Optimal thresholds and algorithms for a model of multi-modal learning in high dimensions