High-temperature Expansions for Learning Models of Nonnegative Data