YAYehudit Aperstein

Yehudit Aperstein

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Yehudit Aperstein has a total of 47 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Point Cloud Outlier Removal with Corruption-Stable Tokenization

    by Maoz Koren, Dovi Yellin, Yehudit Aperstein

    Real-world point clouds often contain spurious observations that obscure the underlying geometry. Modern point-cloud models struggle to distinguish such outliers from sparse but valid local geometry. We propose a two-stage transformer framework for additive outlier removal that combines self-supervised masked reconstruction with supervised point-wise classification. To reduce pre-encoding corruption, the framework replaces data-dependent token centers with fixed spatial centers, drawing on Vision Transformers. This establishes a stable token layout while still allowing token contents to vary. The model retains points classified as structural without changing their coordinates. Experiments on normalized object-level point clouds with synthetic additive corruption show that the method removes a substantial proportion of outliers while preserving most structural points. Self-supervised pretraining is critical to this balance, while alternative tokenization changes the removal-preservation trade-off. Performance remains stable across tested contamination levels, although near-surface outliers and thin or sparsely sampled structures remain challenging.

  • Open Access

    Modeling Normal Is All You Need: Joint Latent Clustering for Anomaly Detection in Multimodal Industrial-IoT Cyber-Physical Systems

    by Alexander Apartsin, Yehudit Aperstein

    The Industrial Internet of Things (IIoT) instruments physical processes with networked sensors and controllers that stream multivariate telemetry, and monitoring this telemetry for faults and cyberattacks is essential to reliability and safety. Machine learning is the natural tool, but failures are rare and seldom labeled, so supervised detection is impractical: a detector must learn normal behavior from unlabeled telemetry alone. In cyber-physical systems, this normal law is structured: a union of many imbalanced operating regimes shaped by physics and control, formalized as ten Massive, Implicit, Imbalanced Multimodality (MIIM) assumptions. That structure exposes a blind spot of the dominant approach: a fault can be individually normal on every sensor and reconstruct accurately, yet improbable under normal joint operation, so reconstruction-based detectors measure reachability rather than probability and miss it. Our detector, LatAD (Latent-density Anomaly Detector), learns a latent Gaussian-mixture clustering of regimes (VaDE, Variational Deep Embedding), scores anomalies by density, and factorizes that density over correlation-community subsystems combined by a cohesion-weighted Higher Criticism statistic. On three real IIoT testbeds, LatAD attains the best All-subset AUROC of any learned detector on HAI (0.948) and leads the difficult subset on HAI (0.845; a significant +0.088, 95% CI [0.042, 0.157]); USAD and TranAD fall to 0.44–0.50 on these reconstructable-but-improbable faults.

  • Open Access

    Toward Controllable Learner Simulation for Teacher Education: Benchmarking Skill-Profile Control in LLM Answer Behavior

    by Omri Sasson, Alexander Apartsin, Yehudit Aperstein

    Teacher preparation depends on deliberate practice with learners who have identifiable strengths and weaknesses, yet classroom experience accumulates slowly and is hard to standardize. Language-model-simulated students could offer scalable, repeatable practice, but only if a simulated learner can exhibit a specified answer-performance profile rather than behaving as a uniformly capable assistant. We introduce a benchmark for controllable learner simulation in school mathematics: the target condition is a binary vector of retained and suppressed curricular skill families, and we measure how faithfully a language model’s answer behavior realizes that profile via profile-alignment, retained-versus-forgotten, and cross-skill analyses. Across three models and two grades, selective suppression can be induced and measured, yet higher mathematical capability did not imply stronger controllability among the evaluated models: a model can solve the mathematics well yet be the weakest at selective suppression (bootstrap selectivity index 0.35 to 0.70). Induced errors are systematic: independent models converge on the same wrong answers, and an exploratory single-coder analysis found many compaible with documented error categories. Evidence is answer-level (single turn, multiple choice) at Grades 4 and 5. These results show that answer-level skill-profile controllability can be measured as one prerequisite for studying LLM-based learner simulation in teacher-preparation and assessment research.

