AI Model Optimization for Smart Manufacturing

Read how Deeplite provided an optimized model zoo for smart manufacturing applications in partnership with a leading brand in IoT intelligent systems.

READ FULL CASE STUDY
>10x DNN Model Compression
Model Compression

Using Deeplite's industry-leading optimizer, your trained models can become over 10x smaller and 5x faster on your edge device.

+30 DNN Models Zoo
DNN Models

Deeplite's Optimized model zoo features over 30 SOTA classification, detection and segmentation models.

Problem Statement

Deep Neural Networks (DNNs) inside your AI can be large, process-heavy and power-hungry. Moreover, they might require obtaining domain expertise to be deployed on your Automated Optical Inspection (AOI) cameras, robotic arms, smart glasses, or other edge devices. These factors create barriers to enabling machine vision AI applications at the edge and can be costly to your digitalization program and manufacturing process.

Deeplite offers techniques that enable easy-to-use and automated AI development and implementation workflows, critical for a fast AI implementation. Our software solution truly supports you in widespread machine vision AI deployments, as we improve the energy-efficiency and throughput of your DNNs while maintaining the accuracy of your AI applications and systems.

Challenges

calculatorHigh Computational Complexity

Billions of expensive floating-point operations for each input classification are needed.

floppy diskMemory Footprint

Huge amounts of weights and activations with limited on-chip
memory and bandwidth. Which results in Data Center & Data Transfer costs in the $Ms.

batteryPower Consumption

Deep learning requires significant resources and can easily consume battery life.

thumbs downManual Trial & Error

Designing accurate and efficient models requires domain expertise, time and considerable
effort to create acceptable results.

Solution

Smart Manufacturing Model Zoo

The smart manufacturing model zoo includes a set of 30+ pre-trained deep learning models for various classification, detection and segmentation-based applications such as person detection, automated optical inspection, defect detection, mask detection and more. This enables the easy and efficient deployment of AI-powered solutions on various hardware backends (GPU, x86 CPU, ARM CPU, RISC-V CPU, MCU, Accelerators) with optimized model inference.

Performance Benchmarks

We have tested our optimizing engine on several off-the-shelf models available in Deeplite’s Neutrino Zoo™ with impressive results. These models may be purposed and trained for specific applications, like person detection or defect detection, general-purpose classifiers, object detection and segmentation models.

READ FULL CASE STUDY