17895857. Spatially Preserving Flattening in Deep Learning Neural Networks simplified abstract (International Business Machines Corporation)

Spatially Preserving Flattening in Deep Learning Neural Networks

Organization Name

International Business Machines Corporation

Inventor(s)

Tanveer Syeda-mahmood of Cupertino CA (US)

Neha Srivathsa of Mercer Island WA (US)

Raziuddin Mustafa Mahmood of Cupertino CA (US)

Spatially Preserving Flattening in Deep Learning Neural Networks - A simplified explanation of the abstract

This abstract first appeared for US patent application 17895857 titled 'Spatially Preserving Flattening in Deep Learning Neural Networks

Simplified Explanation

The patent application describes techniques for spatially preserving flattening in deep learning neural networks.

• Predictor generates image feature maps from convolutional layers of a neural network.
• Auto-encoder produces encodings of the image feature maps that preserve location and shape information of objects.
• Flattener concatenates the encodings to form a spatially preserving flattened encoding vector.

• • Potential Applications:**
• Image recognition and classification tasks
• Object detection in computer vision
• Natural language processing for preserving spatial information in text data

• • Problems Solved:**
• Maintaining spatial information in feature maps during flattening process
• Improving accuracy of deep learning models by preserving object shapes and locations

• • Benefits:**
• Enhanced performance in tasks requiring spatial information preservation
• Improved interpretability of neural network predictions
• Potential for better generalization to new data instances

Original Abstract Submitted

Techniques for spatially preserving flattening in deep learning neural networks are provided. In one aspect, a spatially preserving flattening module includes: a predictor for generating image feature maps from at least one convolutional layer of a feature extraction phase of a deep learning neural network applied to input image data; an auto-encoder for producing encodings of the image feature maps that preserve location and shape information associated with objects in the input image data; and a flattener for concatenating the encodings of the image feature maps to form a spatially preserving flattened encoding vector. A deep learning neural network that includes the present spatially preserving flattening module is also provided, as is a method for spatially preserving flattening.