2026 · paper
Adversarial Attacks in Weight-Space Classifiers
Examines adversarial attacks against classifiers that operate on neural-network weight representations.[73]
Machine learning, generative models, and AI safety
Ethan Fetaya
איתן פתיה
Identity: verifiedThe roster maps to BIU’s source-spelled Ethan Fetaya; the laboratory page supplies the matching identity and research program. [73]
Documented foundation
Research & experience
Fetaya’s lab spans machine learning and vision, including generative models, AI safety, geometric and Bayesian learning, and federated methods.[73]
geometric deep learninggenerative modelingBayesian learningfederated learningadversarial robustness
CV and official profile
The current lab profile and publication list were inspected; no standalone CV was verified.[73]
Selected work
Representative records, not a complete publication list. Metadata confirms attribution; it does not independently replicate a result.
Patent evidence
2 catalogued patent records · 2 identified families
Coverage: Partial inventor search
Two family representatives found: GM computer vision and joint GM/Bar-Ilan speaker separation. The historical academic homepage independently matches GM co-inventors Dan Levi/Noa Garnett and the obstacle-detection subject. Current BIU lab page was inaccessible on this attempt; attribution is not based on that failure.
Low-latency speaker separation
US20250087217A1 · Published 2025-03-13
Published patent document inspected
Publication assignee: Bar Ilan University; GM Global Technology Operations LLC
Named inventor Ethan Fetaya; exact name and technical/co-inventor continuity, including the historical StixelNet author list for the GM record. Assignee is the captured publication metadata, not a current-ownership determination.[234][254]
Unified deep convolutional neural net for free-space estimation, object detection and object pose estimation
US10474908B2 · Published 2019-11-12
Published patent document inspected
Publication assignee: GM Global Technology Operations LLC
Named inventor Ethan Fetaya; exact name and technical/co-inventor continuity, including the historical StixelNet author list for the GM record. Assignee is the captured publication metadata, not a current-ownership determination.[234][253]
Original report snapshot
Original evidence: not verified
No attributable patent record was verified in the bounded search.
Records are counted separately from identified families. Author-reported entries are labelled and may still need publication verification. Inventorship, publication-time applicant and current ownership are different facts. No legal-status, patentability or freedom-to-operate conclusion is made.
Scores prioritize research fit from 1–10; they are not probabilities.
Review: Reviewed with limitations
Proposed capability matches, not confirmed relationships. Scores are analyst judgments with low forecast confidence; researcher interests, capacity and feasibility need confirmation.
Internal connections
15 candidates
Connection 1
Jacob Goldberger
Original proposal
Proposal hypothesis: Fetaya's adversarial and Bayesian learning and Goldberger's domain-shift and uncertainty work meet on detecting confidently wrong predictions.[73][75][76]
First test and score details
First test
Proposed first test: Apply a fixed corruption suite to one public image benchmark and compare calibration, abstention and retained accuracy against an uncalibrated baseline.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 4
Why this rank
Rank 1/15; fit 10/10 (4 topic overlap + 3 complementarity + 3 feasible first test). Preserved original co-membership proposal in o01. No strictly higher-scoring candidate displaces this original. Direct methodological overlap and a bounded benchmark justify a high test-fit score, without claiming robustness in deployment. No automatic score boost for original membership. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Connection 2
Ori Ernst
Original proposal
Proposal hypothesis: Ernst's proposition-to-source alignment supplies auditable units for Fetaya's adversarial evaluation; the shared question is whether a summarizer preserves support under misleading input.[73][93][94][95]
First test and score details
First test
Proposed first test: Corrupt one supporting passage in an open multi-document corpus and compare unsupported propositions and retained counterevidence against an unaligned summarizer.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 4
Why this rank
Rank 2/15; fit 10/10 (4 topic overlap + 3 complementarity + 3 feasible first test). Preserved original co-membership proposal in o01. No strictly higher-scoring candidate displaces this original. A public-text benchmark is straightforward; robustness beyond the selected corruption family remains untested. No automatic score boost for original membership. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Connection 3
Amir Leshem
Proposal hypothesis: Fetaya's adversarial/federated learning and Leshem's robust distributed learning directly meet on poisoned updates over constrained networks.[73][111][112]
First test and score details
First test
