Abdalrhman0000/Dual-Expert-Self-Attention

.

license: cc-by-4.0 pretty_name: Dual-Expert Self-Attention language: [en, ar] tags: [attention, dual-expert, relational-reasoning, graded-alpha, pytorch]

Dual-Expert Attention v2.8.3: reference PyTorch implementation

Learning When to Look Like vs. Work Like: A Dual-Expert Attention with a Graded Alpha Knob Abdulrahman Al-Hayali, Mosul, Iraq, v2.8.3 (May 2026). This repo implements Eq. (1)-(15) with the reviewer fixes, as a full Transformer with no standard attention anywhere.

The idea in one screen

Standard attention asks one question: do q and k look alike? (E0 = q·k/√d). Great for dog → dogs, useless for Paris → France or 5+3 → 8: they don't look alike, they work together. So every head has two experts:

expert	formula	meaning
E0, looks-like	q·k / √d_head	unbounded similarity
E1, works-like	-(1/d_head) ‖W_rel q − k‖²	lock-and-key: does my key W_rel turn q into k? Best = 0

The scales are different (E0 unbounded, E1 ≤ 0), so each one is normalised per query across keys with its own stats (Ẽ = (E − μ_k)/(σ_k + 1e-5)). That is not nn.LayerNorm(d_model), and all mixing uses Ẽ, never raw E.

The DJ mixer α

S*(q,k;α) = (1−α)·Ẽ0 + α·Ẽ1. α is a graded, continuous knob, not a 0.2/0.9 label:

α	mix	example	the feel
≈0 [0, 0.05], 45% of mass	100% looks	dog → dogs	twins
~0.2 [0.15, 0.30], 25%	80% looks / 20% works	car → road	the cousin who looks like you but is actually your accountant
~0.9 [0.75, 0.95], 30%	10% looks / 90% works	king → queen, car → wheel, Paris → France, 5+3 → 8	relation dominates

Never 1.0. In March, α=1 made retrieval collapse (queen started retrieving queen bee), because pure relation forgets what the thing is. Both α and α̂ are clamped to [0, 0.95], so ≥5% similarity anchor always survives.

The oracle α comes from a frozen library, never from the trainable model (otherwise W_rel could cheat): α = σ(Φ_rel − γ_nli·Sim), γ_nli = 0.3, Φ_rel = DeBERTa-v3-large-MNLI entailment logit (premise=text(q), hypothesis=text(k)), Sim = all-MiniLM cosine. It's pre-computed offline with torch.no_grad() (calling DeBERTa inside an O(N²) attention is infeasible), optionally isotonic-calibrated on human labels.

Predicting α̂

α̂ = σ(clip(W_gate[q; k; Ẽ0; Ẽ1] + c_gate, −10, 10)), then clamp to 0.95. The logit is clipped (not the weights) so α̂(1−α̂) never vanishes. W_gate is per head, 1×(2·d_head+2), Xavier init; c_gate = −0.5 ≈ σ⁻¹(mean α). The concat is never materialised: the linear map splits exactly into q·w_q + k·w_k + w_E0·Ẽ0 + w_E1·Ẽ1.

β, the photo-finish

For a query, target t and distractor p are two straight lines as you turn the knob: S*_t(α') = (1−α')Ẽ0,t + α'Ẽ1,t, S*_p(α') = (1−α')Ẽ0,p + α'Ẽ1,p. They cross at β = (Ẽ0,p − Ẽ0,t) / (D + ε·sign(D)), D = (Ẽ1,t − Ẽ0,t) − (Ẽ1,p − Ẽ0,p), ε = 1e-6. |D| < ε means parallel lines (β = None); β ∉ [0,1] means one horse always wins (skip). β depends only on Ẽ0, Ẽ1, never on the oracle. This was the v2.7 "3-day flip-flop bug": not a bug, just two horses crossing.

Spotlight + decision loss (training only)

d_b        = sqrt((α − sg[β])² + 1e-4)
w          = sg[1 + 2·exp(−d_b / 0.1)]                 # shine light near the cliff
s_β        = tanh((α − sg[β]) / 0.05)                   # SIGNED (no abs)
L_decision = softplus(−s_β · (α̂ − sg[β]) / 0.1)        # softplus, hinge exploded in v2.7
L_v2.8     = w·(α̂ − α)² + 0.5·L_decision

If α is left of β, α̂ must be left too, otherwise the winner flips.

Exposure-bias fix

S_mix = (1−α_mix)Ẽ0 + α_mix Ẽ1, α_mix = γ_mix·α + (1−γ_mix)·α̂, γ_mix annealed 1 → 0 over 10k steps: early training uses the stable oracle, late training matches inference (fixes the 8% drop).

Gradient isolation

• W_rel (and W_q/W_k/W_v, FFN, embeddings) learn only from task CE on S_mix.
• W_gate, c_gate learn only from L_v2.8.

