The ever-increasing use of edge devices – including Internet of Things (IoT) devices, such as smart cameras, mobile point-of-sale kiosks, medical sensors, and industrial PCs to gateways and computing infrastructure – is driving continuous growth in the amount of data generated and collected. As users continue to generate massive amounts of data, storing and using all that data in cloud data centers has begun to strain network bandwidth requirements to the limit. Additionally, with increased consumer demand for quality of service, latency concerns are also paramount and drive the need for innovative solutions.
Edge computing is a distributed computing architecture that brings computation and data storage closer to the data sources. Moving powerful edge computing closer to where data is generated opens the potential for:
- Lower latency
- Increased reliability
- Better traffic management
- Enhanced security
Intel is committed to making the intelligent Edge a reality, including helping to solve the complexities of IT and operational divides, fine-tuning and validating edge technology and software, and working with our partners to bring hundreds of deployment-ready packages to our customers. As part of this effort, Intel Labs has supported research and development in several areas, including anomaly detection, autonomous mobile robotics, and visual data management. Learn more about how these technologies are already shaping the next generation of edge services.
Anomaly Detection
Anomaly detection is typically framed as an unsupervised one-class classification problem in which a distribution of the normal (defect-free) data is learned, and anomalies are detected as deviations from this model. Supervised approaches are largely unsuitable as they require access to a sufficient number of labeled training examples. In industrial settings, this poses a challenge due to the rare appearance of anomalies and the inherent unpredictability of the specific types of anomalies that may arise during operations.
Intel's Solution
Intel Labs has developed a fast and principled method to solve the problem of visual anomaly detection. In addition to image-level detection, researchers simultaneously generate maps that provide precise pixel-level spatial localization of the anomalies within images, effectively achieving segmentation. This unsupervised approach operates under the assumption of having access solely to anomaly-free training data and can identify anomalies of an arbitrary nature on test data. The method utilizes a shallow linear autoencoder to perform out-of-distribution detection on the intermediate features generated by a pre-trained deep neural network.
More specifically, researchers compute the feature reconstruction error (FRE) and establish it as a principled measure of uncertainty. The technique is rigorously connected to the theory of linear auto-associative networks, which provides a solid theoretical foundation and enables multiple efficient implementation strategies across the continuum of computing platforms, from low-cost edge devices featuring only a CPU to on-premises and cloud servers featuring discrete GPUs. Extensive experimentation demonstrates that Intel’s method outperforms the current state-of-the-art methods; it excels in speed, robustness, and remarkable insensitivity to parameterization. Furthermore, it still performs well in the few-shot setting, where a very small number of defect-free examples are used for training. From training data, the solution can build a custom anomaly detection model in a matter of seconds that is ready to be deployed and supports quality control.
Application to Real-world Use Cases
The unsupervised anomaly detection method is especially important in the medical domain because of the high cost associated with annotating medical images. Therefore, Intel has extended the aforementioned FRE anomaly detection method for 3D medical scans. Intel’s 3D medical scans can work on various 3D medical scans, including but not limited to MRI and CT scans. The proposed solution is a pseudo-3D solution. The main advantage of this solution is that it combines the high efficiency of the 2D FRE AD solution while also using 3D information. At MICCAI 2023, Intel applied this method on the medical out-of-distribution (MOOD) challenge and ranked 4th in anomaly detection performance for the whole scan and 3rd for more challenging anomaly segmentation tasks.
Intel has applied similarly trained algorithms for defect detection on over a dozen other materials, parts, and products. The usages range from textile design defect detection to industrial assembly implementation on internal and external customer factory floors. The benchmarks are available in a publicly accessible MVTec dataset (MVTec Anomaly Detection Dataset - MVTec AD: MVTec Software).
Enabling an Edge-centric Robotics Paradigm
Autonomous Mobile Robots (AMRs) are now increasingly being used in many industrial and consumer applications, and they will be an integral part of our future. Industries are already going through a digital transformation towards connected, flexible, and intelligent factories enabled by autonomous robots. AMRs need advanced AI algorithms for perception, cognition, planning, and control and therefore need massive computing capability.
Therefore, “Edge-centric” Robotics or “Edge Robotics” is emerging as a paradigm where advanced communications and Edge computing technologies enable robots to offload their computationally heavy perception/planning/control computing functions to an on-premise Edge Server with low latency and determinism. Enabling robots to offload computing workloads to the Edge can bring advanced intelligence to constrained mobile robots, improve battery life and cost, reduce computing latency by accelerating workloads on the Edge, and improve multi-robot coordination, manageability, and scalability.
Intel Labs researchers have demonstrated the initial feasibility of time-sensitive robot control via the Edge Server and over a 5G network by prototyping an industrial use case. The following use case demonstrates a robotic arm picking up moving objects from a conveyor belt while the vision perception compute function is running on the Edge Server.
In the demonstration video, cameras in the infrastructure stream their data over a 5G network to an on-premise Edge Server. The Edge server runs the object pose perception function to extract the location and orientation of objects on the conveyer and relays this information back to the robot over a 5G network for grasping control action. Due to varying communication and computing latencies, the robot always gets delayed or stale object state. Researchers use AI and ML techniques with which robots learn to adapt to these E2E latencies by predicting the correct state of the objects from delayed states received from the Edge Server. Likewise, AI/ML techniques are used to improve network link utilization by learning to selectively drop camera frames that are non-essential for the completion of the grasping task, such as when the object is beyond the grasping range of the robot.
Visual Data Management System
Visual computing workloads performing analytics on video or image data, either offline or streaming, have become prolific across a wide range of application domains. This is partly due to the growing ability of machine learning techniques to extract information from visual data, which can subsequently be used for informed decision-making. The insights this information can provide depend on the application: a retail vendor might be interested in the amount of time shoppers spend in front of a specific product. At the same time, a medical expert might want to see the effect of a specific treatment on the size of a tumor.
Intel Labs’ Visual Data Management System (VDMS) is an open-source project designed to enable efficient access to visual data. Since visual data often contains rich metadata (such as objects, locations, and time), VDMS stores this information in a high-performance graph database. VDMS can quickly identify which data is relevant to a given query using this metadata. Additionally, VDMS uses a custom library to store and retrieve visual data, providing an interface for machine-friendly and traditional formats. These new formats are designed to support applications that are often interested in specific areas of images or videos, particularly when the individual object is large. While there are a number of big-data frameworks, systems that can be used to store metadata, and systems that manipulate a specific category of visual data, VDMS is set apart with the following aspects:
- Design for analytics and machine learning: By targeting visual data for use cases that require manipulation of visual information and associated metadata.
- Ease-of-use: By defining a common API that allows applications to combine their complex metadata searches with operations on resulting visual data, and together with full support for feature vectors, VDMS goes beyond the traditional SQL or OpenCV level interfaces that do one or the other.
- Performance: A unified system such as VDMS can outperform an ad-hoc system constructed with well-known discrete components. VDMS handles complex queries significantly better than the ad-hoc system without compromising the performance of simple queries.
Research shows that compared with a combination of industry-standard systems, VDMS demonstrates improvements of up to 364x in certain queries and an average improvement of about 85x when compared to PostgreSQL+Apache. When compared to MySQL+Apache, we see up to 96x speedup and an average improvement of 31x. The design of VDMS, which was conceived as a data management system that treats visual entities as first-class citizens, can remove inefficiencies resulting from re-purposing and combining solutions not designed for the job while providing simpler and richer interfaces. VDMS’ easy-to-use interfaces outperform industry-standard systems with a set of functionalities, which are not available in any other single data management solution for visual data.