DeepFocus-BP: Error-Aware Adaptive Backpropagation via Dynamic Alpha-Beta Routing (Achieving 66% FLOPs Reduction with Improved Accuracy) - SOTA NLP Confirmed v3. (Resnet FAIL)
Fuente:
Zenodo
Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Recurso digital |
| Sprache: | Englisch |
| Veröffentlicht: |
Zenodo
2026
|
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866901555240763392 |
|---|---|
| author | Fernandes, Cláudio |
| author_facet | Fernandes, Cláudio |
| contents | <p>This technical note presents DeepFocus-BP, a novel adaptive backpropagation algorithm that dynamically routes computational resources based on real-time error magnitude. By employing stochastic probing and adaptive Alpha-Beta thresholding, the algorithm categorizes network blocks into Skip, Full Precision, and Low Precision regimes.</p> <p>KEY RESULTS (Seed 42, IMDB Dataset):<br>- Test Accuracy: 84.10% (Surpassing Dense Baseline by +3.3%)<br>- Computational Efficiency: ~66.2% reduction in total FLOPs.<br>- Mechanism: Selective gradient suppression acts as a powerful regularizer.</p> <p>COMMERCIAL POTENTIAL:<br>This technology offers significant potential for reducing cloud infrastructure costs and energy consumption in large-scale AI training. The author is actively seeking strategic partnerships, licensing agreements, or co-founding opportunities to commercialize this algorithm.</p> <p>COPYRIGHT:<br>© 2026 Cláudio Fernandes. All Rights Reserved. Unauthorized commercial use is prohibited without explicit written permission.</p> <p>Contact for partnerships: benficaizeda306@gmail.com</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19415887 |
| institution | Zenodo |
| language | eng |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | DeepFocus-BP: Error-Aware Adaptive Backpropagation via Dynamic Alpha-Beta Routing (Achieving 66% FLOPs Reduction with Improved Accuracy) - SOTA NLP Confirmed v3. (Resnet FAIL) Fernandes, Cláudio <p>This technical note presents DeepFocus-BP, a novel adaptive backpropagation algorithm that dynamically routes computational resources based on real-time error magnitude. By employing stochastic probing and adaptive Alpha-Beta thresholding, the algorithm categorizes network blocks into Skip, Full Precision, and Low Precision regimes.</p> <p>KEY RESULTS (Seed 42, IMDB Dataset):<br>- Test Accuracy: 84.10% (Surpassing Dense Baseline by +3.3%)<br>- Computational Efficiency: ~66.2% reduction in total FLOPs.<br>- Mechanism: Selective gradient suppression acts as a powerful regularizer.</p> <p>COMMERCIAL POTENTIAL:<br>This technology offers significant potential for reducing cloud infrastructure costs and energy consumption in large-scale AI training. The author is actively seeking strategic partnerships, licensing agreements, or co-founding opportunities to commercialize this algorithm.</p> <p>COPYRIGHT:<br>© 2026 Cláudio Fernandes. All Rights Reserved. Unauthorized commercial use is prohibited without explicit written permission.</p> <p>Contact for partnerships: benficaizeda306@gmail.com</p> |
| title | DeepFocus-BP: Error-Aware Adaptive Backpropagation via Dynamic Alpha-Beta Routing (Achieving 66% FLOPs Reduction with Improved Accuracy) - SOTA NLP Confirmed v3. (Resnet FAIL) |
| url | https://doi.org/10.5281/zenodo.19415887 |