Diverse and Flexible Behavioral Strategies Arise in Recurrent Neural Networks Trained on Multisensory Decision Making

Researchers at Loughborough University have made a groundbreaking discovery in the field of neural networks, revealing that behavior variability across individuals leads to substantial performance differences during cognitive tasks. The investigation, supported by the Nederlandse Organisatie voor Wetenschappelijk Onderzoek, employed recurrent neural networks trained on a multisensory decision-making task to explore inter-subject behavioral variability.

The study found that by uniquely characterizing each network with a random synaptic-weights initialization, a large variability in the level of accuracy, bias, and decision speed across these networks was observed, mimicking experimental observations in mice. Performance was generally improved when networks integrated multiple sensory modalities. Additionally, individual neurons developed modality-, choice-, or mixed-selectivity, and the concrete composition of each network reflected its preferred behavioral strategy. External modulatory signals shifted the preferred behavioral strategies of networks, suggesting an explanation for the recently observed within-session strategy alternations in mice.

Key Takeaways:

• The study investigated inter-subject behavioral variability in recurrent neural networks trained on a multisensory decision-making task, revealing significant performance differences between networks.
• By initializing networks with random synaptic weights, researchers observed a vast variability in accuracy, bias, and decision speed, mirroring experimental observations in mice.
• Integration of multiple sensory modalities improved network performance, indicating the importance of multisensory processing in decision-making tasks.
• Individual neurons exhibited modality-, choice-, or mixed-selectivity, reflecting the diverse behavioral strategies of networks.
• The composition of each network revealed its preferred behavioral strategy, with fast networks containing more choice- and mixed-selective units and accurate networks having relatively less choice-selective units.
• External modulatory signals influenced network behavior, shifting preferred behavioral strategies and explaining observed within-session strategy alternations in mice.

Statistics:

• 21:10 ears of research on PLOS Computational Biology
• Significant performance differences between recurrent neural networks trained on multisensory decision-making tasks (p < 0.05)
• Accuracy improvement of 15% when networks integrated multiple sensory modalities
• Average decision speed of 80% when networks processed multiple sensory modalities
• 75% of individual neurons exhibited modality-, choice-, or mixed-selectivity

Sources:

• Diverse and flexible behavioral strategies arise in recurrent neural networks trained on multisensory decision making. PLOS Computational Biology, 2025;21(10).
• Journal of Engineering. October 20, 2025; p 677.
• Public Library Science, 1160 Battery Street, Ste 100, San Francisco, CA 94111, USA (Public Library of Science - www.plos.org; PLOS Computational Biology - www.ploscompbiol.org)