Proposed first test: Compare two robust aggregation methods under identical poisoning and congestion settings on accuracy, bandwidth and recovery.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 4
Why this rank
Rank 3/15; fit 10/10 (4 topic overlap + 3 complementarity + 3 feasible first test). New pairing outside the frozen portfolio co-member graph. The attack budget and data heterogeneity must be frozen before comparing robustness. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Show 12 moreShow fewer internal connections
Connection 4
Orr Levy
Proposal hypothesis: Levy's biological generative-network models give Fetaya a concrete test of geometric constraints and adversarial robustness.[5][6][73]
First test and score details
First test
Proposed first test: Perturb a public biological network while preserving degree and compare a constrained model with an unconstrained generator on held-out failure prediction.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 3
Why this rank
Rank 4/15; fit 9/10 (3 topic overlap + 3 complementarity + 3 feasible first test). New pairing outside the frozen portfolio co-member graph. Biological plausibility checks must come from Levy's domain model rather than graph validity alone. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Connection 5
Yaara Erez
Proposal hypothesis: Erez's neural decoding tasks give Fetaya a setting to test robust learning under subject and sensor shifts.[28][29][73]
First test and score details
First test
Proposed first test: Perturb a public neural benchmark with held-out-subject and channel-dropout tests; compare robust learning with regularized linear decoding.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 3
Why this rank
Rank 5/15; fit 9/10 (3 topic overlap + 3 complementarity + 3 feasible first test). New pairing outside the frozen portfolio co-member graph. Improved prediction does not validate a brain mechanism or a clinical BCI. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Connection 6
Shahar Alon
Proposal hypothesis: Alon can define spatial/genomic measurement invariants and Fetaya can test whether geometric learning respects them under perturbation.[60][61][73]
First test and score details
First test
Proposed first test: Compare a geometric model with a non-spatial baseline under coordinate jitter and expression dropout in a public spatial dataset.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 3
Why this rank
Rank 6/15; fit 9/10 (3 topic overlap + 3 complementarity + 3 feasible first test). New pairing outside the frozen portfolio co-member graph. Biological validity is not guaranteed by geometric invariance. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Connection 7
Gonen Singer
Proposal hypothesis: Fetaya can stress-test learned predictions while Singer measures their downstream cost in prescriptive allocation.[73][79][80]
First test and score details
First test
Proposed first test: Inject covariate shifts into a synthetic demand model and compare predictive accuracy and realized allocation cost with a robust non-learning baseline.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 3
Why this rank
Rank 7/15; fit 9/10 (3 topic overlap + 3 complementarity + 3 feasible first test). New pairing outside the frozen portfolio co-member graph. Better prediction may worsen decisions; downstream cost must be measured explicitly. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Connection 8
Ilan Reuven Cohen
Proposal hypothesis: Fetaya's robust learning could provide uncertain predictions for Ilan Cohen's online resource-constrained decisions.[73][119][120]
First test and score details
First test
Proposed first test: Compare forecast-assisted and forecast-free allocation under adversarial prediction errors on regret and constraint violations.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 3
Why this rank
Rank 8/15; fit 9/10 (3 topic overlap + 3 complementarity + 3 feasible first test). New pairing outside the frozen portfolio co-member graph. A robust predictor is useful only if its downstream online decisions improve. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Connection 9
Yoli Shavit
Proposal hypothesis: Fetaya's robust geometric learning and Shavit's uncertainty/generalization direction meet directly on reliable complex-data representations.[73][127][128]
First test and score details
First test
Proposed first test: Benchmark a public multimodal model under held-out domains and missing modalities; compare calibration and robust accuracy.
Score components
- complementarity
- 2
- feasible first test
- 3
- topic overlap
- 4
Why this rank
Rank 9/15; fit 9/10 (4 topic overlap + 2 complementarity + 3 feasible first test). New pairing outside the frozen portfolio co-member graph. The two ML roles overlap; a distinct division between perturbation design and representation/calibration is needed. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work. Yoli Shavit biological/multimodal programme evidence is a 2025 announcement; current laboratory operation and biological experimental capability are not established.
Connection 10
Shimrit Shtern
Identity check needed: Conditional proposal: confirm the researcher identity and research interests before assessing this match.
