Acting on Machine Inference

Systems that model their users and adapt to them in real time need an estimate of user
state that holds up outside the laboratory, and adaptation policies that support a
person rather than destabilise them.

Estimation in use is a different regime, not a harder version of the same one. In the
laboratory we average over hundreds of repetitions, we know when each stimulus
occurred, and we analyse the recording afterwards; in use there is a single pass, no
trigger, drifting electrodes and a subject who moves. We are building estimation that
works under those constraints — causal, without averaging, inside a latency budget —
at present on the real-time analysis of somatosensory evoked potentials along the
nervous system. Real-time estimation of external state such as head and eye movement
has been part of our work for years and provides the engineering ground for it.

Adaptation raises a problem that measurement does not have. A system that responds to
rising cognitive load reduces that load, and with it the signal from which the
estimate was drawn. We also hold that users should be able to recognise how a system
has understood them and to correct it, which is a requirement on the interface and not
only on the estimator.

J. C. Redlich, P. Kloke, and S. Bosse. ‘DMFO: A Modular Alignment Architecture for Situationally Interpretable State Representations’. In: Proceedings of the 25th InternationalSemantic Web Conference (ISWC) (Bari, Italy). Nov. 2026

M. Gerstenberger, T. Wiegand, P. Eisert, and S. Bosse. ‘But That’s Not Why: Inference Adjustment by Interactive Prototype Revision’. In: Progress in Pattern Recognition, Image Analysis, Computer Vision, and Applications: Proceedings of the 26th Iberoamerican Congress on Pattern Recognition. Ed. by V. Vasconcelos, I. Domingues, and S. Paredes. Vol. 14469. Coimbra, Portugal: Springer Nature Switzerland, 2024, pp. 123–132

K. Dawoud, W. Samek, P. Eisert, S. Lapuschkin, and S. Bosse. ‘Human-Center Evaluation of XAI Methods’. In: Proceedings of the International Conference on Data Mining Workshops (ICDMW). IEEE. Shanghai, China: IEEE, Dec. 4, 2023, pp. 912–921

R. Achtibat, M. Dreyer, I. Eisenbraun, S. Bosse, T. Wiegand, W. Samek, and S. Lapuschkin. ‘From Attribution Maps to Human-Understandable Explanations through Concept Relevance Propagation’. In: Nature Machine Intelligence 5.9 (Sept. 20, 2023), pp. 1006–1019