  • Open Access

    LLM-guided headline rewriting for clickability enhancement without clickbait

    by Yehudit Aperstein, Linoy Halifa, Sagiv Bar, Alexander Apartsin

  • Open Access

    Selective multipathology chest X-ray classification via rejection mechanisms

    by Yehudit Aperstein, Amit Tzahar, Alon Gottlib, Tal Verber, Ravit Shagan Damti, Alexander Apartsin

    Overconfidence in deep learning models poses a significant risk in high-stakes medical imaging tasks, particularly in multi-label classification of chest X-rays, where multiple co-occurring pathologies must be detected simultaneously. This study introduces an uncertainty-aware framework for chest X-ray diagnosis based on a DenseNet-121 backbone, enhanced with an entropy-based selective prediction mechanism that abstains, on a per-pathology basis, from predictions whose binary entropy exceeds a calibrated threshold, deferring those findings to clinical experts. A quantile-based calibration procedure tunes rejection thresholds using either global or class-specific strategies. On NIH ChestX-ray14, MIMIC-CXR, and PadChest, selective prediction is evaluated with risk-coverage analysis and balanced accuracy under non-parametric bootstrap confidence intervals. At a calibrated operating point, abstention improves balanced accuracy for all four target pathologies on NIH ChestX-ray14, with gains of 0.07–0.09 in balanced accuracy at a coverage of 0.70, a relative reduction in balanced error of 25–36% (95% confidence intervals excluding zero). Every target pathology is also supported on MIMIC-CXR and PadChest, on an independently trained model evaluated cross-dataset on a source excluded from its training, and under leave-one-dataset-out domain shift in both transfer directions. These results support the integration of selective prediction into AI-assisted diagnostic workflows, providing a practical step toward safer, uncertainty-aware deployment of deep learning in clinical settings.

  • Open Access

    IRC-Bench: Recognizing Entities from Contextual Cues in First-Person Reminiscences

    by Yehudit Aperstein, Eden Moran, Alexander Apartsin

    When people recount personal memories, they often refer to people, places, and events indirectly, relying on contextual cues rather than explicit names. Such implicit references are central to reminiscence narratives: first-person accounts of lived experience used in therapeutic, archival, and social settings. They pose a difficult computational problem because the intended entity must be inferred from dispersed narrative evidence rather than from a local mention. We introduce IRC-Bench, the Implicit Reminiscence Context Benchmark, for evaluating implicit entity recognition in reminiscence transcripts. The benchmark targets non-locality: entity-identifying cues are distributed across multiple, non-contiguous clauses, unlike named entity recognition, entity linking, or coreference resolution. IRC-Bench comprises 25,136 samples constructed from 12,337 Wikidata-linked entities across 1994 transcripts spanning 11 thematic domains. Each sample pairs an Entity-Grounded Narrative, in which the target entity is explicitly mentioned, with an Entity-Elided Narrative, in which direct mentions are removed. We evaluate 19 configurations across LLM generation, dense retrieval, RAG, and fine-tuning. QLoRA-adapted Llama 3.1 8B performs best in the open-world setting (38.94% exact match; 51.59% Jaccard), while fine-tuned DPR leads in closed-world retrieval (35.38% Hit@1; 71.49% Hit@10). We release IRC-Bench with data, code, and evaluation tools.

  • Open Access

    Mapping License Plate Recoverability Under Extreme Viewing Angles for Opportunistic Urban Sensing

