Experiment Documentation

Detailed description of every experiment behind the manuscript "Relevance-Aware Semantic Multiple Access via Meta-Learned User-Wise Cross-Attention." Each entry states the reviewer concern it addresses, the exact setup, the procedure, the metrics, the key results, and the artifact paths. All experiments run on a single NVIDIA RTX A4500 (PyTorch 2.10, CUDA 12.8), seed 42, and write raw JSON results consumed only by rev2/plot_rev2.py.

Common system model (all rev2 experiments, matching manuscript Eq. (7)): single superimposed uplink y = sum_v |h_v| e^{j dphi_v} (e_v (.) m_v) + n over one d=64-dimensional frame; fixed {0,1} disjoint transmit block masks (d/U coordinates per user, identical for every compared scheme); per-rail noise std = sqrt(mean|y_tx|^2 / SNR); U=4 users, HIGH/LOW/MIX scenarios of Table II unless stated. The decoder-side masks are separate learned soft masks in [0,1]^d initialized at the block pattern. Training is the manuscript's first-order meta-training aggregated over SNR tasks {0,4,...,20} dB, Adam 1e-3 (mask logits 0.1), 250-300 epochs, batch 64; evaluation uses 150-200 Monte Carlo batches of 64 per point.

E1 — Comparison fairness against optimal linear receivers

Reviewer concerns: R1-10, R3-7 (d/U comparison "mathematically unfair"); editor point 4.

Question. Is the OFDMA/SFDMA subspace ceiling an artifact of weak baselines, and how much of the UWCA gain is mere dimensionality?
Setup. For each scenario the true relevance matrix B (B_uv = beta_u beta_v for shared scene, else 0) and the per-realization channel magnitudes are formed. Receivers evaluated on the same received frame: uwca (trained), ofdma, sfdma, noma (full-band power-domain SIC, powers 0.40/0.30/0.20/0.10), lmmse_blind (closed-form Wiener with cross-covariances zeroed), lmmse_genie (closed form with true B and channel gains, manuscript Eq. (23)), tdma_proj (random orthonormal 16-dim projection per user — an arbitrary orthogonal partition).
Key results. lmmse_blind = OFDMA at every SNR in all scenarios (SER 0.509 vs 0.508 @10 dB HIGH), tdma_proj and SFDMA coincide with them; HIGH @20 dB: blind 0.327 -> UWCA 0.110 -> genie 0.044 (UWCA recovers 77% of the blind-to-genie gap); LOW: genie = blind (nothing to exploit), UWCA within 0.02 SER.
Artifacts. rev2/e1_fair_baselines.py -> rev2/data/e1_fair_baselines.json -> fig/fig_fair.pdf (manuscript Fig. 3). Trained checkpoints e1_uwca_{HIGH,LOW,MIX}.pt (reused by E4/E6).

E2 — Full complex-baseband phase errors and CSI error

Reviewer concerns: R1-1 (phase model underestimates IUI), R2-3 (CSI).

Question. Does the multi-user superposition amplify residual phase errors into severe inter-user interference once nothing is absorbed into a noise term?
Setup. Complex channel with per-user residual dphi_u ~ N(0, sigma^2), sigma in {0,5,10,15,20,30} deg; both rails simulated so every leakage path exists. Three decoders: mismatch-trained (never saw phase errors), phase-augmented (trained with sigma ~ U[0,20] deg), and a two-rail variant whose keys/values read [Re;Im] (2d input). CSI sweep: amplitude error h_hat = h(1+eps), eps ~ N(0, sigma_h^2), sigma_h up to 0.2, at 10 deg. Learned-mask overlap (mean pairwise cosine) is also measured.
Key results. Mismatch-trained @10 dB: SER 0.270 -> 0.291 (20 deg) -> 0.332 (30 deg); @20 dB 0.110 -> 0.127; augmentation and I/Q reading change curves by less than Monte Carlo spread -> leakage is second order because it scales with sin(dphi) x mask overlap (measured 0.26). CSI: UWCA unchanged within 0.004 (uses no explicit CSI).
Artifacts. rev2/e2_phase_iui.py -> rev2/data/e2_phase_iui.json; fig/fig_phase2.pdf (repo; numbers narrated in manuscript Sec. VII-E).

E3 — Dynamic user population

Reviewer concerns: R1-6 (fixed U, costly retraining), R2-3 (fixed identities).

Setup. U_max = 8 mask slots (d=64, 8 dims/slot), HIGH-type correlation (beta_u = 0.6, one scene). Active set resampled per frame; inactive users transmit nothing and their softmax scores are masked to -inf (the scheduler announces the active set). Compared: one model trained with the full population always active (never retrained), an activity-sampled model, and per-count oracle models retrained from scratch for |A| in {2,4,6,8}. Metric: mean cosine (the sqrt(1/8)=0.35 subspace ceiling saturates the binary SER at this scale).
Key results. The single full-population model sustains fidelity over every active count (c 0.34 -> 0.39 for |A| = 2 -> 8 @10 dB) and matches or exceeds the per-count retrained models (0.33/0.32/0.36/0.39); per-count retraining is counterproductive (sparse populations train each slot on a fraction of the traffic).
Artifacts. rev2/e3_dynamic_users.py -> rev2/data/e3_dynamic_users.json; fig/fig_dynusers.pdf (repo).

