News & EventsDepartment Events & Announcements
Events
-
Jul24
EVENT DETAILS
lessCameras often deviate from ideal image formation because of optical, sensor, and scene-dependent physics. These deviations are usually treated as errors to suppress, but they can also provide useful structure when modeled during reconstruction. This dissertation develops a unified measurement-model perspective for computational imaging: a physical effect can be treated either as an encoding to invert or as a nuisance to suppress, depending on the scene properties the system aims to recover. Across spectral and thermal imaging, this perspective enables practical cameras to recover richer scene information from measurements that would otherwise appear noisy, ambiguous, or inconsistent.
First, we study hyperspectral imaging, where light must be distributed across many wavelength channels. This creates a photon-efficiency trade-off among spectral resolution, spatial resolution, acquisition time, and noise. Two systems address this trade-off in complementary ways. HyperColorization uses a dense grayscale guide image to propagate sparse spectral measurements, exploiting the high photon efficiency of broadband grayscale imaging while recovering dense spectra. Spectrum from Defocus instead uses longitudinal chromatic aberration as an optical encoding: a focal stack captured with simple refractive optics and a grayscale sensor provides wavelength-dependent blur that can be calibrated and inverted. Together, these systems show how scene structure and optical structure can reduce the cost of spectral imaging without relying on complex spectral hardware.
Later we study thermal imaging for novel view synthesis. Thermal cameras operate in darkness and adverse visibility, but their measurements are not stable temperature images: they contain drift, fixed-pattern noise, weak texture, limited dynamic range, and view-dependent reflection. The first thermal system addresses the sensor-driven part of this problem by characterizing these degradations across public multiview datasets, stabilizing thermal video before reconstruction, and absorbing residual radiometric variation during Gaussian splatting. The second thermal system treats view-dependent thermal appearance not only as an error source but also as a cue. By exploiting the fact that emitted radiance is approximately view-stable while reflected radiance changes with viewpoint, it decomposes reconstructed thermal scenes into temperature-correlated emission and material-dependent reflection.
Across these systems, the same principle recurs: the role of a physical effect is determined by the reconstruction target. Chromatic aberration can be suppressed in a conventional RGB camera or inverted as a spectral measurement; thermal reflection can be absorbed as nuisance variation or estimated as scene content. By designing sensing models and reconstruction algorithms together, this dissertation shows that practical cameras can turn structured deviations from ideal imaging into physically meaningful measurements.TIME Friday, July 24, 2026 at 9:30 AM - 11:30 AM
LOCATION Mudd 3514, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Jensen Smith jensen.smith@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
Jul29
EVENT DETAILS
lessGraphs are mathematical objects that model many complex systems and have become a central element of algorithm design. However, traditional worst-case analysis often yields pessimistic complexity bounds that ignore the typical structure of real-world instances. This thesis studies average-case analysis for various algorithms solving combinatorial optimization problems on graphs. It also investigates the combinatorial structures underlying these problems. These insights lead to a better understanding of the limitations of existing algorithms and provide guidance for developing new algorithmic techniques.
Specifically, my research includes the following problems:
Community Detection on Stochastic Block Models (SBM). Consider a graph in which each vertex belongs to a hidden community, and the probability of an edge between any pair of vertices depends solely on the community assignments of the two vertices. The goal is to recover the community partition from the observed graph. We show that a semidefinite programming (SDP) relaxation designed for exact recovery in symmetric SBMs cannot be directly extended to asymmetric SBMs, and we provide geometric intuition explaining this limitation.
Sum of Leaf Weights in the Minimum Spanning Tree. Let G be a complete graph in which each edge weight is sampled independently from the Uniform(0,1) distribution, and let T be the minimum spanning tree of G. Define a leaf edge as an edge of T that is incident to a leaf of T. We establish tight bounds on the expected sum of the weights of all leaf edges, together with the concentration around it. These results substantially improve the state-of-the-art bounds for several variants of the minimum spanning tree problem, including the probabilistic minimum spanning tree (PMST).
Probabilistic Minimum Spanning Tree (PMST). Given a graph in which each vertex is present independently with probability p, the goal is to find an a priori spanning tree T that minimizes the expected total length after deleting edges while preserving connectivity among the vertices that remain present. We establish new upper and lower bounds on the expected length of the optimal a priori spanning tree.TIME Wednesday, July 29, 2026 at 10:00 AM - 12:00 PM
LOCATION 3501, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Jensen Smith jensen.smith@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
Jul29
EVENT DETAILS
lessData collections—such as arrays, lists, sets, and maps—are among the most commonly used abstractions in modern programs, yet compilers discard their rich semantic structure before performing analysis and optimization.
Once lowered to raw memory, the logical organization of a collection is lost, leaving the compiler to reason about undecidable problems, such as pointer aliasing.I present MEMOIR: a general-purpose compiler intermediate representation (IR) that preserves data collections in static single assignment (SSA) form, making their logical structure explicit for analysis and optimization.
I demonstrate the benefits of MEMOIR with novel layout-agnostic optimizations that operate directly on logical collections.
Furthermore, I show how the compiler can manufacture data properties to enable specialized implementations, such as bitsets, achieving significant performance and memory savings compared to industry-standard implementations.
Representing collections in SSA form makes these analyses efficient through sparse dataflow, but introduces two challenges.First, SSA represents updates functionally.
Lowering them back to imperative code risks introducing copies that turn an O(1) update into an O(n) operation.
