Meta tags:
description=Haggai Maron – Assistant Professor at the Technion and Senior Research Scientist at NVIDIA Research. Research in machine learning and deep learning for structured data.;

Headings (most frequently used words):

haggai, maron, publications, group, news, talks, media, teaching,

Text of the page (most frequently used words):
and (212), the (189), learning (113), for (93), that (81), #haggai (76), maron (76), summary (73), paper (68), graph (67), neural (62), networks (56), conference (53), equivariant (43), our (36), international (36), deep (36), with (35), gnns (32), expressive (31), data (29), subgraph (28), novel (27), from (27), machine (27), propose (27), graphs (26), power (26), models (24), architectures (24), 2025 (24), invariant (23), based (23), icml (22), architecture (22), point (21), representations (21), weight (21), can (20), are (20), network (20), 2024 (20), layers (19), yaron (19), lipman (19), gal (18), processing (18), neurips (18), symmetries (18), efficient (17), show (17), method (17), new (17), information (16), iclr (16), code (16), approach (16), chechik (16), fabrizio (16), frasca (16), 2023 (16), present (15), both (15), two (15), results (15), via (15), using (15), which (15), methods (15), github (14), spectral (14), framework (14), space (14), annual (14), systems (14), all (14), tasks (14), these (14), aviv (14), introduce (13), demonstrate (13), permutation (13), sets (13), node (13), problem (12), shape (12), over (12), image (12), presentation (12), analysis (12), parameter (12), guy (12), matching (11), prove (11), problems (11), this (11), linear (11), representation (11), 2019 (11), blog (11), their (11), between (11), techniques (11), 2021 (11), structure (11), symmetry (11), spaces (11), bar (11), shalom (11), high (10), arxiv (10), post (10), first (10), universal (10), model (10), across (10), 2022 (10), derek (10), lim (10), positional (10), encodings (10), attention (10), 2026 (10), today (10), nadav (9), order (9), yam (9), ethan (9), fetaya (9), develop (9), single (9), workshop (9), existing (9), performance (9), features (9), gnn (9), technion (8), surfaces (8), such (8), shapes (8), algorithm (8), dym (8), than (8), scalable (8), clouds (8), ben (8), multi (8), random (8), more (8), nodes (8), where (8), domain (8), use (8), generalization (8), feature (8), how (8), message (8), passing (8), oral (8), beatrice (8), bevilacqua (8), michael (8), bronstein (8), diverse (8), expressivity (8), moshe (8), eliasof (8), weights (8), eitan (8), kan (8), while (7), elements (7), slides (7), state (7), art (7), datasets (7), classification (7), siggraph (7), well (7), detection (7), provably (7), same (7), outperforms (7), surface (7), spotlight (7), function (7), universality (7), training (7), study (7), consistently (7), navon (7), design (7), mpnns (7), sharing (7), joint (7), significantly (7), yoav (7), gelberg (7), llm (7), context (7), msc (7), simple (6), video (6), acm (6), applying (6), standard (6), relaxations (6), achieving (6), equal (6), contribution (6), cloud (6), functions (6), different (6), benchmarks (6), tensor (6), multiple (6), input (6), generation (6), poster (6), general (6), fundamental (6), several (6), parameters (6), geometric (6), work (6), symmetric (6), loss (6), family (6), them (6), into (6), other (6), prior (6), its (6), expressiveness (6), gradients (6), theoretical (6), shamsian (6), through (6), encoding (6), comprehensive (6), strong (6), transformers (6), yaniv (6), without (5), computational (5), convex (5), relaxation (5), convolutional (5), meaningful (5), optimization (5), points (5), heli (5), hamu (5), concave (5), known (5), derive (5), provide (5), case (5), size (5), any (5), apply (5), building (5), institute (5), under (5), talk (5), learn (5), experiments (5), 2020 (5), principled (5), improve (5), large (5), previous (5), tracks (5), trained (5), hypernetwork (5), guided (5), geometry (5), laplacian (5), policies (5), understanding (5), stefanie (5), jegelka (5), sign (5), normalization (5), zhang (5), pass (5), effectively (5), merge (5), llms (5), hallucination (5), lora (5), fourteenth (5), kans (5), research (4), sensitive (4), solution (4), popular (4), itay (4), applied (4), level (4), properties (4), projection (4), operators (4), evaluation (4), 2018 (4), used (4), generative (4), distribution (4), polytopes (4), they (4), acceptance (4), rate (4), one (4), solve (4), characterization (4), hyper (4), question (4), tensors (4), second (4), groups (4), nimrod (4), segol (4), low (4), computer (4), vision (4), directly (4), test (4), non (4), hadar (4), science (4), approximation (4), class (4), compared (4), characterize (4), images (4), set (4), award (4), baselines (4), adaptation (4), task (4), winter (4), idan (4), achituve (4), supervised (4), implicit (4), designing (4), accuracy (4), have (4), rank (4), approaches (4), eli (4), meirom (4), forward (4), theoretically (4), not (4), log (4), agnostic (4), fine (4), tuning (4), subgraphs (4), address (4), bargaining (4), schemes (4), norm (4), group (4), process (4), polynomials (4), effectiveness (4), diffusion (4), metanetworks (4), product (4), alignment (4), epnn (4), foundation (4), 38th (4), theo (4), putterman (4), empirical (4), granola (4), kernels (4), generate (4), editing (4), limitations (4), topological (4), galron (4), token (4), ran (4), activations (4), spring (4), phd (4), light (3), display (3), flexible (3), content (3), global (3), type (3), semantic (3), segmentation (3), meirav (3), galun (3), quadratic (3), naturally (3), matrices (3), extension (3), mapping (3), three (3), competing (3), matan (3), like (3), focusing (3), parts (3), interpolation (3), modeling (3), short (3), analyze (3), generalize (3), conditionally (3), probably (3), local (3), higher (3), field (3), dimension (3), number (3), basis (3), approximate (3), main (3), there (3), map (3), times (3), haim (3), sampling (3), positions (3), reconstruction (3), matrix (3), applications (3), mlp (3), serviansky (3), studies (3), objects (3), compute (3), maps (3), weizmann (3), formulate (3), sparse (3), unsupervised (3), improved (3), inherent (3), outstanding (3), best (3), set2graph (3), jonathan (3), describe (3), real (3), auxiliary (3), rotation (3), often (3), shie (3), mannor (3), dynamics (3), empirically (3), even (3), able (3), knowledge (3), scene (3), vertex (3), finding (3), camera (3), policy (3), unlike (3), does (3), contraction (3), guarantees (3), leveraging (3), clip (3), text (3), domains (3), natural (3), language (3), few (3), daniel (3), generalizing (3), original (3), selection (3), generalizes (3), extending (3), connection (3), beyond (3), joshua (3), full (3), introducing (3), inspired (3), theory (3), propagation (3), rfp (3), eran (3), treister (3), make (3), current (3), shot (3), then (3), key (3), downstream (3), experimental (3), equivalent (3), pretrained (3), structural (3), pretraining (3), layer (3), specific (3), particularly (3), superior (3), enables (3), temporal (3), fast (3), mlps (3), settings (3), directed (3), inversion (3), quality (3), homomorphism (3), homp (3), centrality (3), gradmetanet (3), act (3), vit (3), activation (3), achieves (3), yftah (3), ziser (3), lol (3), aware (3), output (3), unified (3), histograph (3), openreview (3), kolmogorov (3), arnold (3), derivatives (3), shine (3), tutorial (3), artifacts (3), structured (3), nvidia (2), capable (2), self (2), reproducing (2), viewpoint (2), passive (2), allow (2), simulate (2), 2016 (2), isometric (2), anatomical (2), kezurer (2), seamless (2), convolution (2), operator (2), defined (2), allows (2), 2017 (2), covers (2), stronger (2), solutions (2), additional (2), asia (2), pcnn (2), consists (2), robust (2), varying (2), translation (2), atzmon (2), zero (2), adversarial (2), enabling (2), exploration (2), energies (2), many (2), including (2), euclidean (2), sinkhorn (2), dimensional (2), originated (2), lifted (2), program (2), mrf (2), assignment (2), qap (2), form (2), devise (2), journal (2), edge (2), characteristics (2), consider (2), continuous (2), providing (2), iccv (2), niv (2), improving (2), unseen (2), 33rd (2), controlling (2), upon (2), distinguish (2), reduced (2), multiplication (2), has (2), provable (2), powerful (2), obtained (2), during (2), considering (2), resulted (2), thesis (2), professor (2), open (2), complexity (2), recent (2), amg (2), prolongation (2), train (2), convergence (2), demonstrating (2), potential (2), developing (2), ronen (2), basri (2), algebraic (2), interview (2), litany (2), suggest (2), practical (2), arbitrary (2), find (2), shlomi (2), kyle (2), cranmer (2), eilam (2), gross (2), motivated (2), mixup (2), evaluations (2), auxilearn (2), differentiation (2), useful (2), combining (2), coherent (2), combines (2), improves (2), attributes (2), conditions (2), recently (2), dynamic (2), processes (2), epidemic (2), rlgn (2), should (2), reinforcement (2), related (2), connectivity (2), structures (2), microphone (2), speech (2), dereverberation (2), enhances (2), inside (2), jets (2), jet (2), focus (2), takes (2), account (2), common (2), pre (2), reconstruct (2), yoni (2), kasten (2), esan (2), processed (2), variants (2), aggregation (2), heterogeneous (2), clients (2), require (2), david (2), hypernetworks (2), carefully (2), basic (2), construct (2), step (2), mtl (2), game (2), certain (2), assumptions (2), nash (2), interactive (2), driven (2), those (2), defines (2), formulation (2), generalized (2), most (2), prominent (2), employing (2), shows (2), collections (2), weisfeiler (2), leman (2), discuss (2), christopher (2), morris (2), martin (2), grohe (2), karsten (2), manner (2), invariants (2), proposed (2), top (2), robinson (2), lingxiao (2), zhao (2), hierarchy (2), degree (2), sequences (2), enhance (2), link (2), intermediate (2), iterative (2), seeds (2), metric (2), leading (2), 37th (2), dvir (2), samuel (2), rami (2), ari (2), nir (2), darshan (2), prediction (2), build (2), representing (2), gmns (2), marc (2), law (2), select (2), small (2), learns (2), lack (2), explore (2), substantial (2), neurreps (2), subgraphormer (2), operating (2), itself (2), unifying (2), products (2), align (2), indicate (2), feed (2), called (2), strictly (2), less (2), result (2), bohang (2), capture (2), being (2), embeddings (2), carola (2), bibiane (2), schönlieb (2), towards (2), dataset (2), investigate (2), impact (2), reduce (2), conduct (2), moe (2), adaptive (2), extensive (2), within (2), time (2), tracksto4d (2), casual (2), videos (2), coarsening (2), associated (2), resulting (2), leverage (2), foldable (2), supernet (2), merging (2), limited (2), tmlr (2), edan (2), kinderman (2), soudry (2), heat (2), follow (2), newton (2), raphson (2), lens (2), quantitative (2), highlight (2), architectural (2), wang (2), cellular (2), hymn (2), walk (2), costs (2), outperforming (2), efficiency (2), supra (2), about (2), faster (2), identify (2), curvature (2), gradient (2), 39th (2), transformer (2), traditional (2), setup (2), scores (2), gray (2), box (2), complete (2), probabilities (2), los (2), contamination (2), aaai (2), hallucinations (2), final (2), wise (2), hod (2), end (2), regimes (2), elbaz (2), heuristics (2), attributed (2), charm (2), adir (2), dayan (2), contexts (2), adapters (2), publications (2), wis (2), topics (2), simons (2), faculty (2), biu (2), aiola (2), members (2), google (2), scholar (2), supplementary, presenting, opaque, occlusions, depth, parallax, illumination, variations, shadows, highlights, uses, spatial, modulators, whose, micron, sized, digitally, thin, optical, photography, iccp, anat, levin, michal, yarom, sdp, procrustes, generic, almost, returns, correct, gives, shahar, kovalsky, registration, sphere, parameterization, planar, flat, torus, frameworks, readily, success, human, body, automatic, landmark, noam, aigerman, miri, trope, ersin, yumer, vladimir, kim, toric, programming, scalability, derivation, suggests, integer, extended, doubly, stochastic, partial, injective, constraints, tight, restriction, volumetric, vice, versa, computationally, samplings, densities, central, convincingly, genus, devised, parameterizations, charts, generated, avineri, chart, idea, concepts, encapsulate, instances, minima, probability, extreme, 32nd, markov, energy, minimization, closed, sided, convergent, regularized, siam, imaging, sciences, kushinsky, value, respectively, generally, tuples, bell, numbers, orthogonal, bases, computed, constant, successfully, shamir, tackle, s_n, acts, permuting, coordinates, allowed, unavoidable, obtaining, covering, wraps, around, distortion, coverage, cnn, retrieval, control, relating, samples, sample, attacks, curve, lior, yariv, ofer, israelov, pytorch, isomorphic, tests, containing, just, scaled, identity, augmented, operation, interleaves, regression, review, doctoral, consortium, dissertation, summarizes, guidance, were, considered, irregular, aiming, characterizing, trade, off, showing, contrast, positive, semi, definite, entire, rates, classical, utility, system, solvers, ilay, luz, irad, yavneh, multigrid, element, reordering, composed, dss, approximators, maximal, compact, testing, application, particle, physics, favorable, 