Job Description

Join our Team

About this opportunity:

Ericsson Research’s Artificial Intelligence Research Area combines machine learning and reasoning to enable intelligent, autonomous operations in complex telecom systems.

We are looking for a motivated student to study Latent Neural-Operator Structural Causal Models.

Causal representation learning asks whether hidden causal variables can be recovered from observed measurements. Existing theory often assumes scalar variables and unstructured mixing. In physical systems, causal variables may instead be functions defined on different domains, while the measurement process has known structure.

A radio link is a clear example. Channel response, interference, and the equalised constellation are function-valued and not directly observable. Receiver IQ samples and channel estimates combine them with hardware impairments and transmitter choices.

This project will study a latent structural causal model for such systems. Function-valued latent variables are connected through operator mechanisms, while the measurement chain links latent structure to observed data. The structure is known, the mechanisms are not, and system actions are logged with known targets.

The thesis will assess whether the recovered latent structure is meaningful, not merely whether it reconstructs measurements. Two criteria will be evaluated:

• Identifiability: Do the recovered latents correspond to the true variables, and are they determined by the measurement model, logged interventions, or both? For function-valued variables, ambiguity may involve transformations of entire function spaces.
• Compositional generalisation: Can a causal-graph model predict untested combinations of interventions better than an unstructured model?

You will implement the framework in a link-level simulator with known ground truth and compare it with causal representation learning baselines. Negative results are valid outcomes; showing that the latent structure is not meaningful or that the graph provides no benefit is also valuable.

What you will do:

• Review causal representation learning and identifiability literature.
• Build a controlled physical-layer simulator with known ground truth.
• Implement the proposed model and comparison baselines.
• Evaluate identifiability and prediction of untested interventions.
• Present findings in regular discussions and the final thesis, including negative results.

The skills you bring:

• A largely completed master’s degree with strong academic performance.
• Knowledge of machine learning, linear algebra, and probability.
• Proficiency in Python and experience with PyTorch or JAX.
• Ability to work independently and use AI coding assistants effectively.
• Excellent written and spoken English and teamwork skills.
• Interest in causal inference or signal processing is beneficial.

We expect an analytical, research-oriented mindset, the ability to learn quickly, and the initiative to identify problems and solutions.