RL cannot achieve without new experience (Berridge; Dayan).
Neural Implementation
Where does this integration happen? Evidence points to several key hubs:
- Basal Ganglia: Rossi et al. demonstrated that satiety bidirectionally modulates neurons in the substantia nigra, suggesting the basal ganglia scales its output based on subjective motivation rather than just objective reward probability.
- Thalamus: Machen et al. identified D2+ neurons in the paraventricular nucleus of the thalamus that scale their activity with hunger, acting as a "prediction engine" that links internal needs to environmental cues.
- Prefrontal Cortex: Ehret et al. found that the medial prefrontal cortex can dynamically suppress sensory representations once a goal-directed action is initiated, prioritizing task execution over continuous monitoring.
By treating internal states as active participants in sensory processing, these researchers have shown that "intelligence" is not just the ability to process the world, but the ability to filter and revalue the world to serve the body's changing needs.
Key Foundational Work
Neural Implementation of Motivated Behavior: Feeding in an Artificial Insect (1989) — contributed 2 concept elements to this direction.
Active Semantic Networks In Robots With Independant Control (1975) — contributed 2 concept elements to this direction.