This is enforced in the graph: the gate only sees detached q, k, Ẽ0, Ẽ1; the scores only use a detached α̂; β and w are detached. So one backward() of CE + L_gate equals two separate backward passes. On top of that the optimizer has two AdamW groups (3e-4 main, 1e-3 gate), clipped separately at 1.0 (a joint clip would couple them). verify_gradient_isolation() asserts it on the first batch of every run.

Files

file	what
dual_expert_attention.py	DualExpertAttention, Eq (1)-(15). python dual_expert_attention.py runs the Paris demo + isolation check
transformer.py	PositionalEncoding, EncoderLayer, TransformerModel (cls/mlm), DualExpertBert (bert-base-uncased, 768-d, 12 heads), optimizer, γ schedule
data.py	Wikipedia streaming / D_α (bAbI 16-18, CLUTRR, ConceptNet IsA/PartOf/Capitals, 1:1 negatives, 10/1/1 split), offline DeBERTa+MiniLM oracle, mock_alpha()
train.py	training loop, γ_mix annealing, α̂-vs-α histograms, checkpoints
eval.py	accuracy, α MAE/Pearson/Spearman, band agreement, photo-finish side accuracy, Paris demo
visualize.py	writes attention_analysis/ (heatmaps of Ẽ0, Ẽ1, S̃ per head, photo-finish lines, α histogram, spotlight)

Quickstart

pip install -r requirements.txt
python dual_expert_attention.py            # Paris→France α̂ ≈ 0.9 after gate fit + isolation OK
python transformer.py                      # shapes + isolation + param groups
python train.py --quick                    # CPU smoke test
python train.py --data_source dalpha --oracle mock                          # paper config, synthetic D_α
python train.py --data_source dalpha --dalpha_source real --oracle deberta --oracle_cache cache/alpha.jsonl
python train.py --data_source wikipedia --max_len 128                       # MLM from scratch
python train.py --data_source hf --batch_size 32 --lr_main 2e-5             # fine-tune BERT
python eval.py --ckpt runs/dea/checkpoint.pt
python visualize.py [--ckpt runs/dea/checkpoint.pt] [--oracle deberta]

Defaults follow §8: d_model 512, 8 heads, AdamW 3e-4 (W_rel, orthogonal) / 1e-3 (W_gate, Xavier), c_gate −0.5, clip 1.0, batch 128, seed 42.

Implementation decisions you should know about

1. W_rel is per head, [n_heads, d_head, d_head] (8×64×64 = 32k params/layer, not 262k), and both experts scale by d_head (= d_k).
2. Softmax is applied directly to S (Ẽ are already unit-scale), with no extra 1/√d.
3. Token-level oracle. The pair α is written into the [premise tokens × hypothesis tokens] block; every other cell is NaN (no label). Unlabelled cells fall back to α̂ in α_mix and are excluded from L_v2.8.
4. Gold key t. Pass target_idx; if omitted, t = oracle winner argmax_k S*(q,k;α) among labelled keys.
5. Gate loss = dense (α̂−α)² on all labelled pairs (w = 1, where β = None) + mean over valid (t,p) of the spotlight + decision term.
6. Wikipedia + mock α only exercises the plumbing (random α carries no signal). Use D_α + DeBERTa for real results.
7. Real dataset loaders (bAbI/CLUTRR/ConceptNet on the HF hub) are best-effort; hub names move. On failure each source falls back to a synthetic generator, so the run never crashes. The synthetic ConceptNet bucket is a few hundred hand-written pairs, not 4k.
8. Heatmap file names use E0tilde instead of Ẽ0 for filesystem portability.

Discrepancies found in the source documents (resolved as listed)

item	paper / model card	this code
α_mix order	model card: (1−γ)α + γα̂ (would start on α̂)	γ·α + (1−γ)·α̂, γ 1→0 (paper §6)
s_β	model card: `tanh(	α−β
ε·sign(·) in β	paper: "D = numerator diff"	sign(denominator)
E0/E1 scale	d_model	d_head
gate input / mixing	model card quickstart feeds raw E0,E1 to the gate and clamps after the linear layer	normalised Ẽ everywhere, clip logit before σ
strong band	[0.75,0.95] / [0.80,0.95] / [0.87,0.93]	mock uses [0.75,0.95]
weak band	[0.15,0.30] / [0.15,0.35]	mock uses [0.15,0.30]
licence	model card MIT, prompt CC-BY-4.0	CC-BY-4.0 here, please pick one

Limitations

D_α is small (12k), ConceptNet covers 3 relations, and the oracle inherits DeBERTa/MiniLM biases (IsA may carry gender stereotypes). Random negatives are easy, not hard negatives near β. Memory is O(B·H·T²) for several tensors (E0, E1, Ẽ0, Ẽ1, α̂, β), so it's fine for T ≤ 256; longer contexts need chunking. No d=4096 scaling test yet.

Citation

@misc{alhayali2026dualexpert,
  title  = {Learning When to Look Like vs. Work Like: A Dual-Expert Attention with a Graded Alpha Knob},
  author = {Al-Hayali, Abdulrahman}, year = {2026}, note = {v2.8.3, Mosul, Iraq. Independent researcher (azl\_y80)}
}