Efficient Topology Coding and Payload Partitioning Techniques for Neural Network Compression (NNC) Standard

Abstract

A Neural Network Compression (NNC) standard aims to define a set of coding tools for efficient compression and transmission of neural networks. This paper addresses the high-level syntax (HLS) of NNC and proposes three HLS techniques for network topology coding and payload partitioning. Our first technique provides an efficient way to code prune topology information. It removes redundancy in the bitmask and thereby improves coding efficiency by 4–99% over existing approaches. The second technique processes bitmasks in larger chunks instead of one bit at a time. It is shown to reduce computational complexity of NNC encoding by 63% and NNC decoding by 82%. Our third technique makes use of partial data counters to partition an NNC bitstream into uniformly sized units for more efficient data transmission. Even though the smaller partition sizes introduce some overhead, our network simulations show better throughput due to lower packet retransmission rates. To our knowledge, this the first work to address the practical implementation aspects of HLS. The proposed techniques can be seen as key enabling factors for efficient adaptation and economical deployment of the NNC standard in a plurality of next-generation industrial and academic applications.