34th, ssl, pretext, region, deformations, encountered, sim, transformation, combine, procedure, six, adaptations, synthetic, furniture, improvement, wacv, targets, losses, objective, interactions, sufficient, property, suggested, partially, observed, spread, sequential, decision, tractable, scheme, prioritize, tested, increase, healthy, people, contain, resources, put, important, aspects, very, larger, closely, patterns, formalize, setups, gilad, yehudai, cope, arrays, presumed, applicability, harnesses, reverberant, spectrum, interspeech, yochai, yemini, sharon, gannot, perform, flavor, tagging, implement, determine, pairs, european, physical, sanmay, ganguly, secondary, recovers, minimizing, reprojection, designed, permutations, cameras, scenes, inexpensive, dror, moran, hodaya, koslowsky, motion, each, represented, derived, some, predefined, suitable, isomorphism, lower, bounds, further, increases, balasubramaniam, srinivasan, chen, cai, gopinath, balamurugan, accommodates, client, adopting, limit, makes, external, disclose, technical, report, acuna, jan, kautz, sanja, fidler, federated, combined, ordering, quantum, circuit, simulation, implemented, agent, obtain, significant, improvements, varieties, circuits, optimizing, viewing, combination, negotiate, reach, agreement, direction, update, unique, kenji, kawaguchi, demo, scale, shifting, collecting, prompts, minutes, adapt, generator, multitude, characterized, styles, rinon, patashnik, cohen, stylegan, nada, generators, much, simpler, enjoys, evaluate, modelnet40, topology, algebra, finkelshtein, chaim, baskin, define, laplace, embedding, dissimilarity, rather, instantiated, norms, sohir, maskey, ali, parviz, maximilian, thiessen, hannes, stärk, ylli, sadikaj, requires, smaller, adopted, works, bound, sun, 36th, rethinking, give, overview, setting, regime, background, extensions, outline, jmlr, bastian, rieck, nils, kriege, matthias, fey, borgwardt, story, far, signnet, basisnet, requisite, hence, eigenspaces, previously, notable, tess, smidt, suvrit, sra, alternative, ability, calculate, concrete, algorithmic, tools, evaluating, adding, polynomial, operations, omri, puny, bobak, kiani, draws, iteration, concatenates, steps, computing, dominant, eigenvectors, starting, concatenation, biases, composition, affine, variety, observation, procedures, biased, toward, inputs, narrow, range, values, interpolating, taking, centroid, rare, concept, long, tail, latent, equivariance, orthogonally, analytic, thus, inherit, take, struggle, head, resnet, blocks, lorraine, james, lucas, subset, possible, fashion, families, indistinguishable, exist, bruno, ribeiro, effective, reasons, overfitting, reason, diversity, dws, strategies, augmentation, adapted, similarly, having, yield, gains, contrastive, yan, augmentations, integrates, enhanced, mechanisms, suggesting, formulated, aimed, adheres, respects, labeled, produces, better, alignments, produced, algorithms, eigenspace, essentially, unifies, either, surprising, bounded, implying, argue, community, needs, shift, balanced, thorough, interplay, i̇smail, i̇lkan, ceylan, ron, levie, future, directions, foundations, extent, requiring, purely, still, preliminary, only, enough, depends, quantity, fabian, jogl, ostrovsky, thomas, gärtner, cross, transfer, modifying, assessing, effect, removing, reveal, interesting, observations, instance, observe, mode, monotonic, hild, 2nd, robin, walters, thereof, normalizes, adapting, generating, neighborhood, rnf, underscores, emerging, performing, inferring, originating, tailored, way, comparable, drastically, reducing, runtime, wuyue, encoder, track, cluster, super, coarsened, reveals, whereby, running, implements, yet, discover, exhibits, unexplored, optimizes, fuse, widths, against, various, sizes, modalities, sota, scenarios, hubara, supernets, initializations, denoising, autoencoder, exploits, special, exponential, discrete, kernel, banach, multimodal, kreis, roots, equation, fastly, converges, provides, reconstructions, inverts, sec, a100, gpu, sdxl, turbo, flux, opening, door, barak, meiri, yoad, tewel, shai, avidan, lightning, count, exactly, tree, referred, parallel, trees, significance, establishing, hierarchies, jingchu, gai, yiheng, liwei, aim, strengths, weaknesses, subsequently, inability, diameter, orientability, planarity, homology, classes, tdl, mcn, smcn, mitigate, blind, spots, ses, measures, drawing, perturbation, reduces, burden, addresses, mitigating, competitive, bag, southern, balancing, challenges, eigendecomposition, ambiguity, when, offer, connecting, per, slice, overhead, 