    by Igor Adamenko, Orpaz Ben Aharon, Yehudit Aperstein, Alexander Apartsin

    Urban environments are saturated with imaging sensors deployed for purposes unrelated to vehicle identification, from ATM and dashboard cameras to pole-mounted CCTV and smartphones. We term the use of such non-purpose-built sensors for secondary inference “opportunistic sensing”; its central question is where, under uncontrolled capture conditions, AI-enabled restoration remains reliable. This paper introduces recoverability maps, a task-agnostic methodology for quantifying that boundary, and applies it to oblique-view license plate recognition (LPR). It pairs a full-grid synthetic sweep of the degradation space with two summary measures: a boundary area-under-curve for coverage and a reliability score F for the frequency and depth of interior unrecovered pockets. For LPR, the space is the oblique-angle grid [0°,89°]2 sampled by Scrambled Sobol sequences, and the utility is plate-level optical character recognition (OCR) accuracy. Within this synthetic benchmark, approximately 90–92% of the angle grid is recoverable (best single model to union of restoration arms), recovery degrades sharply beyond roughly 80° in both axes, and lateral rotations are harder to reconstruct than elevational ones. Five restoration architectures cluster within a narrow AUC band of 0.89–0.93, and share the same α/β asymmetry, so the recoverable region is set primarily by sensing geometry, with architecture affecting efficiency and interior consistency; discriminative architectures outperform generative models. The methodology is validated on real plates: on CCPD and the Brazilian legacy and Mercosur layouts of RodoSol-ALPR, restoration raises held-out extreme-angle recognition by +15 to +38 exact-match points under plate-specialized recognizers, and the discriminative-over-generative ordering reproduces on real data.

  • Open Access

    CalexNet: Soft Cascade-Aligned Training and Calibration for Lightweight Early-Exit Branches

    by Yehudit Aperstein, Alexander Apartsin

    Early-exit cascades over a frozen convolutional backbone enable adaptive inference but suffer from three sources of train–inference mismatch: branches train on samples they will never see at inference; their per-class precision thresholds are calibrated on the wrong distribution; the standard cross-entropy target on backbone argmax labels discards the backbone’s uncertainty signal. We close all three gaps with CalexNet (cascade-aligned early exits), a training-recipe-only modification: branches train under continuously weighted importance sampling that matches the cascade-survivor distribution; per-class precision thresholds are calibrated on the actual cascade-survivor subset of the validation set; the classification head is trained against the backbone’s full softmax via a temperature-scaled KL objective. Combined with an augmented prototype-pooling branch head, CalexNet is evaluated on ResNet18 and ResNet50 backbones across CIFAR-100 (20-supe-class coarse, the harder primary setting) and CINIC-10 (10-class, the easier cross-validation counterpart). On the accuracy–FLOPs Pareto frontier, CalexNet matches or exceeds three published baselines (PTEEnet, ZTW, BoostNet) and a within-paper “no-alignment, no-KD” reference. The largest gains appear in the practically relevant 30–70% FLOPs-reduction regime and show consistent trends across n=3 training seeds. CalexNet requires no inference-time architectural change and is a drop-in for any frozen-backbone early-exit cascade.


  • Open Access

    A benchmark for evaluating diagnostic questioning efficiency of LLMs in patient conversations

    by Mai Werthaim, Maya Kimhi, Alexander Apartsin, Yehudit Aperstein

    Physician–patient diagnostic interactions often rely on incomplete, uncertain, and noisy symptom descriptions expressed in non-specialist language. To reach an accurate diagnosis efficiently, physicians employ adaptive sequences of focused questions in which each inquiry depends on prior patient responses. Similar requirements apply to AI assistants engaged in clinical dialogue, which must strategically select questions to elicit diagnostic information efficiently. While several datasets and benchmarks address clinical reasoning, few focus on evaluating the strategic inquiry process itself. To address this gap, we introduce Q4Dx (Question-Driven Diagnosis), a benchmark for assessing goal-directed diagnostic questioning. Q4Dx consists of synthetically generated patient cases derived from curated symptom–disease knowledge. Each case is instantiated at multiple information levels (100%, 80%, and 50% symptom exposure) to evaluate performance under varying informational constraints. We simulate patient–physician interactions using GPT-4.1 and GPT-4o-mini agents to generate clinician questions, patient responses, and intermediate diagnostic hypotheses. Performance is evaluated using Zero-shot Diagnostic Accuracy (ZDA), Mean Questions to Correct Diagnosis (MQD), and Inquire Sequence Efficiency (ISE), which measures convergence toward the correct diagnosis. Q4Dx provides a reusable framework for benchmarking large language models in strategic clinical inquiry and supports future research on AI-assisted diagnostic training. The dataset and benchmark are publicly available at: https://github.com/MaiWert/MedQDx.