Proposal hypothesis: Conditional on confirmation, Shtern's continuous robust optimization could provide a tractable reference for Fetaya's adversarial learning tests.[73][83][84][85]
First test and score details
First test
Conditional proposed first test: After identity confirmation, compare a small robust convex surrogate with empirical adversarial training on loss and constraint satisfaction.
Score components
- complementarity
- 3
- feasible first test
- 2
- topic overlap
- 3
Why this rank
Rank 10/15; fit 8/10 (3 topic overlap + 3 complementarity + 2 feasible first test). New pairing outside the frozen portfolio co-member graph. Optimization expertise is documented; equivalence to a non-convex neural robustness problem is not assumed. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work. Conditional: confirm Shimrit Shtern identity mapping and current institutional affiliation before any internal team assignment. The BIU directory and current Technion appointment leave affiliation unresolved. Documented optimization expertise is conditional on that mapping; excluded from confirmed-team claims.
Connection 11
Tamar Goldzak Mizrahi
Proposal hypothesis: Goldzak Mizrahi can define symmetry and physics constraints for Fetaya's geometric or generative learning on material structures.[73][132][133]
First test and score details
First test
Proposed first test: Test a small generative model on public structures and compare invalidity under symmetry-preserving perturbations with an unconstrained baseline.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 2
Why this rank
Rank 11/15; fit 8/10 (2 topic overlap + 3 complementarity + 3 feasible first test). New pairing outside the frozen portfolio co-member graph. Valid structures and low prediction error do not establish new stable materials. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Connection 12
Hila Chalutz-Ben Gal
Original proposal
Proposal hypothesis: Chalutz-Ben Gal can operationalize organizational responses to AI errors while Fetaya supplies controlled adversarial failure cases for an adoption study.[22][23][73]
First test and score details
First test
Proposed first test: Build matched task vignettes with robust and deliberately brittle model outputs; pretest whether the rubric distinguishes justified refusal from blind acceptance.
Score components
- complementarity
- 3
- feasible first test
- 2
- topic overlap
- 2
Why this rank
Rank 12/15; fit 7/10 (2 topic overlap + 3 complementarity + 2 feasible first test). Preserved original co-membership proposal in o01. An added candidate, Amir Leshem (10/10), ranks above this original because its stated pair-specific roles and first test score higher; this original is limited as follows: Organizational study design and approval precede recruitment; model robustness and worker adoption are different outcomes. No automatic score boost for original membership. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work. Original initiative conditions remain: o01 Hila Chalutz-Ben Gal: Conditional later role: design a separately approved researcher-use study of noticing and correcting unsupported recommendations.
Connection 13
Mor Weiss
Original proposal
Proposal hypothesis: Fetaya can define a federated-learning update constraint and Mor Weiss can formalize how to prove compliance without revealing client examples.[73][123]
First test and score details
First test
Proposed first test: For a toy update-norm constraint, compare a proof-based audit with plaintext checking on proof overhead and accepted malicious updates.
Score components
- complementarity
- 3
- feasible first test
- 2
- topic overlap
- 2
Why this rank
Rank 13/15; fit 7/10 (2 topic overlap + 3 complementarity + 2 feasible first test). Preserved original co-membership proposal in o01. An added candidate, Amir Leshem (10/10), ranks above this original because its stated pair-specific roles and first test score higher; this original is limited as follows: Norm compliance is not benign model behaviour; a security model is needed before claiming protection. No automatic score boost for original membership. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work. Original initiative conditions remain: o01 Mor Weiss: Conditional later role: only after a protected-input threat model exists, specify a computation or access-policy statement that can be proven without revealing inputs.
Connection 14
Moti Fridman
Proposal hypothesis: Fridman can define a temporal-cavity control objective while Fetaya supplies uncertainty-aware or robust model-based search.[34][73]
First test and score details
First test
Proposed first test: Optimize a low-dimensional cavity simulator with a Bayesian method and random search; compare target error and failed settings under parameter shifts.
Score components
- complementarity
- 3
- feasible first test
- 2
- topic overlap
- 2
Why this rank
Rank 14/15; fit 7/10 (2 topic overlap + 3 complementarity + 2 feasible first test). New pairing outside the frozen portfolio co-member graph. The ML method is transferable, but lab-safe control and physical validation remain separate. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work.
