Smart Engines says new method boosts neural network efficiency by 40%

Scientists from the computer vision software company Smart Engines have announced they have found a way to improve the efficiency of neural networks by 40 percent. The method is based on a new quantization scheme.

Currently, deep neural networks are typically executed on specialized graphics cards, and often require more computing power than most devices have due to their increasing complexity. All user devices do, however, have a central processor, which is the world standard for 8-bit neural networks.

Smart Engines researchers have proposed 4.6-bit networks, which can operate 40 percent faster than the 8-bit model with almost no decrease in quality. It achieves this by more efficiently using the features of CPUs in mobile devices.

The input data and coefficients of the model are quantized to produce an output small enough to fit into 8-bit registers.

A two-level system of 16 and 32 bit accumulators summarize the results, and an average of 4.6 bits of information is assigned to each value.

The quantization method differs from existing methods as it allows for data input bit size to be adjusted and is not bound to powers of two. It can therefore produce higher recognition quality than counterparts like 4-bit models.

Smart Engines products are already using the networks for object detection and text recognition.

PAD, mobile document recognition, among 2023 research milestones

This advancement follows after a busy 2023 for the company. Last year, it published 41 science papers, obtained 6 patents, presented over 20 reports at various conferences, and made several discoveries.

The company’s scientists presented their 20+ reports across three international conferences: the 17th International Conference on Document Analysis and Recognition, the London International Meeting, and the 16th International Conference on Machine Vision.

Their public annotated corpus of the document images MIDV-Holo allows inventors everywhere to teach their algorithms to detect and prevent biometric replay and presentation attacks.

Researchers also presented work in regards to the rapid detection of MRZ for document recognition on mobile devices using a neural network architecture that identifies machine-readable zones in 16 milliseconds on an iPhone SE that uses a processor released in 2019. The detection rate is 62 frames per second. The Smart ID Engine system uses this architecture, entering passport data 20 times faster with twice the accuracy in comparison to a qualified operator.

A method for document identification is among the six patented inventions.

Article Topics

computer vision  |  neural networks  |  object detection  |  OCR  |  Smart Engines