Learning from Duality in Neural Network Field Theories
Organizations: STAG Research Centre, University of Southampton, Southampton, SO17 1BJ, UK · Institute of Physics, University of Amsterdam, 904 Science Park, 1098 XH Amsterdam, Netherlands · Department of Mathematical Sciences, Ulsan National Institute of Science and Technology, 50 UNIST-gil, Ulsan 44919, South Korea
Abstract
We develop neural realizations of positive-level Wess-Zumino-Witten models using bosonic Wakimoto and fermionic Fourier-Berezin fields. Level-rank duality connects the charged sectors of the fermionic construction, where explicit neural observables recover conformal block data and enable the transfer of information about unknown correlators between paired theories. We also show that distinct neural presentations of the same continuum current algebra exhibit different finite-width behavior, with approximation accuracy depending on architecture, regulator, and parameter distribution. These results connect the exact mathematical structure of conformal field theory with the expressive and computational properties of neural field theories.
Figures & tables
| Regulator | |||||
|---|---|---|---|---|---|
| condition | Hardy | Random | Direct | Neural |
|---|---|---|---|---|
| 0.0458 | 7.0854 | 0.0520 | 0.0738 | |
| 0.5792 | 6.5633 | 0.3828 | 0.4162 | |
| 0.3871 | 6.2779 | 0.4054 | 0.4010 | |
| 0.3865 | 9.8797 | 0.3387 | 0.3479 | |
| 0.5470 | 9.8454 | 0.5492 | 0.5579 |
| orbits | Exact/source | Blind | Wrong | Kernel | Median | control |
|---|---|---|---|---|---|---|
| 3 | 0.7577 | 0.6640 | 0.5912 | 0.5872 | 0.5667 | 0.6729 |
| 5 | 0.4510 | 0.6011 | 0.5250 | 0.5002 | 0.6032 | 0.4056 |
| RRMS versus KZ (%) | maximum node error (%) | |||
|---|---|---|---|---|
| grade | ||||
| two | 0.546 | 0.808 | 1.347 | 1.928 |
| three | 0.217 | 0.319 | 0.529 | 0.775 |
| four | 0.118 | 0.153 | 0.234 | 0.336 |
| mean test RRMS (%) | reverse | |
|---|---|---|
| grade-three neural response | 0.322 | 0.346 |
| grade-three KZ reference | 0.466 | 0.400 |
| grade-four KZ reference | 0.363 | 0.367 |
| mean pure-probe RRMS (%) | reverse | |
|---|---|---|
| no added noise | 1.35 | 2.49 |
| measurement noise | 2.65 | 3.46 |
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
| Variables/neuron | |||||
|---|---|---|---|---|---|
| 38 | 0.8633 | 0.15310 | 2.4652 | 3.0647 | |
| 38 | 1.4414 | 2.3759 | 2.2965 | ||
| 76 | 0.8906 | 0.14427 | 2.3858 | 2.9708 | |
| 80 | 0.8906 | 0.14427 | 2.3870 | 2.9718 |
| Operator pattern | Operator pattern | ||
|---|---|---|---|
| 0 | 7 | ||
| 1 | 8 | ||
| 2 | 9 | ||
| 3 | 10 | ||
| 4 | 11 | ||
| 5 | 12 |
| Configuration | ||||
|---|---|---|---|---|
| Correlator type | Coordinate factor |
|---|---|
| Set | Improved | Worsened | |||||
|---|---|---|---|---|---|---|---|
| 0.398214 | 1 | 0.994654 | 144544 | 255 | 0 | ||
| 0.418142 | 1 | 1.049563 | 263027 | 220 | 35 | ||
| 0.473206 | 1 | 1.040337 | 218776 | 226 | 29 | ||
| 0.229583 | 1 | 0.548762 | 21878 | 255 | 0 | ||
| 0.170406 | 7 | 1.365613 | 2188 | 225 | 30 | ||
| 0.152453 | 7 | 1.554804 | 2188 | 220 | 35 |
| Set | Before RMS | After RMS | After/before | ||
|---|---|---|---|---|---|
| 1 | 0.398214 | ||||
| 1 | 0.418142 | ||||
| 1 | 0.473206 | ||||
| 0.458596 | |||||
| 0.938208 | |||||
| 0.921068 |
| Task | time (s) | time (s) | Ratio | |
|---|---|---|---|---|
| Full calculation | ||||
| Adjustable parameters | ||||
| Fixed parameters | ||||
| Full calculation | ||||
| Adjustable parameters | ||||
| Fixed parameters |
| condition | reference/KZ | direct/KZ | neural/KZ |
|---|---|---|---|
| 0.1627 | 0.1765 | 0.1896 | |
| 0.1627 | 0.2206 | 0.2464 | |
| 0.1627 | 0.2577 | 0.2519 | |
| 0.2953 | 0.1945 | 0.2050 | |
| 0.2953 | 0.2599 | 0.2650 |