E4 — Symbol-timing offsets (v3 = block-wise realignment)

Reviewer concern: R1-8 (asynchronous reception / ISI).

Setup. Per-user integer offsets delta_u ~ U{0..Delta} symbols, Delta in {0,1,2,4,8}, shift each user's transmitted block within the frame (edge energy lost). Conditions: uncorrected (UWCA and OFDMA), corrected by block-wise realignment using pilot-estimated offsets (each user's block region shifted back individually), and corrected with a deliberately impaired estimator (+-1 symbol on 20% of users). v1 (e4_async results in json) showed offset-augmented training cannot repair unknown shifts; v2 showed whole-frame realignment breaks cross-block alignment — both superseded by v3.
Key results. Uncorrected offsets are catastrophic for every embedding-level scheme (one symbol: UWCA 0.26 -> 0.73, OFDMA 0.51 -> 0.76) because i.i.d. embedding coordinates fully decorrelate under a one-symbol misalignment. With realignment the degradation is gradual (0.266 -> 0.304 @Delta=1, 0.455 @Delta=8) and realigned UWCA stays below realigned OFDMA (0.536-0.638) at every offset. The impaired estimator costs 0.14 SER — whole-symbol residuals sacrifice the affected block, so timing must be sub-symbol (standard timing advance).
Artifacts. rev2/e4_v3_async.py -> rev2/data/e4_v3_async.json; fig/fig_async.pdf (repo).

E5 — Nonlinear inter-user semantic structure

Reviewer concerns: R1-7, R2-5, R3-2 (linear scalar model too restrictive).

Setup. Embeddings e_u = normalize(g_u([s; p_u])) with fixed random two-layer tanh view networks g_u per user (seed 7): users share the scene s only through independent nonlinear transformations. Cases: shared scene vs independent scenes (control). Schemes: trained UWCA, OFDMA, NOMA-SIC, and the scalar-parameterized genie LMMSE fed the measured mean pairwise cosine.
Key results. Linear correlation is destroyed (mean cosine 0.006), so the genie LMMSE collapses onto the ceiling (0.515 vs OFDMA 0.518 @10 dB) — no scalar/linear receiver can represent the shared structure. UWCA still attains 0.363 @10 dB / 0.220 @20 dB. The independent-scene control (UWCA 0.441) isolates the manifold-prior share, so the further reduction to 0.363 is pure nonlinear cross-user structure.
Artifacts. rev2/e5_nonlinear.py -> rev2/data/e5_nonlinear.json (numbers narrated in manuscript Sec. VII-G).

E6 — Residual orthogonality vs content preservation

Reviewer concerns: R1-4, R2-2, R3-3 (semantic-orthogonality self-contradiction).

Setup. E1's trained HIGH decoder; per-sample Pearson correlation across the 64 dimensions, averaged over user pairs and 100x64 samples per SNR, for: input embeddings, decoded embeddings, decoding residuals r_u = e_hat_u - e_u; OFDMA decoded correlation as reference.
Key results. Decoded-embedding correlation rises with SNR from 0.31 toward the 0.39 input level (shared content preserved, not stripped); residual correlation falls 0.23 -> 0.14 (2.8x below input) — the emergent residual orthogonality. OFDMA's decoded correlation is 0.00 at every SNR: orthogonal access erases the inter-user semantic structure.
Artifacts. rev2/e6_residual_orth.py -> rev2/data/e6_residual_orth.json -> fig/fig_resorth.pdf (manuscript Fig. 4).

E7 — Meta-adaptation beyond the SNR axis (v2)

Reviewer concerns: R1-3, R3-4, R3-5 (MAML overkill for a 1-D lookup); R1-9 (eta stability).

Setup. Task family = 6 SNRs x {Rayleigh, Rician K=5 dB, K=10 dB} x phase residual {0,10} deg (36 tasks). Systems: one initialization meta-trained over the full family (300 epochs, eta and outer-gradient norm logged every epoch); per-SNR specialist bank trained on Rayleigh/no-phase (the 1-D lookup, nearest-SNR index). Held-out tests: Rician K=20 dB + 15 deg (10/18 dB), Nakagami m=3 + 5 deg (a fading law in no training task), Rayleigh + 20 deg @6 dB, plus an in-distribution check; zero-shot and after S in {1,5,10,20} inner steps (SGD 0.02, one 64-sample support batch).
Key results. Family initialization transfers zero-shot with 9-46% lower SER than the lookup on every held-out condition (e.g. 0.114 vs 0.158 Rician-K20-15deg @10 dB; 0.139 vs 0.190 unseen Nakagami); inner-loop adaptation is stable (within 0.005 over 20 steps). eta converges to 0.94 (max 1.00), outer gradient norm <= 0.024 vs clip bound 5 — no softmax saturation or divergence.
Artifacts. rev2/e7_v2_meta.py -> rev2/data/e7_v2_meta.json (includes eta/gradient trajectories).

