https://github.com/saeidemadi/nns
NNS : Neural network surgery | academic assignment
https://github.com/saeidemadi/nns
ann artificial-neural-networks compression-algorithm compression-models compressions pruning pruning-algorithms pruning-structures
Last synced: 11 months ago
JSON representation
NNS : Neural network surgery | academic assignment
- Host: GitHub
- URL: https://github.com/saeidemadi/nns
- Owner: saeidEmadi
- License: gpl-3.0
- Created: 2024-04-24T01:52:33.000Z (almost 2 years ago)
- Default Branch: main
- Last Pushed: 2024-06-15T12:48:51.000Z (over 1 year ago)
- Last Synced: 2025-01-26T19:46:45.900Z (about 1 year ago)
- Topics: ann, artificial-neural-networks, compression-algorithm, compression-models, compressions, pruning, pruning-algorithms, pruning-structures
- Language: Jupyter Notebook
- Homepage:
- Size: 13 MB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# NNS : Neural network surgery
Transfer learning has emerged as a pivotal approach in machine learning, enabling models to leverage knowledge from one domain
and apply it to another, often related, domain. Methods like : Pre-trained Models, Feature Extraction,Fine-tuning
,Multi-task Learning,etc. The use of these methods due to domain dependence, complete and complex transfer of learning
causes lack of understanding of the model, high computational cost, negative transfer and very limited application of our
model. With neural network surgery (NNS) and its in-depth investigation, we provide general knowledge about the specific learned
topic with full coverage of the topic without bias and inter-neural interference in the process of learning transfer. In this method,
by taking a deep look at the events in the memory during calculations, as well as labeling them, stimulating the neural network
case by case, and comparing the results, we tried to discover and limit the network to that particular label. By removing the less
important connected neurons to the rest of the labels, we have been able to prevent overfitting and also prepare general learning for
transfer. By performing surgery on the neural network that has learned the MNIST, we have extracted learning in a special
way to recognize ”3” labels. Finally, it is possible to transfer learning to the next generations in a simple and easy way and to
use it for continuous learning of special labels, and it is also possible to mention the transfer of learning special labels from
very complex models with many labels to small models with High accuracy In this research, new interesting results about the
compression of neural network models are mentioned.
In this article, we tried to perform deep surgeries on neural networks and very interesting results have been obtained
We suggest you take a look [Paper Direct Link](https://github.com/saeidEmadi/NNS/blob/main/NNS.pdf)
> [!NOTE]
> This article and writing is only an academic assignment and has no confirmed scientific validity (it should be noted that the copyright law includes this assignment as well)
Thanks to Zahra.D