Cross-modal applications of a neuromorphic olfactory learning algorithm.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Neuromorphic algorithms, inspired by biological sensory systems, offer potential for efficient AI.
• Adapting algorithms across different sensory modalities remains a challenge in AI research.

Purpose of the Study:

• To adapt an olfactory neuromorphic algorithm for image and sound recognition tasks.
• To evaluate the algorithm's one-shot online learning capabilities across multiple data modalities.
• To assess the effectiveness of principal component analysis (PCA) in feature representation for this algorithm.

Main Methods:

• Adapted an olfactory neuromorphic algorithm using modality-specific preprocessing for image (NIST digits) and sound (Google Speech Commands) data.
• Applied a gammatone filter to audio signals for noise reduction and temporal-to-frequency conversion.
• Modified the single stimulus test algorithm for spectrogram analysis and implemented PCA, retaining 90% variance.

Main Results:

• The adapted algorithm demonstrated successful one-shot online learning for both image and sound recognition over sequential "olfactory" gamma cycles.
• Principal component analysis (PCA) representations did not achieve high similarity to corresponding templates across all three modalities (including the original olfactory modality).

Conclusions:

• The olfactory neuromorphic algorithm can be effectively adapted for image and sound recognition, showcasing robust online learning capabilities.
• While the core algorithm shows promise, PCA's utility as a feature representation method requires further investigation for cross-modal template matching.