Neural Responses to Unexpected Stimulus Repetitions and Omissions in Auditory Cortex Provide Mixed Evidence for Predictive Coding

This study of macaque primary auditory cortex reveals mixed evidence for predictive coding, showing that while neurons exhibit enhanced responses to unexpected stimulus omissions, they fail to show the predicted enhancement for unexpected stimulus repetitions.

The Gist

Imagine your brain is a highly sophisticated weather forecaster. Every second, it tries to predict what's going to happen next based on what it has experienced before. If the sun is shining, it predicts sunshine. If you hear a dog bark, it predicts a dog is nearby.

According to a famous theory called Predictive Coding, the brain doesn't just passively listen to the world; it actively guesses what's coming. When the world surprises the brain (like a sudden thunderclap on a sunny day), the brain is supposed to scream, "Hey! That wasn't supposed to happen!" This "scream" is a spike in neural activity, signaling an error in the prediction.

The Experiment: The Monkey's Ear
In this study, researchers (Shukla et al.) wanted to see if the primary auditory cortex (the brain's "first stop" for hearing, or A1) acts like this smart forecaster. They used two monkeys and played them sounds in a very specific way to test their predictions.

Think of the experiment as a game of "What's Next?" played with musical notes.

The Two Games They Played

Game 1: The "Repeat" Surprise (Pattern Paradigm)

• The Setup: The monkeys heard a rhythm: Note A, Note B, Note A, Note B...
• The Expectation: The brain gets used to this pattern. It expects Note A to be followed by Note B.
• The Surprise: Suddenly, the brain hears Note A again (a repeat). It's a violation of the pattern.
• The Theory's Prediction: If the brain is a perfect predictive coder, it should light up like a Christmas tree when it hears the repeat, shouting, "Wait, that's wrong!"
• What Actually Happened: The brain's "first stop" (A1) was surprisingly bored. It didn't react much differently to the surprise repeat than it did to the expected note. It was like a weather forecaster who sees a tornado but just shrugs and says, "Meh, seen it before."

Game 2: The "Missing" Surprise (Omission Paradigm)

• The Setup: The monkeys heard a steady rhythm: Note A, Note A, Note A, Note A...
• The Expectation: The brain expects another Note A to follow.
• The Surprise: The sound stops. There is silence where a note should be.
• The Theory's Prediction: The brain should scream, "Where did the note go?!"
• What Actually Happened: This time, the brain did react! When the sound was missing, the neurons fired up more than when the sound was expected. It was like the forecaster suddenly shouting, "The sky is empty! Something is wrong!"

The Big Takeaway: A Mixed Bag

The results are a bit of a "mixed bag," which is why the title says "Mixed Evidence."

1. The Bad News for the Theory: When the brain heard a sound it didn't expect (a repeat in a changing pattern), it didn't act like a predictive coder. It didn't get excited. This suggests that the "first stop" of hearing might not be doing the heavy lifting of predicting the future. Maybe it's just a simple recorder that gets tired of hearing the same thing (a phenomenon called "adaptation") rather than a smart guesser.
2. The Good News for the Theory: When the brain heard nothing (an omission), it definitely reacted. This supports the idea that the brain is making predictions about the future. When the prediction fails (the sound is missing), the brain notices.

The Analogy: The Jukebox vs. The DJ

Imagine the brain's auditory system is a Jukebox (the first stop) and a DJ (higher brain areas).

• The Jukebox (A1): When you ask for a song and it plays, the Jukebox just plays it. If you ask for the same song twice in a row, the Jukebox doesn't get excited; it just plays it again. If you ask for a song and nothing plays, the Jukebox might make a weird noise because the mechanism is confused. This is what the researchers saw: the Jukebox didn't care about the pattern repeats, but it did react when the music stopped.
• The DJ (Higher Cortex): The DJ is the one who knows the playlist, predicts what song comes next, and gets hyped when the crowd is surprised. The researchers suspect that while the Jukebox (A1) isn't doing much predicting, the DJ (higher brain areas) is probably doing all the heavy lifting for the "predictive coding" theory.

Why Does This Matter?

This study is important because it tells us that predictive coding isn't happening everywhere in the brain at the same level.

It suggests that the very first place sound enters the brain is more of a "local mechanic" dealing with immediate sounds, rather than a "grand strategist" predicting the future. The "smart" prediction work might happen later in the brain's processing chain.

In short: The brain's hearing center is great at noticing when a song stops unexpectedly, but it's not very good at getting excited when a song repeats unexpectedly. This helps scientists understand exactly where and how our brains build our perception of reality.

Technical Summary

1. Problem Statement

The study addresses a central debate in systems neuroscience regarding the implementation of Predictive Coding (PC) in primary sensory cortices.

• The Theory: PC posits that the brain functions as a Bayesian inference engine, where perception is generated by comparing bottom-up sensory input against top-down predictions. Mismatches generate "prediction errors" (surprise), which update internal models.
• The Conflict: While PC is a leading theory, it is unclear if primary auditory cortex (A1) implements true prediction-error signaling or if observed "surprise" responses are merely artifacts of local bottom-up mechanisms like Stimulus-Specific Adaptation (SSA), forward suppression, or "off" responses.
• The Gap: Previous studies using the classic "oddball" paradigm cannot distinguish between PC (prediction error) and SSA (reduced response to repeated stimuli). This study aims to disentangle these mechanisms by testing specific violations of expectation that SSA cannot explain.

