Hierarchical Tensor Encoding for Adaptive CNN Edge Compression

The patent applies preliminary action by pre-training the bottleneck encoder and decoder layers offline with diverse input data statistics. This pre-training establishes adaptive weights that anticipate various input conditions, reducing the need for real-time adaptation and minimizing processing delays during actual CNN inference on edge devices.

The patent implements dynamics by designing a trainable bottleneck encoder and decoder with adaptive weights that can be refined through training. This dynamic structure allows the compression system to adapt to different input data statistics and CNN architectures, optimizing the balance between compression efficiency and processing speed for distributed CNN execution.

The patent applies segmentation by dividing the CNN processing pipeline into distinct segments: a first portion executed on the edge device with a trainable bottleneck encoder, and a second portion executed on the cloud server. This segmentation allows the bottleneck encoder to be specifically optimized for CNN tensor data characteristics while leveraging standard video compression for the bottlenecked representation, achieving both adaptability and compression efficiency.

The patent implements parameter changes by transforming the CNN tensor data through a trainable bottleneck encoder that modifies key parameters (dimensionality, feature representation) before applying standard video compression. The adaptive weights in the bottleneck encoder adjust compression parameters based on input data statistics, enabling VVC to effectively compress diverse CNN tensors while maintaining adaptability to different data characteristics.

The patent applies the intermediary principle by introducing a bottleneck encoder and decoder as intermediary components between the edge device and cloud server. These intermediaries handle the complex adaptive transformation of tensor data, converting high-dimensional CNN features into a compressed bottleneck representation that can be efficiently transmitted and then reconstructed, isolating the complexity from the main CNN execution path.

The patent implements copying by creating a simplified bottleneck representation of the original tensor data. The bottleneck encoder produces a compressed copy of the essential features that can be transmitted with reduced bitrate, and the decoder reconstructs an approximation of the original tensors. This copying approach maintains adaptability while reducing the computational burden on edge devices compared to processing full-resolution tensors.