Anyone familiar with Hewlett Packard Labs would not be surprised to learn that its interns always score in the internal competitions that Hewlett Packard Enterprise holds every year. But this year was particularly impressive, especially when you consider the size of Labs compared to many of the corporation’s departments.
This year, Labs took first, second, and third place in Best-in-Class Technical Competition at the 2021 Hewlett Packard Enterprise Bay Area Intern Fair and first place Best-in-Class at the Durham Site.
HPE as whole does not hire mediocre people and it casts an absolutely enormous net, one which covers the entire globe. But because Labs is a research and development organization, its personnel are intellectually ferocious – not to mention tons of fun.
Our interns come from all four corners of the earth. We asked them about their internships (which will come to an end in August or September), about their awards, how their work is likely to be used in the future, and how their individual research might fit together with that of their colleagues.
Josue Rivera
Best-in-Class 1st Place, Bay Area: Inverse Design of Grating Couplers Using Physics-Based Reinforcement Neural Network; finishing master’s degree in computer science at the University of Massachusetts, Dartmouth; beginning two PhD programs at Purdue University in the fall, one in aerospace engineering and another in computer science; intern at the Large Scale Integrated Photonics Lab
In his own words. There is a nano-structure called a grating coupler, whose job is to redirect light. So if light comes in into the coupler through a wave guide, then it can be redirected in a certain direction. However, designing those grating couplers is really challenging because you have to think about how light will interact with them. The team and I were working on a machine learning model, which will allow you to specify the parameters you want the grating couplers to have and then, using reinforcement learning, the neural net will create the design for, and manufacture, that device. So in the future, this research has to potential to cut a lot of costs because you won't have to spend money on design. It will also be a lot more robust because once the neural network is trained, you can easily use it to design many devices with many different properties.
My thesis discusses a novel approach for modifying and training neural networks for lifelong learning tasks by first training the networks via conventional approaches and then performing pruning and fine-tuning on the network to turn it into an embedder. This embedder can now be used to map data into a latent space which reflects detected features by the model. Now, new data can be introduced and their feature “DNA” (shared features found in the new data that existed in old ones) can help us understand the new data without the need to retrain our model. Models also have the option to be sparsely/infrequency retrained to learn new features novel in recent data. Additional novelties of the approach are that it can be applied to both Euclidian and non-Euclidian data. Non-Euclidian data often refers to graph structures such as 3D models, social networks, and protein structures.
There are so many things that I want to do. But I do know that I want some aspect of it to include machine learning. It could be reinforcement learning or general intelligence.
M. Jobayer Hossain
Best-in-Class 2nd Place, Bay Area: Development of an Automated Test Setup for Silicon Photonics; PhD candidate at the University of Central Florida’s College of Optics and Photonics, graduating in the fall; intern at the Large Scale Integrated Photonics Lab
In his own words. Once Josue and the other team members design those complicated structures and then fabricate them, it's very important to test them. So what I'm doing in my internship is developing an automated testing set-up for testing those devices and their components. It's very complicated to test because those devices are in the micrometer range and you need to align them properly and synchronization the different devices, for example, power meter, laser, detector, and your rotating stage – all of these. So you cannot really do it manually. You have to automate it.
I wrote some Python codes which helped automate the system. I also developed a graphical user interface to help the researchers in the lab who don't really have any experience in Python. They can still use that graphical user interface to do the test. So you hit run and it tests and give you the result. I hope having this capability will speed up the resource in a lab and that in the future HPE can use it for their research.
Amit Samanta
Best-in-Class 3rd Place, Bay Area: Enabling Persistent Memory for Serverless Applications; PhD student at the University of Utah; intern in the Networking and Distributed Systems Lab
Puneet Sharma, director of the Networking and Distributed Systems Lab at Hewlett Packard Labs, on Amit. Amit and his mentors Faraz Ahmed and Lianjie Cao are tackling a key challenge for serverless computing, which is how to efficiently manage huge amount of ephemeral state. They are striving to build a solution that is not only low latency to meet application service level objectives but also low cost so that it can be scaled for large volume of state to be managed. The judges appreciated the fact that the project co-optimizes the hardware and software for serverless state management.
In his own words. I'm working on a project that aims at enabling persistent memory for serverless applications. I am looking at challenges related to management of ephemeral or intermediate data of serverless applications on different storage mediums including persistent memory.
The advantages are many. Firstly, it provides a better way to load balance. Secondly, suppose somebody is trying to run a machine learning training job. In such a case, the data movement between different stages becomes a bottleneck. Persistent memory could be helpful, as it allows ultrafast access to large datasets. Also, it provides access latencies less than those of flash SSDs and is cheaper than DRAM.
I have developed a test-bed that can be used as a testing environment for deploying different applications while using persistent memory as a storage medium. This setup will allow us to develop solutions for optimizing performance of serverless applications.
Alex Shmakov
Best-in-Class 1st Place, Durham: Automatic Reinforcement Learning for Wave Energy Control; PhD student in machine learning at University of California, Irvine, studying artificial intelligence; intern in Artificial Intelligence Research Lab
In his own words. I'm doing reinforcement learning, which is a cutting edge of research. It's not terribly well known and it hasn’t been exhaustively explored. I am developing multi-agent reinforcement learning solutions in order to lower the cost of wave energy power generation while ensuring safe operation through testing.
My project involves testing and safety because we're using reinforcement learning to control wave energy power generators. This is a new, exciting form of green energy, and it has to be safe and it has to be reliable.
Curt_Hopkins
Managing Editor, Hewlett Packard Labs
