Neural Network Specification Learning
Job Description
Description
Level of qualifications required : A levels + 2 years of higher education or equivalent
Function : Temporary scientific engineer
Context
The research engineer will work in the context of an international collaboration with the University of Toronto, Canada. This collaboration aims at introducing novel ways for specifying, verifying, and explaining machine learning models. The overarching research hypothesis behind the project is that reasoning about large, complex, state-of-the-art models can be achieved through proper abstractions built on top of their internal representations.
Assignment
The engineer will work on designing and developing a new methodology for learning mathematically rigorous, logic-based, and expressive specifications over the latent features of a given machine learning model. Building on preliminary results (Gopinath et al. 2019, Geng et al. 2023), they will explore a broader class of neural specifications.
Main activities
This investigation will focus on innovation and generalization across three key aspects:
- Neural Predicate Design: Previous studies primarily considered neuron-level abstractions using basic binary predicates (active or inactive), which are limited in expressiveness. We will investigate predicates with higher arities—binary or ternary relations—and introduce neural predicates over attention heads for transformer architectures.
- Specification Structure: Current specification structures, such as those in VNN-COMP, are limited to narrow regions around a single input. We will generalize this by redesigning the specification structure into sets of Horn clauses over neural predicates, enabling a more expressive and flexible language for neural network specifications.
- Specification Learning Algorithms: We will explore a statistical approach that collects inference traces, computes distributions for neural predicates, and composes candidate specifications. This approach will be refined to reduce redundancy and incorporate insights from formal methods and machine learning.
Benefits package
- Partial reimbursement of public transport costs
- 7 weeks of annual leave + 10 days off due to RTT + possibility of exceptional leave
- Teleworking (after 6 months) and flexible hours
- Professional equipment (videoconferencing, computers, etc.)
- Social, cultural, and sports activities
- Theme/Domain : Proofs and Verification, Software engineering (BAP E)
Warning
You must enter your e-mail address to save your application on Inria. Applications must be submitted online via the Inria website. Applications sent through other channels may not be processed.
Instruction to apply
Defence Security : This position may be in a restricted area (ZRR), as per Decree No. . Authorization is granted by the unit director following a favourable Ministerial decision. An unfavourable decision would cancel the appointment.
Recruitment Policy : Inria values diversity and encourages applications from people with disabilities.
Inria is the French national research institute dedicated to digital science and technology, employing 2,600 people across numerous disciplines and international collaborations.
#J-18808-Ljbffr
Company
Inria
Location
Toronto
Country
Canada
Salary
100.000
URL