Connection 15
Alexandra Simonovsky
Identity check needed: Conditional proposal: confirm the researcher identity and research interests before assessing this match.
Proposal hypothesis: Conditional on confirmation, Simonovsky could coordinate prototype submissions for a robustness challenge whose adversarial tasks are designed by Fetaya.[16][17][73]
First test and score details
First test
Conditional proposed first test: After identity confirmation, draft a sandboxed challenge with fixed benign corruptions and a scoring rubric; test the instructions on sample outputs.
Score components
- complementarity
- 2
- feasible first test
- 1
- topic overlap
- 1
Why this rank
Rank 15/15; fit 4/10 (1 topic overlap + 2 complementarity + 1 feasible first test). New pairing outside the frozen portfolio co-member graph. The technical evaluation belongs to Fetaya; coordination experience does not establish AI-safety research expertise. Equal scores use existing-first, then stable researcher ID.
Conditions
Analyst proposal hypothesis, not an established collaboration, commitment, evidence-confidence rating or assessment of researcher quality; confirm participation and access before work. Conditional: confirm Alexandra Simonovsky identity and current role before any team assignment. Evidence supports AI-agent event coordination only; research appointment, teaching role and technical research expertise are not verified. Excluded from confirmed-team claims.
External connections
10 candidates
Connection 1
Michael Bronstein
University of Oxford
Original proposal
His geometric deep-learning and protein-design work could complement scientific structure domains and evaluation; no willingness is asserted.[73][74][442]
First test and score details
First test
Benchmark one graph-generative model under symmetry-preserving perturbations and chemically or physically constrained invalidity checks.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 4
Why this rank
Rank 1/10 after semantic revision; analyst score 10 = max(1, 4+3+3): topic overlap 4/4, complementarity 3/3, feasible first test 3/3. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked. Original remains first under these components; original status breaks equal-score ties only, without a prestige bonus.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified. The refreshed Oxford profile also lists an Aithyra scientific-director role; the original Oxford institution string is preserved.
Connection 2
Yarin Gal
University of Oxford
Proposed capability match: Ethan Fetaya's geometric deep learning, generative modeling can be paired with Yarin Gal's documented Bayesian deep learning, uncertainty estimation for geometry-aware safety tests for scientific generative models. The specific contribution is uncertainty and selective prediction; this transfer is an analyst hypothesis.[73][78][452]
First test and score details
First test
Compare uncertainty estimates with calibrated single-model and ensemble baselines under a predefined shift using a public scientific-generative-model benchmark with held-out structures and an explicitly separate textual-explanation subset. Compare calibration error, risk-coverage and confident-error rate with an unstructured model plus existing safety checks under identical data access.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 4
Why this rank
Rank 2/10 after semantic revision; analyst score 10 = max(1, 4+3+3): topic overlap 4/4, complementarity 3/3, feasible first test 3/3. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified.
Connection 3
Andrea Montanari
Stanford University
Proposed capability match: Ethan Fetaya's geometric deep learning, generative modeling can be paired with Andrea Montanari's documented high-dimensional statistics, posterior sampling for geometry-aware safety tests for scientific generative models. The specific contribution is high-dimensional statistical baselines; this transfer is an analyst hypothesis.[73][479]
First test and score details
First test
Compare a regularized low-complexity estimator with a flexible model while varying sample size and dimensionality using a public scientific-generative-model benchmark with held-out structures and an explicitly separate textual-explanation subset. Compare generalization error, calibration and the sample-size threshold with an unstructured model plus existing safety checks under identical data access.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 3
Why this rank
Rank 3/10 after semantic revision; analyst score 9 = max(1, 3+3+3): topic overlap 3/4, complementarity 3/3, feasible first test 3/3. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified.
