Statistical Properties of the King Wen Sequence: An Anti-Habituation Structure That Does Not Improve Neural Network Training
The King Wen sequence of the I-Ching, dating back to approximately 1000 BC, presents a fascinating ordering of 64 hexagrams, which represent states within a six-dimensional binary space. For over three millennia, scholars have pondered the underlying principles of this sequence. Recent research has aimed to statistically characterize this ordering through Monte Carlo permutation analysis, comparing its properties against 100,000 random baselines.
Key Findings
The analysis yielded four statistically significant properties of the King Wen sequence:
- Higher-than-random transition distance: Positioned in the 98.2nd percentile, indicating a unique transition structure.
- Negative lag-1 autocorrelation: This was observed with a p-value of 0.037, suggesting a tendency for transitions that are less predictable.
- Yang-balanced groups of four: A significant finding with a p-value of 0.002, indicating a balance in the grouping of hexagrams.
- Asymmetric within-pair vs. between-pair distances: This property reached the 99.2nd percentile, highlighting a distinct pattern in the arrangement of hexagrams.
Hypothesis and Experiments
These properties bear a resemblance to principles observed in curriculum learning and curiosity-driven exploration, leading researchers to hypothesize that the King Wen sequence could enhance neural network training. To investigate this, three distinct experiments were conducted:
- Learning rate schedule modulation: Testing various learning rate adjustments using the King Wen sequence.
- Curriculum ordering: Evaluating the performance of the King Wen sequence as an ordering method compared to traditional sequential orders.
- Seed sensitivity analysis: Conducting a 30-seed sweep to assess variability in performance outcomes.
Results and Conclusions
The outcomes of these experiments were uniformly negative regarding the effectiveness of the King Wen sequence in enhancing neural network training. Specifically, the modulation of learning rates using the King Wen sequence consistently degraded performance across all tested amplitudes. Furthermore, in terms of curriculum ordering, the King Wen sequence was found to be the least effective non-sequential ordering on one testing platform, while performing comparably to noise in another.
A comprehensive 30-seed sweep confirmed that only the degradation associated with the King Wen sequence surpassed the natural variance found in seed performance. The underlying reason for this detrimental effect appears to stem from the sequence’s high variance. While this high variance contributes to its statistical uniqueness, it simultaneously disrupts gradient-based optimization processes, ultimately undermining training efficacy.
In conclusion, while the King Wen sequence showcases intriguing statistical properties, its application as an anti-habituation structure does not translate into improved performance in neural network training dynamics.