E8 — End-to-end training with anti-collapse (v2)

Reviewer concern: R1-5 (representation collapse prevents joint JSCC).

Setup. Four configurations on HIGH: frozen encoder (reference); naive end-to-end with the distortion measured against the encoder's own output (the collapse-prone moving target); source-anchored end-to-end (distortion vs normalize(x)); moving target + VICReg-style variance-covariance regularizer (hinge at 1/sqrt(d) per-dim std, 25x variance + 100/d covariance weights). Collapse-proof metric: batch nearest-neighbor retrieval accuracy over the encoded gallery, plus the effective rank of the encoder-output covariance.
Key results. Naive collapses exactly as the reviewers expect: retrieval falls to chance (0.005 vs 0.536 frozen @10 dB) — the pathology is the moving-target objective, not the cross-attention. Source anchoring keeps effective rank 59.6/64 and lands within 0.04 SER of frozen; VICReg restores retrieval to 0.509 with a learned task-specific code. E2E training is therefore demonstrated, and the frozen encoder in the main experiments is a controlled-isolation choice.
Artifacts. rev2/e8_v2_e2e.py -> rev2/data/e8_v2_e2e.json.

E9 — Online top-k relevance acquisition and measured overhead

Reviewer concerns: R1-2, R3-6 (circular dependency), R2-4 (cost at large populations).

Setup. U=32, eight relevance clusters of four, k=4, briefly trained full-attention model. Protocol: frames 1-3 full attention (needs no relevance) while the BS accumulates beta_hat by EWMA (zeta=0.5) of decoded-embedding cosines; from frame 4, per-row top-k via argpartition on beta_hat. References: oracle top-k (peers selected from the true clusters) and permanent full attention. Timing: decoder forward vs EWMA-update + selection wall-clock for U in {8,16,32,64,128} (256-sample frames, GPU, 20-run averages).
Key results. From the first post-warm-up frame the online selection matches the oracle exactly and slightly exceeds full attention (discarding irrelevant peers discards their noise); ranking needs far less accuracy than estimation because intra-cluster (~0.42) and inter-cluster (~0) cosines are well separated. Estimation + selection cost 0.02-0.04 ms vs the 0.3-5.4 ms decoder pass, incurred once per relevance coherence interval.
Artifacts. rev2/e9_topk_online.py -> rev2/data/e9_topk_online.json (trajectory + timing).

Legacy experiments (retained from the original study)

Three-scenario SER study (manuscript Fig. 2): analytical simulation legacy/semantic_correlation_sim.py + trained decoder-only overlay legacy/maml_semantic.py --decoder_only; threshold sweep tau in [0.30,0.50] and mean-cosine table in legacy/revision_experiments.py.
beta sweep (monotone gain law): legacy/revision_betasweep.py (figure fig/fig4_beta_sweep.pdf, narrated in Sec. VII-C).
Attention heatmaps: legacy/plot_figures.py (figure fig/fig6_hlm.pdf, narrated in Sec. VII-C).
Hyperparameter ablations (S, lambda, d, H, K): legacy/revision_ablation.py, legacy/revision_dsweep.py, legacy/revision_e2e.py (Sec. VII-D).
Real-data study (manuscript Fig. 5): UCI optical-recognition digits (8x8=64 dims), legacy/revision_realdata_train.py + legacy/revision_realdata_plot.py (Sec. VII-G).

Reviewer-concern -> experiment map

Concern	Experiment(s)	Manuscript
R1-1 phase IUI	E2	Sec. III-A, VII-E
R1-2 / R3-6 top-k circularity	E9	Sec. V-E, VII-H
R1-3 / R3-4 / R3-5 MAML vs lookup	E7	Sec. V-E, VII-F
R1-4 / R2-2 / R3-3 orthogonality contradiction	E6	Def. 3, Prop. 4, Fig. 4
R1-5 representation collapse	E8	Sec. VI-B
R1-6 dynamic U	E3	Sec. IV-B, VII-E
R1-7 / R2-5 / R3-2 linear scalar model	E5 (+ real data)	Sec. III-B, VII-G
R1-8 asynchrony	E4	Sec. III-A, VII-E
R1-9 eta stability	E7 logs	Sec. IV-A, VII-F
R1-10 / R3-7 d/U fairness	E1	Prop. 3, Fig. 3, VII-B
R2-1 idealized proofs	assumption block + E1/E5/real data	Sec. V
R2-3 sync / CSI / identities / encoder	E4 / E2 / E3 / E8+E5	Sec. III, VI-B, VII-E
R2-4 large-population cost	E9 timing	Sec. VII-H
R3-1 DSC novelty	(positioning)	Sec. II-A

14 KiB Executable File Raw Blame History