Joint Workshop of KI Absicherung and KI Wissen sheds light on DNN interpretability
Following an internal workshop in May, project partners of both KI Absicherung and KI Wissen plan to discuss DDN interpretability in a joint workshop reinforcing their collaboration. Goal of the workshop is to set a common ground for further work on this important topic within the KI Familie.
A common topic between the sibling projects KI Absicherung and KI Wissen is deep neural network (DNN) interpretability. Amongst others, it is needed to enable human inspection of a DNN’s inner workings, and verification of requirements formulated in formal or natural language like “a head usually belongs to a pedestrian”. One sub-topic of interpretability that both projects are working on is the examination and analysis of the knowledge encoded within DNNs, more precisely in their latent space intermediate output.
Thus, on May 4, KI Absicherung featured an internal workshop on this topic of DNN latent space interpretability, organized by Continental. The 3h session featured introductions to six diverse methods currently developed by partners in KI Absicherung, ranging from black-box to white-box approaches, and post-hoc to inherent interpretability methods.
Some of the discussed work already has been published, e.g. on the SAIAD 2021 workshop collocated with CVPR’21:
- https://compvis.github.io/iin/
- https://compvis.github.io/invariances/
- https://compvis.github.io/net2net/
- https://arxiv.org/abs/2105.06727
Due to the great success and appreciation of the first workshop and the interest both in KI Wissen and KI Absicherung, the topic of DNN latent space interpretability will be followed-up in a joint workshop of KI Wissen and KI Absicherung. This is intended to foster the scientific exchange as well as collaboration between the project partners both within and between the projects. It will be an extended version of the first one with more speaker time for the by now 6 submissions.
A submission of a future workshop to be collocated with a scientific conference is in prospect, in order to broaden the audience of scientific exchange. We are looking forward to further feedback and fruitful discussions (contact person: Gesina Schwalbe, gesina.schwalbe(at)continental-corporation.com, Continental AG).
Image: Unsplash / Dylan Gillis