On the Sparse Structure of Natural Sounds and Natural Images: Similarities, Differences, and Implications for Neural...

Area of Science:

• Computational neuroscience
• Machine learning
• Sensory processing

Background:

• Sparse coding models successfully predict neural responses in visual and auditory systems.
• The universality of learned structures across sensory modalities remains unclear.

Purpose of the Study:

• To investigate differences in statistical structures learned by sparse coding models from natural images and speech.
• To understand how these differences affect neural representations in different sensory modalities.

Main Methods:

• Fitting complete and overcomplete sparse coding models to natural images and speech spectrograms.
• Analyzing the distributions and sparseness of learned features.
• Evaluating impact on a biologically plausible sparse coding network with local plasticity.

Main Results:

• Natural images yield features with similar, Laplace-like distributions.
• Speech features exhibit a wider range of asymmetric distributions.
• Overcomplete models show higher lifetime sparseness for images, greater population sparseness for speech.
• Biologically plausible networks learn distinct speech features compared to conventional algorithms.

Conclusions:

• The statistical structure learned from natural images and speech differs significantly.
• These modality-specific statistics influence neural coding and representation.
• Sparse coding models provide insights into sensory processing across different modalities.