I formally prove that MEMOIR programs can always be lowered without such copies, and show that standard compiler transformations can be made to maintain this guarantee with 100% effectiveness and negligible overhead.
Second, references and iterators, which are ubiquitous in languages like C++, identify elements independently of the collection that owns them.
However, SSA value semantics have no way to express this.
I introduce bookmark semantics to resolve this tension.TIME Wednesday, July 29, 2026 at 2:00 PM - 4:00 PM
LOCATION Mudd 3514, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Jensen Smith jensen.smith@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
Jul30
EVENT DETAILS
lessModern inference problems often involve randomly generated data with structure that is not directly observed. In network data, the structure may take the form of hidden communities or unknown correspondences between vertices across correlated observations of the same underlying network. In statistical models, a low-dimensional parameter of interest may be obscured by an unknown and potentially high-dimensional component of the data distribution. Ignoring this structure can lead to overly pessimistic conclusions, while modeling every unknown component explicitly may be infeasible. A central challenge is therefore to identify structural properties that enable inference without requiring complete knowledge of the data-generating process.
This thesis develops algorithmic and statistical methods for inference in various structured probabilistic models. One direction concerns the recovery of latent combinatorial structure in random graphs, including graph matching and community detection in correlated and growing network models such as correlated stochastic block models and the preferential attachment block model. We establish recovery guarantees, including guarantees without computational constraints, and develop efficient algorithms. A second direction concerns inference under structured distributional uncertainty, including the construction of optimal adaptive confidence intervals in the Gaussian mean-shift contamination model and the estimation of watermark proportions in language-model-generated text. Together, these results show how structures under random observations help design provable algorithms for recovering hidden objects and estimating parameters.TIME Thursday, July 30, 2026 at 1:00 PM - 3:00 PM
LOCATION 3501, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Jensen Smith jensen.smith@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
Aug20
EVENT DETAILS
lessThe component-by-component migration of a program from untyped to typed can trigger unintended performance degradations. When such a degradation occurs, typing well-chosen components can lessen the cost of type enforcement, while typing poorly chosen components can exacerbate it. In this talk, I examine whether off-the-shelf profiling tools deliver information that helps programmers navigate these migration choices effectively in Typed Racket.
TIME Thursday, August 20, 2026 at 1:00 PM - 4:00 PM
LOCATION Mudd 3514, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Jensen Smith jensen.smith@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
Sep24
EVENT DETAILS
lesstba
TIME Thursday, September 24, 2026 at 9:00 AM - 11:00 AM
LOCATION 3514, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Wynante R Charles wynante.charles@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
Sep28
EVENT DETAILS
lessMonday / CS Seminar
September 28 / 12:00 PM
Hybrid / Mudd 3514Speaker
TBATalk Title
TBAAbstract
TBABiography
TBA---
Zoom Link
Panopto LinkTIME Monday, September 28, 2026 at 12:00 PM - 1:00 PM
LOCATION 3514, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Wynante R Charles wynante.charles@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
Oct5
EVENT DETAILS
lessMonday / CS Seminar
October 5 / 12:00 PM
Hybrid / Mudd 3514Speaker
Brian Suchy, Software Engineer Google DeepMindTalk Title
Formal Relational Equivalence for SQL, GenAI, and BeyondAbstract
Verifying that complex query rewrites from database optimizers or AI-driven generators preserve exact bag semantics under three-valued logic is an NP-hard challenge. To address this, we present an MLIR-native compiler framework that formally reasons about relational algebra. By decoupling query semantics from specific execution engines and lowering queries into a unified Relational Algebra Intermediate Representation, our language-agnostic methodology definitively proves semantic equivalence across all possible database states.The core of the presentation will deep-dive into our multi-tiered proving architecture, which synthesizes several advanced academic methodologies. First, we utilize E-Graphs and Equality Saturation to rapidly explore the equivalence space and detect structural congruence between query abstract syntax trees using fast, algebraic rewrite rules. Second, we employ Semiring Arithmetic, treating relational algebra as expressions over K-relations to leverage algebraic simplification and canonical forms under semiring laws. Finally, we implement a First-Order Logic and SMT translation path, lowering Relational Algebra into Relational Calculus and then into First-Order Logic to evaluate constraints and domain-specific axioms using parallel solvers like Z3 and CVC5, which either formally proves equivalence or synthesizes concrete counter-examples.
Finally, we will discuss the practical implications of combining these formal mathematical methods with modern compiler design. Attendees will leave with a comprehensive understanding of how bridging database theory, equality saturation, and SMT solving can create robust solutions for verifying query optimizers, enforcing semantic correctness, and validating automated SQL generation at scale.
Biography
Brian Suchy is a Software Engineer within Google DeepMind.
In his time at Google he has worked on F1 Query (Google's internal SQL query engine), hardware development, and (of course) AI.
Prior to joining Google, Brian received his PhD student at Northwestern University, advised by Peter Dinda, with a focus on hardware/software codesign and memory management.Research Interests: Artificial Intelligence, Query Processing and Formal Logic
---
Zoom Link
Panopto LinkTIME Monday, October 5, 2026 at 12:00 PM - 1:00 PM
LOCATION 3514, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Wynante R Charles wynante.charles@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
Oct29
EVENT DETAILS
lesstba
TIME Thursday, October 29, 2026 at 9:00 AM - 11:00 AM
LOCATION 3514, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Wynante R Charles wynante.charles@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)