Challenges to Implementing Artificial Intelligence

Artificial intelligence (AI) is one of the wider trends in engineering and manufacturing applications. The AI field is nebulous with challenges that include finding enough data to train models, navigating the current data regulations, and keeping a watchful eye on data security and privacy. Securing enough computing power to run the AI system can also be a problem, and AI engineers can find themselves constantly working on these ancillary issues instead of focusing on the development of models and applications.

Hewlett Packard Enterprise recently announced two tools designed to help engineers bring the benefits of AI into a company. HPE Swarm Learning is an AI training tool designed to increase accuracy and protect data privacy at the source. The HPE Machine Learning Development System focuses on scaling up AI models and giving customers the ability to lean on HPE for systems that approach complete turnkey solutions.

Swarm Learning Shares the Knowledge but Not the Data

The basics of HPE Swarm Learning is that machine learning starts at the edge. What is learned from the data is then brought back to a centralized location without sharing the actual data points.

HPE says that this swarm method has several advantages to traditional AI learning in which data is often shared to the hub, and then insights are returned to the edge.

The largest and most obvious advantage is the sharing of training algorithms instead of training data. Sensitive private information can stay on locations, but the knowledge gains can come back to the swarm. The second advantage is a decentralized architecture in which blockchain secures what is learned and trained. Keeping the infrastructure on the HPE side allows customers to scale operations based on immediate needs. The third advantage that HPE touts is its vendor and enterprise support and services. This frees up AI engineers to focus on training AI instead of worrying about IT issues.

The Machine Learning Development System

Machine learning and AI are always going to be computation heavy. Finding the data, storing the data, working through the data and then processing the results are all important pieces of the AI process. HPE says that too many companies want the benefit of AI tools without understanding the need for AI infrastructure. The two big pillars of the machine learning development system are the hardware itself and the existing HPE AI systems.

The Apollo 6500 Gen10 Plus system is built for high-performance computing loads, outfitted with NVIDIA HGX A100 Tensor Core GPUs and AMD Instinct MI100 accelerators. The hardware uses HPE’s direct liquid cooling system and 15 80-millimeter chassis fans to regulate the heat generated from those heavy calculations. Building for the exascale era means that HPE is continuing its path with the idea that AI needs to be faster, more secure and more powerful.

The machine learning development environment is another place where this new system takes existing HPE tools and centers them around AI development. The short summary of the system lists faster model training, higher accuracy, lower costs, and more consistent methods of archiving and viewing datasets and results.

Much of this HPE AI system is built around the Determined platform, an acquisition made in June 2021. One calendar year in this current AI landscape can mean several fundamental iterations, but this 2021 acquisition is still running moderately autonomously while HPE has adopted its methods for experiment tracking, resource management and distributed training.

What Does It All Mean?

HPE is introducing these new tools as a mix of new functions and small incremental changes made to its existing product line. The AI platform landscape right now is full of changes, and many companies are trying to carve out pieces of the monetary landscape.

Engineers need to decide how to get data for an AI project. Again, several datasets can be found without a cost attached, but most require payment for access. Engineers trying to navigate these competing constraints need to look at several factors before deciding on a path.

Swarm Learning monetizes smaller peer-to-peer interactions, calling to mind BitTorrent sharing from the early 2000s. Instead of focusing on getting all the hardware and data, the work is distributed among the swarm.

The Swarm Learning announcement also feels full of possibility. Being able to take disparate data and pull it together without sharing the data itself is big, and the fact that each edge tool can learn from each other is even bigger.

The privacy concerns are addressed here, but on some level this could be a benefit to hackers. Instead of worrying about the sensitive information in one central location, engineers now need to worry about thousands of security risks at remote locations. This sounds like a challenge.

The most prominent use of Swarm Learning would be staving off another pandemic and finding solutions to big medical issues, such as Alzheimer’s disease, heart disease and cancer.

An HPE Swarm Learning white paper lists several possible benefits of this tool. Breast cancer research can be shared between the United States, Europe, Africa and Asia without bias on race or social class. Urban mobility applications could combat traffic congestion by proposing and incentivizing small changes to departure times or routes.

Deep space exploration is another area ripe for swarm learning applications. As the world gets more and more data from the James Webb Telescope, swarms of data centers can study each point at its source. As both manned and unmanned exploration ventures farther from the Earth, latency from mission control to the spacecraft becomes more of a problem. Having AI applications on location could help solve mission issues faster, saving time and keeping missions on target.

These are all exciting possibilities and, hopefully, the next few years will push engineers to even greater AI applications.