Mechanical and Civil Engineering Seminar
Unsupervised identification and segmentation of surgical robotic operative data with movement primitives and Gibbs sampling
PhD Thesis Defense
This work addresses unsupervised machine learning as motivated by an interest towards increased autonomy in surgical robotics, and particularly towards increasingly autonomous learning that could be potentially deployed at scale to leverage a collective surgical consciousness across a global network of surgical experts and devices.
This work addresses low-level movement learning of bio-inspired and human-demonstrated motions through modeling of both continuous movements and discrete task phases in a hybrid dynamical system framework. Learning is addressed through a latent variable representation of the discrete modes and then the simultaneous continuous identification and discrete segmentation problem is approached probabilistically via Gibbs sampling. Dynamic movement primitives are leveraged for the continuous modeling and their multi-shot and one-shot learning is demonstrated from laboratory kinematic data on a benchtop pick-and-place task.
Finally, conditionally specified probability distributions are investigated fundamentally and the reliance on sampling based methods is re-examined. The conditional specification problem is shown to be often solvable and it is proven to be a special case of non-negative matrix factorization. Accordingly, an optimization approach is advocated for as a potential alternative to sampling based methods, enabling direct and deterministic solving of conditionally specified distributions. A proof-of-concept is shown wherein a classic Gibbs sampling task is solved numerically deterministically. Lastly, the theory of conditional compatibility is applied to identifying stationary distributions in nonlinear probabilistic dynamical systems in discrete time.
Please virtually attend this thesis defense:
Zoom meeting ID: 943 672 5035