Sabitlenmiş Tweet

PsiSCN v6e: We replaced Transformer attention with a coherent physical field — now with longer sequences + distractors.
1/ Upgraded the experiment:
seq_len = 32 (+33%)
1 distractor token
Same CoherenceAttn (1D field ψ evolving with Laplacian, double-well potential αψ + βψ³, thermal noise, energy coupling).
Still no dot-product. No softmax over Q/K. Just real field dynamics.
2/ Trained only via imitation (MSE on weights + context) from a standard teacher.
After 20 epochs:
Cosine sim. (weights) → 0.9701
Cosine sim. (context) → 0.9897
3/ Zero-shot swap (replace attention module without any label fine-tuning):
Original teacher: 87.07%
CoherenceAttn (zero-shot): 88.23% → Gap +0.0117 (even better than previous version)
4/ After quick 12-epoch adaptation (CoherenceAttn + classifier only): → 99.97%val_acc (basically solved the task)
The field not only imitates the classic attention — it outperforms it on a harder setting.
5/ This is strong evidence that the core computation of attention can be distilled into a physically-grounded dynamical system that is more robust and scalable.
Full public Colab (one-click): colab.research.google.com/drive/1QglX81p…
(just change the script to v6e_stageC_len32_d1.py if you want to reproduce)
Code + plots + summary JSON available in the outputs folder.
Next? Scaling to 128+ tokens and real reasoning benchmarks.
What do you think?
#PsiSCN #NeuralFields #AlternativeAttention #PhysicsInspiredAI #xAI

English





















