News & EventsDepartment Events & Announcements
Events
-
May26
EVENT DETAILS
lessTraditional database management systems (DBMSs) rely on data-dependent observability, where the optimizer utilizes selectivities and intermediate cardinalities to select efficient execution plans. As privacy concerns increase and regulatory requirements are enforced, privacy-preserving DBMSs lose access to this information. While secure query execution becomes feasible, it often incurs high computational costs. An additional challenge arises in production environments. In multi-tenant cloud platforms, when a tenant experiences a slow query, developers would typically re-execute the query on the original data to diagnose performance regressions. However, confidentiality requirements prevent this approach, leading to genuine but unreproducible performance regressions. This thesis proposes that database metadata can serve as a substrate to address both challenges. From the perspective of DBMS builders, public schema constraints and protocol-visible information can substitute for the private statistics used by conventional optimizers. From the perspective of DBMS operators, differentially private releases of physical metadata can reproduce execution behavior on substitute datasets.
My prior work develops three systems based on this principle. Alchemy derives a circuit-aware cost optimizer for oblivious SQL using public schema constraints. HAMMER extends the principle beyond a single privacy primitive, routing public operators to plaintext, slot-wise arithmetic to fully homomorphic encryption (FHE) on GPUs, and control-flow-heavy operators to secure multi-party computation (MPC). ScanTwin generates differentially private sketches of Parquet footers, allowing scan-level performance regressions to be reproduced on synthetic data without accessing tenant records. Building on these results, the proposed thesis extends ScanTwin to PerfTwin, releasing operator-level differentially private sketches for additional operators to enable reproducing full-pipeline performance regressions. Overall, the goal is to demonstrate that, despite privacy constraints, the database stack can be efficiently built and reliably operated from safe metadata alone.
TIME Tuesday, May 26, 2026 at 1:00 PM - 3:00 PM
LOCATION Mudd 3001, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Jensen Smith jensen.smith@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
May26
EVENT DETAILS
lessTuesday / CS Seminar
May 26 / 1:00 PM
Hybrid / Mudd 3514Speaker
Amil Dravid, UC BerkeleyTalk Title
Deep Learning as a Natural Science: From Neurons to Scaling LawsAbstract
"As neural networks grow larger, they increasingly resemble complex systems with systematic behavior rather than isolated engineered artifacts. This is reflected in neural scaling laws, which show that performance follows a precise power law as a function of model size, data, and compute. Yet, we still lack a principled scientific understanding of how these models organize information internally.
In this talk, I will describe an approach to studying neural networks as scientific objects, drawing on ideas from neuroscience and statistical physics. I will focus on neurons that respond to similar inputs across independently trained models of the same modality — for example, neurons that fire for similar words across language models or similar objects and textures across vision models. We find that the population of shared neurons grows as a power law with model size, but represents a shrinking fraction of the overall population. We account for this trend with a simple theoretical model, which also predicts a polarization effect: as models grow, shared neurons become increasingly selective for specific concepts, such as code or visual textures, while the remaining neurons become less selective and encode multiple features. We then test this prediction in large-scale language and vision models. I will conclude with a case study showing how neuron selectivity can guide targeted data selection to improve training. Together, these results suggest that scaling laws are not restricted to external model behavior; they also describe regularities in how neural networks are organized internally."
Biography
Amil Dravid is a third-year PhD student in computer science at UC Berkeley, advised by Alyosha Efros. His research develops tools for understanding the internal structure of deep learning systems, with the goal of informing the design of more effective algorithms and models. He previously earned his BS in Computer Science from Northwestern University. His research is supported by the U.S. Department of Energy Computational Science Graduate Fellowship.Research Areas: Deep learning, computer vision, natural language processing
---
Zoom Link
Panopto LinkTIME Tuesday, May 26, 2026 at 1:00 PM - 2:00 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)
-
May26
EVENT DETAILS
lessCSPAC and CSSI are excited to announce our upcoming series of events on public speaking! All are welcome and encouraged to participate. See below for details:
Lightning talk competition @ May 26, 3:30-5 PM [sign up to present]
Present a 3-minute lightning talk about your research to a panel of judges. Winners will be awarded at the department awards ceremony on May 27. Snacks and refreshments will be provided. While we encourage all students to present, you are also welcome to join to watch and support your peers!
