Colloquia and Seminars

Accurate retrieval of cloud geometric and dynamical properties from satellite observations is essential for understanding atmospheric processes and improving weather and climate models. While existing stereoscopic techniques can estimate cloud-top height and largescale atmospheric motion, recovering dense, cloud-scale vertical velocity fields at tens-of meters spatial resolution from passive satellite imagery remains a challenging inverse problem because the required atmospheric dynamics are not directly observable. This thesis presents a deep learning framework for estimating physically meaningful atmospheric quantities directly from passive multi-view satellite observations. A comprehensive simulation pipeline was developed by combining large-eddy simulations of shallow cumulus clouds with physically based volumetric Monte Carlo rendering, atmospheric radiometric correction, sensor noise modeling, and image alignment to generate realistic synthetic satellite imagery together with corresponding ground-truth atmospheric fields. These data were used to train a fully convolutional encoder-decoder network operating on two temporally consecutive three-view observations.
The proposed framework was evaluated on three retrieval tasks: cloud-top vertical velocity, cloud-top height, and vertical velocity at multiple fixed altitude layers throughout the atmospheric column. Experimental results demonstrate accurate retrieval of cloud-top geometry together with successful estimation of cloud-top and volumetric atmospheric motion at the native 20~m spatial resolution of the simulations. The framework recovers physically meaningful atmospheric velocity fields from passive multi-view observations and successfully infers vertical motion throughout much of the cloud-containing atmosphere using a single network architecture and observational input. The presented results demonstrate that passive multi-view satellite imagery contains sufficient information to support dense retrieval of both geometric and dynamical cloud properties without explicit reconstruction of the three-dimensional cloud structure or intermediate physical modeling. These findings establish the feasibility of learning-based atmospheric retrieval from physically realistic simulations and provide a foundation for future development of operational data-driven cloud retrieval methods.

Udi Gal is an M.Sc. student under the supervision of Prof. Yoav Schechner.

Pipeline parallelism is a key technique for training large language models in distributed settings, but existing schedules face a fundamental trade off between hardware utilization and optimization consistency. Synchronous methods preserve correct optimization but suffer from pipeline bubbles, while asynchronous methods eliminate these bubbles at the cost of stale model weights and additional mechanisms such as weight stashing or prediction. In this work, we present PACI (Pipeline Asynchronous training with Controlled Inconsistency), an asynchronous pipeline training algorithm that achieves bubble free execution while maintaining bounded weight inconsistency without weight stashing, weight prediction, additional parameter copies, or global synchronization. Evaluated on GPT-2 pretraining, PACI matches the convergence, model quality, and memory footprint of a synchronous schedule while reducing training time-to-accuracy by up to 1.69x.

Many drugs produce clear clinical or cellular effects, but the molecular targets underlying these effects often remain unknown. Identifying these targets is essential for understanding mechanisms of action, anticipating side effects, and guiding drug optimization.

Direct biochemical validation methods can provide strong evidence for drug–target interactions, but they are often time-consuming, expensive, and difficult to apply at large scale. This creates a need for high-throughput approaches that can prioritize candidate target proteins for focused downstream validation. Limited Proteolysis Coupled to Mass Spectrometry, LiP-MS, is a proteome-wide approach for target identification that detects drug-induced changes in protein structure or accessibility through changes in protease-generated peptide patterns.

However, existing LiP-MS scoring strategies have mainly been tested on small sets of well-characterized, high-affinity ligands, leaving their performance on broader and more heterogeneous drug–target panels unclear. In this thesis, we evaluate LiP-MS-based target identification on a diverse panel of drug–target pairs and develop a computational framework for improving target prioritization.

The framework combines per-drug statistical evidence with cross-drug background patterns in order to recover known target proteins while reducing the impact of recurrent false-positive responders. We further incorporate biological knowledge about protein structure and function by testing whether high-scoring peptides fall within known active or binding domains. This additional layer helps distinguish biologically plausible candidates from proteins whose statistical signal is less likely to reflect a true drug–target interaction.

Together, this work provides a systematic analysis of factors affecting LiP-MS target identification and introduces a pipeline that integrates statistical scoring with protein domain information. By improving recovery of known targets and supporting more confident interpretation of ambiguous hits, the proposed framework can help guide future validation experiments toward the most relevant candidate target proteins.

Mechanistic interpretability (MI) has emerged as a powerful approach for analyzing and intervening in inference computations, with a growing number of applications such as jailbreak attempt detection, truthfulness evaluation, and hallucination detection. Unfortunately, MI deployment in production model-serving systems is currently not practical, as most existing MI frameworks introduce prohibitively high runtime overheads. The fundamental problem is that MI functions do not compose cleanly with served models: they fragment deployment, often force draining requests and rebuilding serving state, and conflict with critical performance optimizations such as continuous batching and CUDA-graph execution, essential for production deployments.

In this thesis, we present a comprehensive approach for runtime logic integration into LLM serving. Culminating in implementing xMIx, a serving-native framework for deploying MI applications in production inference serving environments. xMIx enables attaching MI functions to a predefined set of locations in the model runtime, interposing on activations within the layers and submodules. xMIx supports conditional invocation of MI functions depending on the outputs in preceding model layers. Multiple MI applications can be deployed in a single model instance. xMIx compiles them all into the serving path but activates them dynamically at runtime only when necessary, with negligible performance cost, and without requiring a separate model instance or alternative execution stack.

We integrate xMIx with the vLLM serving system and evaluate it across three major models and seven diverse MI applications. xMIx achieves performance comparable to native vLLM execution, incurring a slowdown of 1.3% mean inter-token latency (ITL), 1.2% for tail P99 ITL, 2.6% for mean time to first token (TTFT), and 1.6% for mean total token throughput (TTT).

DNA-based storage offers exceptional density and durability, but its scalability depends on efficiently synthesizing many strands in parallel under physical constraints. We study two complementary models capturing limitations arising from nucleotide availability and spatial interactions in large strand arrays.

First, we introduce complex synthesis sequences, a hybrid framework that bridges enzymatic synthesis, which permits strand-specific nucleotide additions, and photolithographic synthesis, which applies common additions across many strands. We characterize the achievable information rate through an analogue of the deletion ball, derive tight expressions for the maximal rate and its asymptotic behavior, and present a dynamic programming algorithm for computing an optimal synthesis sequence for known strands.

We then consider a two-dimensional array model motivated by optical and chemical coupling, in which strands follow a fixed global synthesis sequence and at most one strand per row may advance in each cycle. For two strands in a single row, we show that the laggard-first policy is asymptotically optimal among online policies without look-ahead, while one-symbol look-ahead yields a strict improvement in the binary case.

We further show that even globally optimal scheduling incurs an unavoidable expected overhead that grows linearly with strand length. These results are complemented by an offline dynamic programming algorithm and a constant-redundancy binary coding scheme with a deterministic worst-case synthesis-time guarantee. Together, the two works establish a unified framework for understanding the information-theoretic and scheduling limits of constrained DNA synthesis.

My research focuses on mechanistic interpretability-understanding and improving multi-modal models by analyzing their internal computations. I will present recent studies, including Diffusion Lens (ACL 2024), Padding Tone (NAACL 2025), and Follow the Flow (ACL 2026), and DeLeaker (ICL 2026). These works propose methods to reveal what text-to-image models learn and how computations are performed internally. These insights are then applied to solve concrete problems, such as semantic leakage, paving the way for more reliable and transparent generative AI.