2. Methodology

The researchers conducted electrophysiological recordings in the primary auditory cortex (A1) of two awake, behaving adult male rhesus macaques (Macaca mulatta).

Subjects and Recording:

• Subjects: Monkey C (14 years) and Monkey W (14 years).
• Data: 198 single units from Monkey C and 229 single units from Monkey W across 113 recording sessions.
• Localization: A1 was identified using Current-Source Density (CSD) profiles (showing characteristic laminar sinks in granular layers) and tonotopic mapping.
• Stimuli: Tone bursts centered on each neuron's Best Frequency (BF), determined via Spectrotemporal Receptive Fields (STRFs) derived from Dynamic Moving Ripple (DMR) noise.

Experimental Paradigms:
To isolate PC from SSA, three paradigms were employed:

1. Oddball Paradigm (Control): Standard vs. Deviant tones. (Expected to show SSA/PC effects but not diagnostic between them).
2. Pattern Paradigm (The Critical Test for Repetition):
   • Standard: Alternating tone pairs (A-B-A-B), creating an expectation for alternation.
   • Violation: Unexpected repetition (A-A), violating the alternation expectation.
   • PC Prediction: Enhanced response to the unexpected repetition (Prediction Error).
   • SSA Prediction: Reduced or unchanged response (due to adaptation to the repeated frequency).
3. Omission Paradigm (The Critical Test for Absence):
   • Standard: Repetitive tone pairs where the second tone is always omitted (A-0-A-0), creating an expectation for omission.
   • Violation: Unexpected presence of the second tone (A-A), or conversely, an unexpected omission in a sequence where repetition was expected.
   • PC Prediction: Enhanced response to the unexpected omission (Prediction Error).
   • SSA Constraint: SSA relies on physical stimulus repetition; it cannot easily explain enhanced responses to the absence of a stimulus.

Analysis Metrics:

• Response Index (RESI): Quantified enhancement vs. suppression in tone pairs.
• Prediction-Error Index (PEI): Calculated as the difference between the response to the violation and the standard condition.
  • $PEI > 0$: Enhanced response to violation (Supports PC).
  • $PEI \leq 0$: Reduced/No response to violation (Inconsistent with PC).

3. Key Contributions

• Methodological Rigor: The study moves beyond the ambiguous oddball paradigm by utilizing pattern and omission paradigms specifically designed to dissociate top-down prediction errors from bottom-up adaptation.
• Laminar Specificity: Data was restricted to neurons in A1 with confirmed laminar profiles (granular layer sinks), ensuring the findings are specific to primary sensory processing rather than higher-order association areas.
• Dissociation of Mechanisms: The study provides empirical evidence that different types of "surprise" (repetition vs. omission) are processed differently within the same cortical area (A1).

4. Key Results

The results presented mixed evidence for Predictive Coding in A1:

A. Pattern Violations (Unexpected Repetitions):

• Finding: Contrary to PC theory, A1 neurons did not show enhanced responses to unexpected repetitions.
• Data: The median PEI was significantly negative (Monkey W: -0.05; Monkey C: -0.04).
• Interpretation: Responses to unexpected repetitions were not enhanced; in fact, they were slightly suppressed or unchanged. This aligns with Stimulus-Specific Adaptation (SSA) or forward suppression rather than a prediction-error signal.

B. Omission Violations (Unexpected Absences):

• Finding: Consistent with PC theory, A1 neurons showed enhanced responses to unexpected stimulus omissions.
• Data: The median PEI was significantly positive (Monkey W: 0.05; Monkey C: 0.10).
• Interpretation: The enhanced firing rate to an omitted stimulus cannot be explained by SSA (which requires physical stimulus repetition). This suggests a genuine prediction-error signal or a strong top-down expectation signal.

C. Neuron-by-Neuron Comparison:

• A direct comparison (Fig. 8) showed that for individual neurons, the PEI for omissions was significantly larger than the PEI for pattern violations. The majority of neurons exhibited a "mixed" profile: suppressed/neutral for repetition violations but enhanced for omission violations.

5. Significance and Implications

• Constraints on Predictive Coding: The study refines the understanding of PC by demonstrating that it is not uniformly implemented across all types of sensory violations in A1. While A1 appears capable of signaling "missing" expected events (omissions), it does not appear to signal "unexpected" repetitions via enhanced spiking.
• Re-evaluating Primary Sensory Cortex: The findings suggest that in primary auditory cortex, "surprise" regarding stimulus presence is largely governed by local adaptation mechanisms (SSA), whereas "surprise" regarding stimulus absence may involve top-down feedback or distinct circuitry.
• Theoretical Impact: The results challenge the notion that A1 is a primary generator of prediction-error signals for all types of statistical violations. Instead, it supports a model where PC mechanisms in A1 are selective, potentially limited to specific temporal or feature-based violations (like omissions) that cannot be explained by local feedforward suppression.
• Future Directions: The authors note that while single-unit spiking in A1 shows mixed results, population-level activity or activity in non-primary auditory areas might still support broader PC theories. Additionally, the role of active task engagement (vs. passive listening) remains a variable to explore.

Conclusion:
Shukla et al. provide crucial empirical evidence that the primary auditory cortex implements a nuanced form of predictive processing: it robustly signals the absence of expected stimuli but fails to signal the unexpected presence of repeated stimuli, thereby placing strict boundaries on how predictive coding is neurobiologically realized in early sensory cortices.