Show 7 moreShow fewer external connections
Connection 4
Gregory Wornell
Massachusetts Institute of Technology
Proposed capability match: Ethan Fetaya's geometric deep learning, generative modeling can be paired with Gregory Wornell's documented signal processing, statistical inference for geometry-aware safety tests for scientific generative models. The specific contribution is joint statistical inference and information constraints; this transfer is an analyst hypothesis.[73][110][503]
First test and score details
First test
Compare full-data inference with task-specific compressed statistics at fixed communication or storage budget using a public scientific-generative-model benchmark with held-out structures and an explicitly separate textual-explanation subset. Compare estimation error, calibration and bits per valid decision with an unstructured model plus existing safety checks under identical data access.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 3
Why this rank
Rank 4/10 after semantic revision; analyst score 9 = max(1, 3+3+3): topic overlap 3/4, complementarity 3/3, feasible first test 3/3. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified.
Connection 5
Dan Jurafsky
Stanford University
Proposed capability match: Ethan Fetaya's geometric deep learning, generative modeling can be paired with Dan Jurafsky's documented natural-language processing, language and society for geometry-aware safety tests for scientific generative models. The specific contribution is language-output robustness, conditional on scientific explanations; this transfer is an analyst hypothesis.[73][464]
First test and score details
First test
Vary the framing and incentives of explanation prompts while keeping a generated structure fixed using a public scientific-generative-model benchmark with held-out structures and an explicitly separate textual-explanation subset. Compare explanation consistency and sycophancy; no molecule-validity claim with an unstructured model plus existing safety checks under identical data access.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 2
Why this rank
Rank 5/10 after semantic revision; analyst score 8 = max(1, 2+3+3): topic overlap 2/4, complementarity 3/3, feasible first test 3/3. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified. The official page lists sabbatical during 2026-2027; availability is not inferred. Contribution is restricted to the scope stated in the first test.
Connection 6
Mirella Lapata
University of Edinburgh
Proposed capability match: Ethan Fetaya's geometric deep learning, generative modeling can be paired with Mirella Lapata's documented natural-language processing, long-context understanding for geometry-aware safety tests for scientific generative models. The specific contribution is source-grounded scientific explanations; this transfer is an analyst hypothesis.[73][96][467]
First test and score details
First test
Compare extracted and generated evidence summaries accompanying fixed scientific-model predictions using a public scientific-generative-model benchmark with held-out structures and an explicitly separate textual-explanation subset. Compare unsupported claims and loss of contradictory evidence with an unstructured model plus existing safety checks under identical data access.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 2
Why this rank
Rank 6/10 after semantic revision; analyst score 8 = max(1, 2+3+3): topic overlap 2/4, complementarity 3/3, feasible first test 3/3. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified. Contribution is restricted to the scope stated in the first test.
Connection 7
Christopher Manning
Stanford University
Proposed capability match: Ethan Fetaya's geometric deep learning, generative modeling can be paired with Christopher Manning's documented natural-language inference, summarization for geometry-aware safety tests for scientific generative models. The specific contribution is entailment of generated scientific explanations, conditional on textual outputs; this transfer is an analyst hypothesis.[73][472]
First test and score details
First test
Audit the textual-explanation subset using contradicted and paraphrased source passages using a public scientific-generative-model benchmark with held-out structures and an explicitly separate textual-explanation subset. Compare unsupported explanation rate without claiming physical-model validity with an unstructured model plus existing safety checks under identical data access.
Score components
- complementarity
- 3
- feasible first test
- 3
- topic overlap
- 2
Why this rank
Rank 7/10 after semantic revision; analyst score 8 = max(1, 2+3+3): topic overlap 2/4, complementarity 3/3, feasible first test 3/3. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified. Contribution is restricted to the scope stated in the first test.
Connection 8
Caroline Uhler
Massachusetts Institute of Technology
Proposed capability match for Ethan Fetaya with Caroline Uhler: Causal modelling can test a scientific model's response to a defined intervention; an unspecified benchmark and modality removal do not establish causal ground truth.[73][129][497]
First test and score details
First test
Define a small synthetic structure-property causal graph with one controllable structural variable, nuisance acquisition variables and known intervention outcomes. Compare a predictive model and graph-constrained causal model trained on identical observational samples, then test intervention-property error on held-out structures. Keep any textual explanation subset out of this numeric causal test.
