A Gestalt inference model for auditory scene segregation.

Area of Science:

• Computational Neuroscience
• Auditory Perception
• Machine Learning

Background:

• Auditory scene segregation, the brain's ability to parse complex soundscapes, is often explained by Gestalt principles.
• These principles suggest sound attributes are extracted and grouped, mirroring visual object segmentation from natural image statistics.
• Existing models often lack a unified framework for emergent cue learning from naturalistic auditory data.

Purpose of the Study:

• To investigate emergent auditory grouping cues using stochastic neural networks.
• To model the binding of spectro-temporal attributes into segregated auditory objects.
• To emulate and quantify phenomena of auditory scene segregation.

Main Methods:

• Leveraged inference in stochastic neural networks to learn grouping cues directly from natural soundscapes (speech, music, nature sounds).
• Developed a hierarchical model learning local and global spectro-temporal attributes, akin to Gestalt cues.
• Employed a Hebbian network for fusing features and binding coherent attributes into auditory objects.

Main Results:

• The neural network model learned hierarchical spectro-temporal attributes, reflecting simultaneous and sequential Gestalt principles.
• The architecture successfully emulated established auditory scene segregation phenomena across multiple time scales.
• Quantified the complementary roles of segregation and binding cues in auditory perception.

Conclusions:

• Stochastic neural networks can learn emergent auditory grouping cues from natural soundscapes.
• The proposed model provides a computational framework for understanding auditory scene segregation based on Gestalt principles.
• This approach highlights the interplay between feature segregation and binding mechanisms in auditory object formation.