Deliberate Practice Stabilizes Working Memory in the Brain

Q: What is representational drift in neuroscience?

Representational drift describes how only 0.2 ± 0.2% of M2 neurons retained late-delay selectivity across the naive-to-expert learning period in Bellafard et al. (2024), even though decoding accuracy remained reliable. Despite this turnover, the same neural population could still be decoded each day. The phenomenon reflects continuous neuron reassignment rather than a global loss of activity.

Q: How does practice affect the stability of working-memory representations in the brain?

73,307 M2 neurons were imaged simultaneously by Bellafard et al. (2024) to test whether continued practice stabilizes working-memory representations in mice. Decoding accuracy for the late-delay epoch was significantly higher on days 6-10 of expert performance than on days 1-5 (P < 0.0001). The result indicates representations stabilize only after extended practice beyond initial competence.

Q: What method did researchers use to measure working-memory information in neural activity?

90% of trials were used to train linear SVM decoders, with the remaining 10% used for testing, in Bellafard et al. (2024). The classification was repeated at least 32 times with randomized trial splits to avoid overfitting. A nonlinear LSTM decoder produced similar results, confirming the SVM findings.

Q: What brain region was most important for working memory in this study?

M2 neurons could decode the first odour during the late-delay epoch, but M1 and retrosplenial cortex (RSA) neurons could not, per Bellafard et al. (2024). Muscimol inactivation of M2 reduced expert task performance from 92.8 ± 1.2% to 70.5 ± 1.8%. Adding M1 or RSA activity to M2 did not improve decoding accuracy.

Deliberate Practice Stabilizes Working Memory in the Brain

Summary by Robert Gorak · Published June 18, 2026 · Last reviewed June 18, 2026

Arash Bellafard and Ghazal Namvar and Jonathan C. Kao and Alipasha Vaziri and Peyman Golshani·2024·Nature

Sample: Up to 73,307 M2 neurons recorded simultaneously (volumetric imaging, n = 4 mice); 622 ± 295 neurons per session (single-plane L2/3 imaging)Period: Behavioral training to expert level over approximately 7 sessions; representation stability tracked over 10 consecutive days of single-plane imaging and 10 consecutive days of volumetric imaging

Representational drift describes the gradual change in which neurons encode the same information over repeated days, even as behavior remains stable. In "Volatile working memory representations crystallize with practice," Bellafard, Namvar, Kao, Vaziri, and Golshani (2024) tracked M2 neurons in mice learning an olfactory working-memory task. They recorded up to 73,307 neurons simultaneously using volumetric light-bead microscopy across 10 days of expert performance. Late-delay decoding accuracy was significantly higher on days 6-10 of expert performance than on days 1-5 (P < 0.0001).

What the Study Found

Optogenetic inhibition of M2 during the fourth delay second cut performance by 24.7 ± 9.6%, and during the fifth delay second by 29.2 ± 5.4%. Inhibition during the second odour reduced performance by 34.6 ± 5.9%, and during early choice by 31.9 ± 3.3%. Muscimol inactivation of M2 lowered expert performance to 70.5 ± 1.8%, versus 92.8 ± 1.2% with saline. Cross-day late-delay decoding, using up to 73,307 M2 neurons, was significantly higher on days 6-10 than days 1-5 (P < 0.0001). In single-plane L2/3 imaging, late-delay decoding showed no such improvement across days 1-3 versus days 5-7 (P = 0.12).

Methodology

The study recorded secondary motor cortex (M2) activity in head-fixed mice performing an olfactory delayed-association working-memory task. Researchers tracked up to 73,307 M2 neurons simultaneously using volumetric light-bead microscopy, and 622 ± 295 neurons per session using single-plane L2/3 imaging. Neural stability was assessed over 10 consecutive days of single-plane imaging and 10 consecutive days of volumetric imaging during expert performance. Key controls included mCherry- and EGFP-expressing mice for optogenetic experiments, saline injections for muscimol experiments, and a non-working-memory go/no-go task.

Key Statistics

Metric	Finding	Context
Behavioral accuracy after learning	94.2 ± 1.3% (D' > 3)	After approximately 7 training sessions
M2 optogenetic inhibition effect	24.7-34.6% performance reduction	Late-delay seconds 4 and 5, second odour, and early-choice epochs
Muscimol M2 inactivation	70.5 ± 1.8% vs. 92.8 ± 1.2% (saline)	Expert mice, tested 1 hour post-injection
Cross-day late-delay decoding (volumetric)	Significantly higher days 6-10 vs. days 1-5 (P < 0.0001)	Up to 73,307 M2 neurons, light-bead microscopy
Cross-day late-delay decoding (single-plane)	No significant change, days 1-3 vs. days 5-7 (P = 0.12)	L2/3 imaging, early expert phase
Neurons correlated with limb movement	1.2 ± 0.2% (32 of 2,611 cells)	5 expert mice, DeepLabCut paw tracking

Why This Matters

A newly learned skill appears to be initially carried by neural patterns that fluctuate from day to day before settling into a stable form. For anyone developing a rule-based decision skill through paper trading, early competence may not yet reflect a durable internal model. Repetition beyond the point of basic proficiency, not just initial mastery, may be required to consolidate a decision process into a stable representation. Deliberate practice may therefore work as a consolidation period rather than a single competence threshold.

Frequently Asked Questions

Representational drift describes how only 0.2 ± 0.2% of M2 neurons retained late-delay selectivity across the naive-to-expert learning period in Bellafard et al. (2024), even though decoding accuracy remained reliable. Despite this turnover, the same neural population could still be decoded each day. The phenomenon reflects continuous neuron reassignment rather than a global loss of activity.

73,307 M2 neurons were imaged simultaneously by Bellafard et al. (2024) to test whether continued practice stabilizes working-memory representations in mice. Decoding accuracy for the late-delay epoch was significantly higher on days 6-10 of expert performance than on days 1-5 (P < 0.0001). The result indicates representations stabilize only after extended practice beyond initial competence.

90% of trials were used to train linear SVM decoders, with the remaining 10% used for testing, in Bellafard et al. (2024). The classification was repeated at least 32 times with randomized trial splits to avoid overfitting. A nonlinear LSTM decoder produced similar results, confirming the SVM findings.

M2 neurons could decode the first odour during the late-delay epoch, but M1 and retrosplenial cortex (RSA) neurons could not, per Bellafard et al. (2024). Muscimol inactivation of M2 reduced expert task performance from 92.8 ± 1.2% to 70.5 ± 1.8%. Adding M1 or RSA activity to M2 did not improve decoding accuracy.