18203552. DYNAMIC NEURAL NETWORK MODEL SPARSIFICATION simplified abstract (NVIDIA Corporation)

From WikiPatents
Jump to navigation Jump to search

DYNAMIC NEURAL NETWORK MODEL SPARSIFICATION

Organization Name

NVIDIA Corporation

Inventor(s)

Jose M. Alvarez Lopez of Mountain View CA (US)

Pavlo Molchanov of Mountain View CA (US)

Hongxu Yin of San Jose CA (US)

Maying Shen of Santa Clara CA (US)

Lei Mao of San Jose CA (US)

Xinglong Sun of Menlo Park CA (US)

DYNAMIC NEURAL NETWORK MODEL SPARSIFICATION - A simplified explanation of the abstract

This abstract first appeared for US patent application 18203552 titled 'DYNAMIC NEURAL NETWORK MODEL SPARSIFICATION

Simplified Explanation

Machine learning is a process that learns a neural network model from a given dataset, where the model can then be used to make a prediction about new data. In order to reduce the size, computation, and latency of a neural network model, a compression technique can be employed which includes model sparsification. To avoid the negative consequences of pruning a fully pretrained neural network model and on the other hand of training a sparse model in the first place without any recovery option, the present disclosure provides a dynamic neural network model sparsification process which allows for recovery of previously pruned parts to improve the quality of the sparse neural network model.

  • Model sparsification is a compression technique used to reduce the size, computation, and latency of a neural network model.
  • The present disclosure introduces a dynamic neural network model sparsification process that allows for recovery of previously pruned parts to enhance the quality of the sparse neural network model.

Potential Applications

The technology of dynamic neural network model sparsification can be applied in various fields such as:

  • Image recognition
  • Natural language processing
  • Autonomous vehicles
  • Healthcare diagnostics

Problems Solved

The technology of dynamic neural network model sparsification addresses the following issues:

  • Reducing the size, computation, and latency of neural network models
  • Avoiding negative consequences of pruning fully pretrained models
  • Improving the quality of sparse neural network models

Benefits

The benefits of dynamic neural network model sparsification include:

  • Enhanced model efficiency
  • Improved prediction accuracy
  • Faster inference times

Potential Commercial Applications

The technology of dynamic neural network model sparsification has potential commercial applications in:

  • Cloud computing
  • Edge devices
  • IoT devices

Possible Prior Art

One possible prior art in the field of neural network model compression is the use of quantization techniques to reduce the precision of weights and activations in a neural network, thereby reducing the model size and computation requirements.

Unanswered Questions

How does the dynamic neural network model sparsification process compare to other model compression techniques in terms of efficiency and accuracy?

The article does not provide a direct comparison between dynamic neural network model sparsification and other model compression techniques.

What are the limitations of the dynamic neural network model sparsification process in terms of scalability and complexity?

The article does not discuss the scalability and complexity limitations of the dynamic neural network model sparsification process.


Original Abstract Submitted

Machine learning is a process that learns a neural network model from a given dataset, where the model can then be used to make a prediction about new data. In order to reduce the size, computation, and latency of a neural network model, a compression technique can be employed which includes model sparsification. To avoid the negative consequences of pruning a fully pretrained neural network model and on the other hand of training a sparse model in the first place without any recovery option, the present disclosure provides a dynamic neural network model sparsification process which allows for recovery of previously pruned parts to improve the quality of the sparse neural network model.

Retrieved from "http://wikipatents.org/index.php?title=18203552._DYNAMIC_NEURAL_NETWORK_MODEL_SPARSIFICATION_simplified_abstract_(NVIDIA_Corporation)&oldid=53763"