17817552. DESPARSIFIED CONVOLUTION FOR SPARSE ACTIVATIONS simplified abstract (QUALCOMM Incorporated)
Contents
DESPARSIFIED CONVOLUTION FOR SPARSE ACTIVATIONS
Organization Name
Inventor(s)
Jamie Menjay Lin of San Diego CA (US)
Jian Shen of San Diego CA (US)
Fatih Murat Porikli of San Diego CA (US)
DESPARSIFIED CONVOLUTION FOR SPARSE ACTIVATIONS - A simplified explanation of the abstract
This abstract first appeared for US patent application 17817552 titled 'DESPARSIFIED CONVOLUTION FOR SPARSE ACTIVATIONS
Simplified Explanation
Certain aspects of the present disclosure provide techniques for desparsified convolution. This involves receiving an activation tensor and generating a convolution output for the activation tensor. The technique includes selecting a subset of weight elements from a weight tensor, which correspond to a set of non-zero elements in the activation tensor. The set of non-zero elements and the set of weight elements are then multiplied.
- Techniques for desparsified convolution
- Activation tensor is received
- Convolution output is generated for the activation tensor
- Subset of weight elements is selected from a weight tensor
- Subset corresponds to non-zero elements in the activation tensor
- Non-zero elements and weight elements are multiplied
Potential Applications:
- Image and video processing
- Natural language processing
- Signal processing
- Machine learning and deep learning algorithms
Problems Solved:
- Sparse convolution can be computationally expensive
- Sparse convolution may require specialized hardware
- Sparse convolution can lead to memory inefficiency
Benefits:
- Improved computational efficiency
- Reduced memory requirements
- Compatibility with existing convolution algorithms
- Potential for faster and more accurate processing
Original Abstract Submitted
Certain aspects of the present disclosure provide techniques for desparsified convolution. An activation tensor is received, and a convolution output is generated for the activation tensor, comprising: selecting a subset of weight elements, corresponding to a set of non-zero elements in the activation tensor, from a weight tensor, and multiplying the set of non-zero elements and the set of weight elements.
