Neural Network Drop Layer for Compressed Feature Inference

The patent applies the extraction principle by introducing a drop layer that selectively removes (drops) task-irrelevant feature values from the input data. The drop layer identifies and extracts only the relevant features needed for the target task, discarding harmful task-irrelevant content. This is achieved through drop probability assignments that determine which feature values to retain and which to remove, thereby improving inference performance by eliminating harmful factors.

The patent implements local quality by assigning different drop probabilities to different feature values based on their relevance to the target task. Instead of uniformly processing all features, the system applies localized quality control where each feature value is evaluated individually. Features with high relevance receive low drop probabilities (preserved), while task-irrelevant features receive high drop probabilities (removed), creating a non-uniform, quality-based filtering mechanism.

The drop layer extracts and removes redundant or less important feature values from the input data before processing. By selectively dropping feature values based on their relevance to the target task, the system reduces the volume of data that needs to be processed by subsequent neural network layers, thereby improving computation efficiency without losing critical information needed for accurate inference.

The patent applies partial action by not preserving all feature values, but only the necessary subset required for the target task. The drop layer performs partial processing by retaining only relevant features and discarding others, achieving a balance between maintaining sufficient information for accurate results and reducing computational burden through selective feature preservation.