56x, runtimes, benefit, ecml, pkdd, principles, preserves, neuron, decomposition, learned, inr, landscape, estimation, efficiently, supports, simultaneously, encompassing, probing, remaining, extremely, deployment, benefits, substantially, transferred, probes, loras, predictions, decompositions, predict, finetuning, 000, fourth, observable, consisting, term, signature, nuanced, quantify, uncertainty, 40th, artificial, intelligence, next, learnable, distributions, stage, applies, followed, evolution, leverages, history, refine, offers, robustness, sinai, historical, share, computation, accounts, zoo, margin, ami, tavory, evron, maya, bechler, speicher, ido, meta, derivative, base, trainable, aligns, also, draw, connections, introduces, constructing, construction, exhibit, making, excellent, choice, incorrect, unsupported, rely, isolated, traces, unify, signals, tokens, edges, attentional, flows, carry, casts, tackles, above, subsumes, experimentally, operate, systematic, establish, mild, longer, holds, together, reusing, frozen, own, innovations, overcomes, relatively, instruction, pipeline, updates, immediately, complex, answering, accessing, transforming, yewei, liu, xiyuan, yansheng, mao, muhan, introduction, teaching, hebrew, podcast, cirm, independent, internal, talks, media, received, see, recipient, alon, scholarship, integration, keynotes, tag, vatat, grant, years, 88m, nis, pis, unlocking, democratized, moonshot, anyone, interested, joining, contributions, detecting, misbehaviors, out, news, postdoc, alina, sudakov, yonathan, wolanowsky, yuval, aidan, ofir, berkeley, purdue, mit, csail, cambridge, bgu, close, collaborators, ayellet, tal, past, haggaimaron, email, completed, supervision, prof, primary, interest, specifically, mathematical, goal, twofold, understand, perspective, example, analyzing, world, involving, senior, scientist, tel, electrical, engineering, assistant,

Text of the page (random words):
both the activation history of nodes and the graph structure to refine features used for final prediction empirical results demonstrate that histograph offers strong performance that consistently improves traditional techniques with particularly strong robustness in deep gnns paper beyond next token probabilities learnable fast detection of hallucinations and data contamination on llm output distributions guy bar shalom fabrizio frasca derek lim yoav gelberg yftah ziser ran el yaniv gal chechik haggai maron the 40th annual aaai conference on artificial intelligence aaai 2026 spotlight at iclr 2025 workshop quantify uncertainty and hallucination in foundation models summary we propose that gray box analysis should leverage the complete observable output of llms consisting of both token probabilities and the complete token distribution sequences a unified data type we term los llm output signature we develop a transformer based approach to process los that theoretically guarantees approximation of existing techniques while enabling more nuanced analysis our approach achieves superior performance on hallucination and data contamination detection in gray box settings significantly outperforming existing baselines paper code blog post gl equivariant metanetworks for learning on low rank weight spaces theo putterman derek lim yoav gelberg michael m bronstein stefanie jegelka haggai maron the fourth learning on graphs conference log 2025 oral presentation summary we investigate the potential of learning on loras lol a setup where machine learning models learn and make predictions on datasets of lora weights we first identify the inherent parameter symmetries of low rank decompositions of weights and develop several symmetry aware invariant or equivariant lol models in diverse experiments we show that our lol architectures can process lora weights to predict clip scores finetuning