Ten years ago, no one would have expected that neural networks would deliver such impressive results on computer vision problems. Since 2012, when AlexNet, a deep convolutional neural network, produced a significantly lower error rate on the ImageNet Large Scale Visual Recognition Challenge than traditional feature extraction approaches, investigation of neural network approaches for visual understanding has intensified. These efforts culminated in 2015 when a deep residual network (ResNet) bested typical human performance by achieving an ImageNet error rate of 3.6%. These and other milestones in the success of deep learning techniques have spurred a flurry of investment in algorithms, software and hardware.
Even in the bastion of traditional computer vision, automated visual inspection systems for manufacturing, deep learning is gaining influence: earlier this year, Cognex, a leader in machine vision, acquired ViDi Systems SA, a Swiss developer of software that uses deep learning techniques for identification of manufacturing defects. In its consulting practice, BDTI is seeing a considerable increase in inquiries about the use of deep learning. In some of these cases, the client had the idea that a deep neural network would provide all of the functionality required. In some of these cases, once we dug deeper into the applications, we found that a more nuanced approach was called for.
Recently, for example, BDTI engaged with an infrastructure management solution provider. The company had a challenging problem that they expected to solve with deep learning. From the requirements, BDTI identified several key capabilities needed to create a solution, including several machine perception tasks. Upon careful analysis of these tasks, it became clear that some were well-suited for deep learning solutions while others were not.
One key challenge in creating deep learning solutions is the need for extensive training data. Training data determines the success or failure of a solution based on deep learning; neural network-based systems can only adapt to new situations when they have been sufficiently trained.
In addition, deep learning algorithms tend to consume quite a bit of computation and memory—too much for some cost- and power-sensitive applications. For its infrastructure management client, BDTI proposed a hybrid approach that combined deep learning, image processing, and traditional computer vision techniques. The resulting design can be deployed despite insufficient training data for some tasks; moreover, it will continue to learn and improve after deployment. Key to the success of this engagement was BDTI's understanding of the strengths and limitations of deep learning and its ability to design an appropriate solution using all the tools—old and new—in its toolbox.
If your company is considering the use of deep learning techniques in existing and new products and systems, BDTI can make sure you make the right design decisions. Contact Jeremy Giddings for a confidential consultation by phone at +1 (925) 954-1411 or via BDTI’s website.
Add new comment