Score components
- complementarity
- 3
- feasible first test
- 2
- topic overlap
- 3
Why this rank
Rank 8/10 after semantic revision; analyst score 8 = max(1, 3+3+2): topic overlap 3/4, complementarity 3/3, feasible first test 2/3. Causal modelling can test a scientific model's response to a defined intervention; an unspecified benchmark and modality removal do not establish causal ground truth. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified. Post-review scope: Causal modelling can test a scientific model's response to a defined intervention; an unspecified benchmark and modality removal do not establish causal ground truth. This revised proposal awaits independent targeted re-review; simulated outcomes would establish model behavior only, not biological, clinical or deployed benefit.
Connection 9
Susan Athey
Stanford University
Proposed capability match for Ethan Fetaya with Susan Athey: Causal evaluation is an indirect method for assessing a scientific workflow once treatment and validity outcomes are specified, not a general claim about scientific prediction.[73][433]
First test and score details
First test
In a synthetic design-selection experiment, define treatment as use of a safety filter and outcome as accepted valid structures per fixed proposal budget. Randomize filter assignment across simulated batches with planted validity labels. Compare stratified versus pooled treatment-effect estimates against known effects, keeping generator and candidate batches identical; no real safety benefit is inferred.
Score components
- complementarity
- 2
- feasible first test
- 2
- topic overlap
- 2
Why this rank
Rank 9/10 after semantic revision; analyst score 6 = max(1, 2+2+2): topic overlap 2/4, complementarity 2/3, feasible first test 2/3. Causal evaluation is an indirect method for assessing a scientific workflow once treatment and validity outcomes are specified, not a general claim about scientific prediction. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified. Post-review scope: Causal evaluation is an indirect method for assessing a scientific workflow once treatment and validity outcomes are specified, not a general claim about scientific prediction. This revised proposal awaits independent targeted re-review; simulated outcomes would establish model behavior only, not biological, clinical or deployed benefit.
Connection 10
Michael Mitzenmacher
Harvard University
Proposed capability match for Ethan Fetaya with Michael Mitzenmacher: Randomized scheduling addresses benchmark execution infrastructure; it does not change or validate the scientific model's predictions.[73][101][478]
First test and score details
First test
Represent scientific-model requests as jobs with explicit arrival times, graph sizes, estimated compute demands and a fixed worker pool. Compare two-choice load balancing with uniform routing under the same seeded job stream; measure tail completion time and dropped requests while holding model outputs and validation policy fixed.
Score components
- complementarity
- 2
- feasible first test
- 2
- topic overlap
- 2
Why this rank
Rank 10/10 after semantic revision; analyst score 6 = max(1, 2+2+2): topic overlap 2/4, complementarity 2/3, feasible first test 2/3. Randomized scheduling addresses benchmark execution infrastructure; it does not change or validate the scientific model's predictions. A bounded offline comparison is specified; required datasets and domain assumptions must still be checked.
Conditions
Proposed fit, not an assertion of a new or active relationship. Independent review pending; forecast confidence low. Partner interest, capacity, data access and any required experimental approvals/resources are unverified. Post-review scope: Randomized scheduling addresses benchmark execution infrastructure; it does not change or validate the scientific model's predictions. This revised proposal awaits independent targeted re-review; simulated outcomes would establish model behavior only, not biological, clinical or deployed benefit.
Evidence & open questions
- No standalone CV or attributable patent record was verified.
Inspect claim ratings and independent review
Ethan Fetaya has source-grounded capabilities in machine learning, generative models, and ai safety, represented here by geometric deep learning, generative modeling, Bayesian learning. [73]
Moderate confidenceReview: reviewedThe sources establish public professional activity, not comparative quality, future performance, or willingness to participate.
Review record
- profiles_a: supports. A successful direct re-open of the BIU lab page states Fetaya's interests in AI safety, generative models, deep Bayesian models and geometric deep learning, and lists the 2026 weight-space adversarial-attacks paper.
Hypothesis for 2027–2031: Ethan Fetaya could explore geometry-aware safety tests for scientific generative models through the bounded first test described in this profile. [73][74]
Low confidenceReview: reviewedThe relevant threat model and domain-validity oracle must be defined before safety claims are possible.
Review record
- profiles_a: supports. The hypothesis is grounded in Fetaya's documented AI-safety and geometric-learning work and Bronstein's Oxford profile covering graph neural networks and protein design. A domain validity oracle is still required.