data attributes and accuracy on downstream tasks over 50 000 times faster than standard evaluation paper beyond token probes hallucination detection via activation tensors with act vit guy bar shalom fabrizio frasca yaniv galron yftah ziser haggai maron 39th annual conference on neural information processing systems neurips 2025 summary we introduce act vit a vision transformer inspired model that can be effectively and efficiently applied to activation tensors and supports training on data from multiple llms simultaneously through comprehensive experiments encompassing diverse llms and datasets we demonstrate that act vit consistently outperforms traditional probing techniques while remaining extremely efficient for deployment our architecture benefits substantially from multi llm training achieves strong zero shot performance on unseen datasets and can be transferred effectively to new llms through fine tuning paper blog post gradmetanet an equivariant architecture for learning on gradients yoav gelberg yam eitan aviv navon aviv shamsian theo putterman michael m bronstein haggai maron 39th annual conference on neural information processing systems neurips 2025 summary we present a principled approach for designing architectures that process gradients guided by three principles 1 equivariant design that preserves neuron permutation symmetries 2 processing sets of gradients across multiple data points to capture curvature information and 3 efficient gradient representation through rank 1 decomposition we introduce gradmetanet a novel architecture for learning on gradients and prove universality results we demonstrate gradmetanet s effectiveness on diverse gradient based tasks for mlps and transformers such as learned optimization inr editing and loss landscape curvature estimation paper understanding and improving laplacian positional encodings for temporal gnns yaniv galron fabrizio frasca haggai maron eran treister moshe eliasof ecml pkdd 2025 summary we address key challenges in supra laplacian positional encodings for temporal graphs high eigendecomposition costs limited theoretical understanding and ambiguity about when to apply these encodings we offer a theoretical framework connecting supra laplacian encodings to per time slice encodings introduce novel methods to reduce computational overhead achieving up to 56x faster runtimes and conduct an extensive experimental study to identify which models tasks and datasets benefit most from these encodings paper balancing efficiency and expressiveness subgraph gnns with walk based centrality joshua southern yam eitan guy bar shalom michael bronstein haggai maron fabrizio frasca international conference on machine learning icml 2025 summary we propose hymn an expressive and efficient approach that combines subgraph gnns and structural encodings ses via walk based centrality measures by drawing a connection to perturbation analysis we highlight the effectiveness of centrality based sampling which significantly reduces the computational burden of subgraph gnns hymn effectively addresses the expressiveness limitations of mpnns while mitigating the computational costs of subgraph gnns outperforming other subgraph sampling approaches while being competitive with full bag subgraph gnns paper code topological blind spots understanding and extending topological deep learning through the lens of expressivity yam eitan yoav gelberg guy bar shalom fabrizio frasca michael bronstein haggai maron international conference on learning representations iclr 2025 oral presentation summary we aim to explore both the strengths and weaknesses of higher order message passing homp s expressive power and subsequently design novel architectures to address its limitations we demonstrate homp s inability to distinguish between topological objects based on fundamental topological and metric properties such as diameter orientability planarity and homology we then develop two new classes of tdl models multi cellular networks mcn and scalable multi cellular networks smcn which mitigate many of homp s expressivity limitations paper homomorphism expressivity of spectral invariant graph neural networks jingchu gai yiheng du bohang zhang haggai maron liwei wang international conference on learning representations iclr 2025 oral presentation summary we address the