Patrick Henriksen stands next to poster of his work at IJCNN 2023

Patrick Henriksen presents paper, ‘Verification-friendly Networks: the Case for Parametric ReLUs’, at IJCNN 2023

14th August 2023 | News

News > Patrick Henriksen presents paper, ‘Verification-friendly Networks: the Case for Parametric ReLUs’, at IJCNN 2023

Patrick Henriksen (2019 cohort) presented a paper, ‘Verification-friendly Networks: the Case for Parametric ReLUs’, at the 36th International Joint Conference on Neural Networks (IJCNN23) at the Gold Coast, Australia. IJCNN is the premier international conference in the area of neural networks theory, analysis and applications, and drew hundreds of researchers from all over the world. 

Patrick explains the focus of the paper: “Neural Networks are state-of-the-art machine learning models applicable to a wide range of use cases; however, they have been shown to be fragile to even tiny disturbances to their inputs. Formal verification methods may be applied to identify such fragilities, yet current methods are computationally expensive and, thus, have limited applicability for the larger networks often used in practical applications. Our paper proposes small architectural and training adjustments for neural networks to ease formal verification. In experiments, the resulting networks are computationally significantly cheaper to verify, thus facilitating the verification of larger networks”. 

It was Patrick’s second in-person conference and he found the experience highly rewarding. He said, “Engaging in discussions with peers sparked fresh ideas and introduced several opportunities for future collaborations. The Gold Coast’s stunning beaches also offered relaxation during our downtime”. 

Patrick really enjoys being part of the STAI CDT and said, “Through the CDT, I have access to a range of seminars, retreats and other activities that complement my research. Moreover, it is also very motivating to be a part of a community of students in similar positions as myself”. 

You can hear more from Patrick and one of the co-authors of the paper, Francesco Leofante, here about how the paper contributes to research in the area of safe neural networks.