Case Study: How Intertech Is Developing Neural Networks to Identify Electronics Components for an International, Multi-Billion Dollar Distributor

by Tom Helvick

From time to time, we like to share case studies of what we’ve been working on here at Intertech. It gives insight into the type of work we do, what you can expect if you work with Intertech to solve your software needs, and what it’s like to work as a consultant for Intertech on challenging projects.

In our most recent case study, we take a look at the initial work an Intertech consultant has done in the field of neural networks and computer vision. Specifically, the client wants an application that, given a photo of an electronics component, can identify a match or replacement part for that component. Making this vision a reality involves training a neural network on millions of images to match parts to the client’s catalog of over 8 million components. This project is still ongoing, but initial prototypes have shown great promise.

Creating, Training, & Refining Neural Networks

The initial trials and research phases of this project involve training a neural network on a small subset of components from the client’s catalog. To do so, our consultant and the in-house expert team took 96–192 images of each component in the target sample. Then, they augmented each of those photos into various versions with blur, distortion, and lighting alterations. Ultimately, each component had 500–1000 matching images on which the researchers trained the neural network.

Each training session could take hours, and at first, the model’s accuracy was not very high. Over time, with consecutive training rounds and tweaks to the model, the neural network’s accuracy has improved on the subset of components. Now, the project team is looking to expand the problem set to include a wider range of components.

Data Infrastructure

Of course, scaling a neural network comes with its own challenges in terms of data engineering. Our Intertech consultant has worked to make model training as efficient as possible so that future refinements of the model don’t take days or months to complete training.

This is a major undertaking, as the client’s catalog includes 8 million components, each of which will have 96–192 images. Those hundreds of millions of images need to be stored somewhere. But they also need to be retrieved rapidly, as I/O bounds could limit the speed of training the model. To handle these challenges, our consultant assisted with building on-premise hardware infrastructure for the neural network training, including high-end GPUs and storage media.

Learn More

To find out more about our project with this client, check out our newly released case study. To hear about other solutions we’ve implemented for past clients in wide-ranging industries, dive into our past case studies.

Find out how Intertech’s consulting services can help your business. Get in touch today to learn more.

Originally published at https://www.intertech.com on August 28, 2019.