fundamental question of the power of spectral invariants in gnns through the lens of homomorphism expressivity providing a comprehensive and quantitative analysis we prove that spectral invariant gnns can homomorphism count exactly a class of specific tree like graphs referred to as parallel trees and highlight the significance of this result in establishing quantitative expressiveness hierarchies across different architectural variants paper lightning fast image inversion and editing for text to image diffusion models dvir samuel barak meiri haggai maron yoad tewel nir darshan shai avidan gal chechik rami ben ari international conference on learning representations iclr 2025 summary we formulate diffusion inversion as finding the roots of an implicit equation and design a method based on newton raphson nr we derive an efficient guided formulation that fastly converges and provides high quality reconstructions and editing our solution guided newton raphson inversion inverts an image within 0 4 sec on an a100 gpu for few step models sdxl turbo and flux 1 opening the door for interactive image editing paper directed graph generation with heat kernels marc t law karsten kreis haggai maron tmlr 2025 summary we propose a denoising autoencoder based generative model that exploits the global structure of directed graphs via their laplacian dynamics and enables one shot generation our approach generalizes a special class of exponential kernels over discrete structures called diffusion kernels or heat kernels to the non symmetric case via reproducing kernel banach spaces we demonstrate that our model is able to generate directed graphs that follow the distribution of the training dataset even if it is multimodal paper foldable supernets scalable merging of transformers with different initializations and tasks edan kinderman itay hubara haggai maron daniel soudry tmlr 2025 summary we propose foldable supernet merge fs merge a method that optimizes a supernet to fuse original models using a feature reconstruction loss fs merge is simple data efficient and capable of merging models of varying widths we test fs merge against existing methods on mlps and transformers across various settings sizes tasks and modalities fs merge consistently outperforms them achieving sota results particularly in limited data scenarios paper a flexible equivariant framework for subgraph gnns via graph products and graph coarsening guy bar shalom yam eitan fabrizio frasca haggai maron 38th annual conference on neural information processing systems neurips 2024 best paper award at symmetry and geometry in neural representations workshop neurips 2024 summary we introduce a new subgraph gnns framework employing a graph coarsening function to cluster nodes into super nodes the product between the coarsened and original graph reveals an implicit structure whereby subgraphs are associated with specific sets of nodes by running generalized message passing on such graph product our method effectively implements an efficient yet powerful subgraph gnn we discover that the resulting node feature tensor exhibits new unexplored permutation symmetries and leverage this structure to characterize the associated linear equivariant layers paper fast encoder based 3d from casual videos via point track processing yoni kasten wuyue lu haggai maron 38th annual conference on neural information processing systems neurips 2024 summary we present tracksto4d a learning based approach that enables inferring 3d structure and camera positions from dynamic content originating from casual videos using a single efficient feed forward pass we propose operating directly over 2d point tracks as input and designing an architecture tailored for processing 2d point tracks tracksto4d is trained in an unsupervised way and can reconstruct a temporal point cloud and camera positions with accuracy comparable to state of the art methods while drastically reducing runtime by up to 95 paper granola adaptive normalization for graph neural networks moshe eliasof beatrice bevilacqua carola bibiane schönlieb haggai maron 38th annual conference on neural information processing systems neurips 2024 summary we propose granola a novel graph adaptive normalization layer unlike existing normalization layers granola normalizes node features by adapting to the specific characteristics