Students at any stage in their program are highly encouraged to participate! This is a great opportunity to practice an essential research skill, share your research with others, and receive constructive feedback from supportive faculty!
Questions? Email melissac@u.northwestern.edu.
--
What is CSPAC?We are the CS PhD Advisory Council. We are a PhD student-led organization, and our mandate is to interface between PhD students and faculty on academic issues. We want to advocate for PhD students in the department, so if there is some way we can support you, please come talk to us. We welcome PhD students to our weekly meetings on Wednesday 9:30am-10:30am in Mudd 3501 and on zoom. We also welcome anonymous concerns/feedback at any time via this form. Anyone in the community can reach us at cspac@u.northwestern.edu.
TIME Tuesday, May 26, 2026 at 3:30 PM - 5:00 PM
LOCATION 3514, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Melissa Chen melissac@u.northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
May27
EVENT DETAILS
lessDear CS Community:
We are excited to invite you to our End of Year Awards Celebration! Join us on Wednesday, May 27th in TGS Commons at 3pm to help recognize the amazing dedication and hard work our students, faculty, and staff have done this year. Light refreshments and snacks will be served.
Please RSVP no later than May 25th by following this link: Computer Science End of Year Awards Celebration
TIME Wednesday, May 27, 2026 at 3:00 PM - 5:00 PM
LOCATION TGS Commons, 2122 Sheridan Road map it
CONTACT Wynante Charles wynante.charles@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
May27
EVENT DETAILS
lessWhen viewers read a data visualization, they are translating visual marks into quantitative judgments that are systematically imperfect. Decades of graphical perception research have documented these errors, but most findings take the form of ordinal rankings or categorical taxonomies. While helpful, these descriptions do not make quantitative predictions: they cannot predict how large a viewer's error will be for a specific chart design, dataset, and task, nor can they generalize to combinations not yet tested experimentally.
This dissertation develops computational models that make quantitative predictions about how viewers interpret data visualizations. Four studies model different aspects of visualization perception with increasing generality. First, I develop a formal model of y-axis truncation in bar charts that defines task-dependent conditions under which truncation preserves or distorts data structure, replacing heuristics with formally grounded, computable design guidance. Second, I test whether deep neural network features trained on natural images can serve as computational proxies for human similarity judgments of visualizations. Third, I apply signal detection theory (SDT) to visual lineup analysis, demonstrating that SDT provides a richer computational model of lineup perception than accuracy-based approaches by separating viewer sensitivity from decision criterion. Finally, I propose visual decoding operators --- composable perceptual primitives, each with estimable bias and variance --- and provide an existence proof that operators characterized on PDF and CDF charts compose to predict scatterplot mean-estimation performance with no parameters fit to the target data.
Together, these studies demonstrate that computational models of visualization perception are both feasible and productive: they predict quantities that ordinal rankings cannot, expose mechanisms that holistic accuracy measures obscure, and generalize across chart types and tasks.TIME Wednesday, May 27, 2026 at 3:00 PM - 5:00 PM
LOCATION Mudd 3501, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Jensen Smith jensen.smith@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
May28
EVENT DETAILS
lessJoin us for free bagels and coffee followed by an informal discussion hosted by CSPAC and CSSI.
TIME Thursday, May 28, 2026 at 9:00 AM - 11:00 PM
LOCATION 3514, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Wynante Charles wynante.charles@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
May29
EVENT DETAILS
lessPrivate set operators are functions over private inputs that multiple parties can jointly evaluate while revealing only the prescribed output. Two such operators are intersection and join-and-compute, realized respectively by private set intersection (PSI), outputting the intersection of two sets, and private join and compute (PJC), outputting aggregates such as cardinality, sum, and inner product over matching records. Classical PSI and PJC target one-shot two-party settings where each party holds its full input. Real deployments rarely fit this model: servers maintain persistent datasets reused across many clients, and inputs are often split across multiple data owners. Existing protocols fall short: they lack cross-execution consistency, require per-execution server reprocessing, or incur substantial overhead for distributed inputs.
