Bell Labs Internship on Cross-Model Feature Transfer

An increasing body of work has demonstrated that behavior in neural networks can often be explained and controlled through linear operations on their hidden activations. Model steering, for instance, is a technique where a specific target concept is elicited in the predictions of a generative model by adding a concept vector to its activations

Recent work suggests that these concept vectors can be transferred between models.  In this PhD internship, you will investigate the feasibility of cross-model feature transfer for improving the performance and reliability of machine learning models.
 

• You will review scientific literature

• You will design and implement a cross-model feature transfer method(s)

• Evaluate the effectiveness of the proposed technique(s) on representative problems

• Consolidate your research in a scientific publication and/or a patent application

Location: Antwerp (Belgium)  

• Student enrolled in Ph.D. Computer Science/Engineering in Machine Learning 

• Strong programming skills in Python  

• Language skills: English  

• Experience with representation engineering, mechanistic interpretability, or explainable AI is a big plus.

• A strong publication record is a big plus.