Evangelos Chatzipantazis

Hello visitor! I am Evangelos. My name stems from the Greek Ev (< εύ) + angelos (< ἄγγελος) which translates to "the messenger of good news". So yes; if you are looking for good news you are in the right place.

I am currently a PhD Student in Computer Science at the legendary Grasp Lab at the University of Pennsylvania. Under the advising of Kostas Daniilidis, I work on Geometric Deep Learning and its applications to 3D Computer Vision and Robotics.

Moreover, I am currently doing an internship at Boston Dynamics AI Institute under Prof. Robert Platt and Prof. Robin Walters where I work on policy learning for robot manipulation.

I hold a Master of Science in Robotics from UPenn and a Master of Science in Statistics and Data Science from Wharton. Prior to that, I completed my undergraduate studies in the field of Electrical Engineering and Computer Science at the National Technical University of Athens, under the supervision of Prof. Petros Maragos, where I conducted research on spectral methods for image segmentation.

Research

My current research focus lies on Equivariant Deep Learning for 3D Computer Vision and Robotics. More broadly, I am interested in problems that fuse geometry, statistics and physics especially in the form of inductive biases on deep neural networks. I am also very interested in the use of Artificial Intelligence for Science.

During my PhD I have worked on 3D perception tasks, mainly on 3d reconstruction and point cloud registration.
On the theory side, I have designed frameworks to learn symmetries from data as well as optimization frameworks for equivariant deep networks.

I have collaborated on a project on system identification that received the Best Student Paper Award in ACC 2024 and on a project on graph neural networks for active information acquisition that received the Best Paper Award in ICRA 2023 in the category of multi-robot systems sponsored by Amazon Robotics.

Neural networks are seeing rapid adoption in purely inertial odometry, where accelerometer and gyroscope measurements from commodity inertial measurement units (IMU) are used to regress displacements and associated uncertainties. They can learn informative displacement priors, which can be directly fused with the raw data with off-the-shelf non-linear filters. Nevertheless, these networks do not consider the physical roto-reflective symmetries inherent in IMU data, leading to the need to memorize the same priors for every possible motion direction, which hinders generalization. In this work, we characterize these symmetries and show that the IMU data and the resulting displacement and covariance transform equivariantly, when rotated around the gravity vector and reflected with respect to arbitrary planes parallel to gravity. We design a neural network that respects these symmetries by design through equivariant processing in three steps: First, it estimates an equivariant gravity-aligned frame from equivariant vectors and invariant scalars derived from IMU data, leveraging expressive linear and non-linear layers tailored to commute with the underlying symmetry transformation. We then map the IMU data into this frame, thereby achieving an invariant canonicalization that can be directly used with off-the-shelf inertial odometry networks. Finally, we map these network outputs back into the original frame, thereby obtaining equivariant covariances and displacements. We demonstrate the generality of our framework by applying it to the filter-based approach based on TLIO, and the end-to-end RONIN architecture, and show better performance on the TLIO, Aria, RIDI and OxIOD datasets than existing methods.

Recent advances in learning or identification of nonlinear dynamics focus on learning a suitable model within a pre-specified model class. However, a key difficulty that remains is the choice of the model class from which the dynamics will be learned. The fundamental challenge is trading the richness of the model class with the learnability within the model class. Toward addressing the so-called model selection problem, we introduce a novel notion of Structural Risk Minimization (SRM) for learning nonlinear dynamics. Inspired by classical SRM for classification, we minimize a bound on the true prediction error over hierarchies of model classes. The class selected by our SRM scheme is shown to achieve a nearly optimal learning guarantee among all model classes contained in the hierarchy. Employing the proposed scheme along with computable model class complexity bounds, we derive explicit SRM schemes for learning nonlinear dynamics under hierarchies of: i) norm-constrained Reproducing Kernel Hilbert Spaces, and ii) norm-constrained Neural Network classes. We empirically show that even though too loose to be used as absolute estimates, our SRM bounds on the true prediction error are able to track its relative behavior across different model classes of the hierarchy.

Local shape modeling and SE(3)-equivariance are strong inductive biases to reconstruct scenes of arbitrarily many objects appearing in random poses even when a network is trained on single objects in canonical pose.

We propose a new Transformed Risk Minimization (TRM) framework as an extension of classical risk minimization. Our TRM method (1) jointly learns transformations and models in a single training loop, (2) works with any training algorithm applicable to standard risk minimization, and (3) handles any transforms, such as discrete and continuous classes of augmentations. To avoid overfitting when implementing empirical transformed risk minimization, we propose a novel regularizer based on PAC-Bayes theory. We propose a new parametrization of the space of augmentations via a stochastic composition of blocks of geometric transforms. The performance compares favorably to prior methods on CIFAR10/100. Additionally, we show empirically that we can correctly learn certain symmetries in the data distribution (recovering rotations on rotated MNIST) and can also improve calibration of the learned model.

We propose the Information-aware Graph Block Network (I-GBNet), an Active Information Acquisition adaptation of Graph Neural Networks, that aggregates information over the graph representation and provides sequential-decision making in a distributed manner. Numerical simulations on significantly larger graphs and dimensionality of the hidden state and more complex environments than those seen in training validate the properties of the proposed architecture and its efficacy in the application of localization and tracking of dynamic targets.

Talks

Tutorial: ”How to get started with equivariant deep learning”

• Video
• Slides

Organizer

Recording

Teaching

I am very passionate about teaching. Both from the mentoring perspective and as a means to convey knowledge in a clear, concise manner. I am a big fan of Richard Feynman's teaching techniques.

Class Notes (Co-authored with Prof.Pratik Chaudhari)

Website (under Prof. Jianbo Shi)

Credits for the template Jon Barron