Separate the effects of labels, inputs, and fusion
The architecture alone does not describe the experiment. Unit labels can construct positive pairs, type labels can supervise the rhythm branch, and ID labels can supervise the spectral branch. The matched grid varies these choices while using uniform shuffled sampling and the same evaluation protocol.
U/T/I indicate unit pairing, type loss, and ID loss. All eight combinations are retained below. Absence of task labels in the representation objective does not remove the annotated ICI input or supervision used by downstream probes.
| U/T/I | Random unit | Random type | Random ID | Grouped ID |
|---|---|---|---|---|
| 000 | 0.693 ± 0.027 | 0.583 ± 0.042 | 0.624 ± 0.064 | 0.181 ± 0.133 |
| 001 | 0.674 ± 0.016 | 0.566 ± 0.017 | 0.592 ± 0.061 | 0.159 ± 0.054 |
| 010 | 0.709 ± 0.027 | 0.770 ± 0.086 | 0.591 ± 0.057 | 0.209 ± 0.091 |
| 011 | 0.725 ± 0.013 | 0.802 ± 0.046 | 0.600 ± 0.080 | 0.174 ± 0.068 |
| 100 | 0.861 ± 0.001 | 0.493 ± 0.091 | 0.434 ± 0.067 | 0.245 ± 0.084 |
| 101 | 0.885 ± 0.019 | 0.348 ± 0.112 | 0.474 ± 0.032 | 0.134 ± 0.039 |
| 110 | 0.901 ± 0.020 | 0.865 ± 0.115 | 0.597 ± 0.051 | 0.248 ± 0.130 |
| 111 | 0.907 ± 0.010 | 0.735 ± 0.091 | 0.581 ± 0.012 | 0.223 ± 0.175 |
Pairing imposes task-dependent tradeoffs
Without auxiliary losses, same-unit pairing changes random unit/ID means from 0.693/0.624 (000) to 0.861/0.434 (100). The grouped-ID ordering reverses: 0.181 versus 0.245. A pairing objective can favor unit-level consistency without uniformly improving individual discrimination.
These are matched contrasts within this exploratory batch. They do not identify a universal best objective and do not isolate the biological mechanism producing the differences.
Channel controls and joint concatenation
The single-channel controls are trained architectures. They are different from taking a branch out of an already trained dual encoder and probing it afterward. Both auxiliary heads use the available branch in single-channel models. Joint concatenation retains 128 dimensions without a fusion MLP.
| Architecture / UTI | Random type | Random ID | Grouped ID |
|---|---|---|---|
| Rhythm 110 | 0.978 ± 0.021 | 0.477 ± 0.037 | 0.352 ± 0.134 |
| Rhythm 111 | 0.973 ± 0.013 | 0.454 ± 0.061 | 0.331 ± 0.143 |
| Spectrum 110 | 0.202 ± 0.021 | 0.358 ± 0.103 | 0.054 ± 0.047 |
| Spectrum 111 | 0.199 ± 0.037 | 0.444 ± 0.092 | 0.092 ± 0.016 |
| Joint concat. 110 | 0.939 ± 0.062 | 0.636 ± 0.015 | 0.275 ± 0.117 |
| Joint concat. 111 | 0.927 ± 0.060 | 0.696 ± 0.055 | 0.239 ± 0.043 |
The table reproduces the final manuscript’s six architecture rows. The dimensionality-different concatenation control does not establish a necessary benefit of learned fusion. Neither the single-channel scores nor the dual branches establish an exclusively spectral identity code.
Actual swapping and bandwidth
The primary model pairs whole codas from the same unit; it does not swap rhythm and spectrum between codas. The separate swapping condition makes that exchange explicitly. Relative to whole-coda U+T+I, actual swapping raises random ID from 0.581 to 0.657, while grouped ID falls from 0.223 to 0.189.
Increasing the mel ceiling from 8 to 16 kHz raises random ID to 0.631 without ID loss and 0.623 with ID loss. Grouped values are 0.250 and 0.173. The changed ceiling also changes filter spacing at fixed mel-band count. These are bandwidth sensitivities, not localization of the frequencies carrying biological identity.
Post-hoc branch diagnostics
For 12 saved dual checkpoints, rhythm and spectral embeddings were extracted separately and evaluated under the same frozen-probe protocol. Re-extracted fusion embeddings were checked against the saved ones. No encoder was retrained for this diagnostic.
| Condition / output | Random unit | Random type | Random ID |
|---|---|---|---|
| U+T / rhythm | 0.639 ± 0.027 | 0.983 ± 0.023 | 0.459 ± 0.052 |
| U+T / spectral | 0.863 ± 0.013 | 0.158 ± 0.023 | 0.367 ± 0.116 |
| U+T / fusion | 0.901 ± 0.020 | 0.865 ± 0.115 | 0.597 ± 0.051 |
| U+T+I / rhythm | 0.632 ± 0.028 | 0.982 ± 0.019 | 0.461 ± 0.054 |
| U+T+I / spectral | 0.861 ± 0.001 | 0.169 ± 0.021 | 0.443 ± 0.059 |
| U+T+I / fusion | 0.907 ± 0.010 | 0.735 ± 0.091 | 0.581 ± 0.012 |
The spectral-branch improvement with ID supervision does not survive as an improvement in the fused-ID mean. Objective interactions or compression are possible explanations, but those mechanisms are not established by these experiments.
Evidence: all supervision definitions, branch diagnostic report, paired contrasts, and full experiment evidence.