Overview
Keysight is at the forefront of technology innovation, delivering breakthroughs and trusted insights in electronic design, simulation, prototyping, test, manufacturing, and optimization. Our ~15,000 employees create world-class solutions in communications, 5G, automotive, energy, quantum, aerospace, defense, and semiconductor markets for customers in over 100 countries.
Our award-winning culture embraces a bold vision of where technology can take us and a passion for tackling challenging problems with industry-first solutions. We believe that when people feel a sense of belonging, they can be more creative, innovative, and thrive at all points in their careers.
About the Program
Keysight’s Applied AI Research group is developing adaptive neural modeling frameworks that enable AI systems to evolve across changing simulation, measurement, and experimental conditions. A central challenge is to design neural models that can dynamically adapt their input and output structures—adding or removing signals, channels, or targets—without full retraining or costly reconfiguration.
This internship focuses on building dynamic I/O schema mechanisms for neural networks implemented in libtorch (C++ API for PyTorch). Your work will enable AI models to flexibly adapt to new use cases, reduce retraining time, and lower compute costs while improving scalability and reusability across Keysight’s engineering AI ecosystem.
Role Overview
As a PhD Intern in Machine Learning Systems, you will design and prototype runtime-adaptive neural models capable of modifying their input/output schemas dynamically.
You will explore how libtorch-based architectures can adjust tensor structures, model heads, and data mappings without full reinitialization—enabling schema-aware, modular AI models.
You’ll collaborate with machine learning, runtime, and simulation experts to integrate these mechanisms into Keysight’s AI modeling stack, driving more efficient, versatile, and future-proof model workflows.
What This Internship Offers
- A research opportunity to develop next-generation adaptive neural architectures in C++/libtorch.
- Mentorship from experts in AI runtime systems, model architecture design, and computational modeling.
- Direct impact on improving AI model adaptability, reducing retraining costs, and enabling modular reusability.
- A portfolio-defining PhD project at the intersection of deep learning architecture, runtime engineering, and systems optimization.
Responsibilities
Key Responsibilities
- Design and implement dynamic I/O schema functionality in libtorch-based architectures, supporting runtime addition or removal of inputs and outputs.
- Develop modular components that maintain consistent state and gradient flow across evolving input/output configurations.
- Create schema translation and mapping utilities to maintain backward compatibility and incremental fine-tuning.
- Integrate schema management into training, checkpointing, and inference workflows.
- Benchmark adaptability and retraining efficiency, quantifying improvements in compute utilization and convergence time.
- Collaborate with runtime engineers to ensure performance, memory stability, and model safety under dynamic schema changes.
- Document and publish experimental results and architecture designs for internal and research dissemination.
Learning Outcomes
- Gain hands-on experience building dynamic, modular AI models in C++ using libtorch.
- Learn advanced techniques for schema evolution, runtime adaptation, and incremental retraining.
- Explore strategies for balancing architectural flexibility with numerical stability and performance.
- Collaborate within a multidisciplinary research team spanning ML, data systems, and simulation modeling.
Qualifications
Required Qualifications
- Current PhD student (or recently graduated PhD) in Machine Learning, Computer Science, Applied Mathematics, or Electrical/Mechanical Engineering.
- Strong proficiency in C/C++ and libtorch (C++ PyTorch API) for neural network implementation.
- Understanding of dynamic computation graphs, model serialization, and runtime configuration management.
- Experience designing or training modular neural architectures or runtime-adaptive ML systems.
- Familiarity with schema evolution, metadata management, or flexible I/O processing.
- Strong analytical and software engineering skills with attention to efficiency, safety, and reusability.
- Experience designing and training GNN and GCN neural architectures.
Preferred Qualifications
- Experience with dynamic-shape models using TorchScript, TensorRT, or ONNX Runtime.
- Background in graph- or operator-based architectures that support variable topologies.
- Understanding of parameter-efficient fine-tuning (PEFT), adapter layers, or meta-learning strategies.
- Experience profiling or optimizing GPU-based C++ inference and training pipelines.
Prerequisites
- Strong experience in C++/CUDA development using libtorch and modern CMake workflows.
- Familiarity with multi-threading, async I/O, and memory management for high-performance ML applications.
- Knowledge of data marshaling, tensor allocation, and layout optimization in C++.
- Competence with version control (Git), profilers, and testing frameworks.
- Commitment to creating robust, extensible systems that make neural modeling more adaptive and efficient.