This thesis develops efficient and provably secure protocols for private set operators in practical client-server settings, through three schemes together with new cryptographic primitives:
(1) Inspired by password-checkup applications, we study client-output PSI in which the server publishes a one-time, linear-size encoding of its set, after which each client executes PSI with the server at cost linear only in its own set, with simulation-based security against malicious adversaries. A key ingredient is an efficient oblivious verifiable unpredictable function (OVUF).
(2) We introduce committed vector oblivious linear evaluation (C-VOLE), which generates VOLE correlations on a pre-committed vector and serves as a unifying tool for zero-knowledge proofs of committed values and actively secure multi-party computation. Built on a tailored LPN-based commitment, our matching C-VOLE protocols exploit the commitment structure to minimize the cost of binding the committed vector to the VOLE correlation, and efficiently instantiate a maliciously secure server-output PSI protocol.
(3) Beyond intersection, we study computation over matching records from distributed datasets, motivated by applications such as privacy-preserving ad conversion measurement. We propose the first efficient approximate PJC protocol with communication sublinear in the input size. Its core is a new adaptation of the Alon-Matias-Szegedy (AMS) sketch, redesigned for efficient evaluation under fully homomorphic encryption via structured randomness.
TIME Friday, May 29, 2026 at 10:00 AM - 12:00 PM
LOCATION Mudd 3501, Mudd Hall ( formerly Seeley G. Mudd Library) map it
CONTACT Jensen Smith jensen.smith@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)
-
May29
EVENT DETAILS
lessWhile the ubiquitous adoption of algorithms in decision-making and content generation greatly improves societal efficiency, it has also raised various regulatory concerns, including antitrust, nondiscrimination, and intellectual property protection.
This dissertation investigates techniques from three different computer science research areas to support the regulation of algorithms: First, based on ideas from online learning theory, we propose a framework for the regulation of algorithmic collusion by auditing from data. Second, by adapting principles from information flow control with dynamic policies, we design a type system to reason about iteration (probabilistic) independence to support the regulation of the fairness of classification algorithms. Third, we study a proper data attribution notion informed by data privacy concepts for the regulation of credit attribution of generative models. These results demonstrate that the foundations of the regulation of algorithms can benefit from techniques from these distinct areas of computer science research and point to future research directions.
TIME Friday, May 29, 2026 at 1:00 PM - 3:00 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)
-
Jun5
EVENT DETAILS
lessThere is growing momentum—from industry, government, and academia—to use AI for automating cybersecurity tasks. Yet practitioners remain skeptical: while 87% of security leaders expect AI to enhance their roles, only 9% believe it will replace significant parts of them. This gap stems from two fundamental barriers: limited capability and lack of trust. In this talk, I present my research on addressing these barriers through explainable AI. I first introduce StateMask, a method that automatically identifies critical decision steps in AI agent trajectories, enabling security professionals to understand why an AI-generated patch succeeded or failed. A user study with 41 experienced developers shows that 89% find our explanations aligned with their reasoning. I then present GPO, which leverages these explanations to synthesize high-quality training data without expensive expert annotation, thereby improving model capability. GPO-trained open-source models achieve performance competitive with leading commercial models on vulnerability patching, and its extension, EntroPO, ranks 1st on SWE-Bench Lite among all open-weight models. I conclude by discussing future directions toward building AI systems that are robust to imperfect data, trusted by security professionals, and capable of tackling real-world cybersecurity challenges.
TIME Friday, June 5, 2026 at 9:00 AM - 11:00 AM
CONTACT Jensen Smith jensen.smith@northwestern.edu EMAIL
CALENDAR Department of Computer Science (CS)