of the graph particularly by generating expressive representations of its neighborhood structure via random node features rnf our extensive empirical evaluation underscores the superior performance of granola over existing normalization techniques emerging as the top performing method within the same time complexity of mpnns paper the empirical impact of neural parameter symmetries or lack thereof derek lim theo moe putterman robin walters haggai maron stefanie jegelka 38th annual conference on neural information processing systems neurips 2024 best paper award at hild 2024 2nd workshop on high dimensional learning dynamics icml 2024 summary we empirically investigate the impact of neural parameter symmetries by introducing new neural network architectures that have reduced parameter space symmetries we develop two methods of modifying standard neural networks to reduce parameter space symmetries and conduct a comprehensive experimental study assessing the effect of removing these symmetries our experiments reveal several interesting observations for instance we observe linear mode connectivity and monotonic linear interpolation in our networks without any alignment of weight spaces paper workshop paper towards foundation models on graphs an analysis on cross dataset transfer of pretrained gnns fabrizio frasca fabian jogl moshe eliasof matan ostrovsky carola bibiane schönlieb thomas gärtner haggai maron symmetry and geometry in neural representations neurreps neurips 2024 summary we study the extent to which pretrained graph neural networks can be applied across datasets requiring feature agnostic approaches we build upon a purely structural pretraining approach and propose an extension to capture feature information while still being feature agnostic our preliminary results indicate that embeddings from pretrained models improve generalization only with enough downstream data points and in a degree which depends on the quantity and properties of pretraining data paper future directions in foundations of graph machine learning christopher morris nadav dym haggai maron i̇smail i̇lkan ceylan fabrizio frasca ron levie derek lim michael bronstein martin grohe stefanie jegelka international conference on machine learning icml 2024 summary we argue that the graph machine learning community needs to shift its attention to developing a more balanced theory of graph machine learning focusing on a more thorough understanding of the interplay of expressive power generalization and optimization paper on the expressive power of spectral invariant graph neural networks bohang zhang lingxiao zhao haggai maron international conference on machine learning icml 2024 summary we introduce a novel message passing framework for designing spectral invariant gnns called eigenspace projection gnn epnn our comprehensive analysis shows that epnn essentially unifies all prior spectral invariant architectures and they are either strictly less expressive or equivalent to epnn we present a surprising result that epnn itself is bounded by a recently proposed class of subgraph gnns implying that all spectral invariant architectures are strictly less expressive than 3 wl paper equivariant deep weight space alignment aviv navon aviv shamsian ethan fetaya gal chechik nadav dym haggai maron international conference on machine learning icml 2024 summary we propose deep align a novel framework aimed at learning to solve the weight alignment problem we demonstrate that weight alignment adheres to two fundamental symmetries and propose a deep architecture that respects these symmetries our framework does not require any labeled data our experimental results indicate that a feed forward pass with deep align produces better or equivalent alignments compared to those produced by current optimization algorithms paper github subgraphormer unifying subgraph gnns and graph transformers via graph products guy bar shalom beatrice bevilacqua haggai maron international conference on machine learning icml 2024 summary we propose subgraphormer an architecture that integrates subgraph gnns and graph transformers combining enhanced expressive power and message passing mechanisms from subgraph gnns with attention and positional encodings from graph transformers our method is based on a new connection between subgraph gnns and product gra...