1<!DOCTYPE html> 2<html lang=""> 3 4 <!-- Head --> 5 <head><meta charset="utf-8"> 6<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no"> 7<meta http-equiv="X-UA-Compatible" content="IE=edge"> 8 9<title>Hubert Baniecki</title> 10<meta name="description" content=""> 11<meta name="author" content="Hubert Baniecki"> 12<meta name="keywords" content="Hubert Baniecki, hbaniecki, Baniecki, dalex, modelStudio"> 13<meta http-equiv="Cache-Control" content="no-store, max-age=0"> 14 15<!-- Open Graph --> 16 17<meta property="og:site_name" content="" /> 18<meta property="og:type" content="object" /> 19<meta property="og:title" content="" /> 20<meta property="og:url" content="https://hbaniecki.com/" /> 21<meta property="og:description" content="" /> 22<meta property="og:image" content="" /> 23 24 25<!-- Bootstrap & MDB --> 26<link href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" rel="stylesheet" integrity="sha256-DF7Zhf293AJxJNTmh5zhoYYIMs2oXitRfBjY+9L//AY=" crossorigin="anonymous"> 27<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/css/mdb.min.css" integrity="sha256-jpjYvU3G3N6nrrBwXJoVEYI/0zw8htfFnhT9ljN3JJw=" crossorigin="anonymous" /> 28 29<!-- Fonts & Icons --> 30<link rel="stylesheet" type="text/css" href="https://use.typekit.net/kcm7mlp.css"> 31<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/@fortawesome/[email protected]/css/all.min.css" integrity="sha256-CTSx/A06dm1B063156EVh15m6Y67pAjZZaQc89LLSrU=" crossorigin="anonymous"> 32<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/css/academicons.min.css" integrity="sha256-i1+4qU2G2860dGGIOJscdC30s9beBXjFfzjWLjBRsBg=" crossorigin="anonymous"> 33<!-- <link rel="stylesheet" type="text/css" href="https://fonts.googleapis.com/css?family=Roboto:300,400,500,700|Roboto+Slab:100,300,400,500,700|Roboto+Serif:100,300,400,500,700|Material+Icons"> --> 34<!-- <link rel="stylesheet" type="text/css" href="https://fonts.googleapis.com/css?family=Lato:300,400,500,700|Material+Icons"> --> 35<!-- <link rel="stylesheet" type="text/css" href="https://fonts.googleapis.com/css?family=Roboto:300,400,500,700|Material+Icons"> --> 36<!-- <link rel="stylesheet" href="https://fonts.cdnfonts.com/css/linux-libertine-o"> --> 37 38<!-- tooltip instead of title --> 39<link rel="stylesheet" href="https://unpkg.com/balloon-css/balloon.min.css"> 40 41<!-- Code Syntax Highlighting --> 42 43 44<!-- Styles --> 45 46<link rel="shortcut icon" href="/assets/img/favicon.ico"> 47 48<link rel="stylesheet" href="/assets/css/main.css"> 49<link rel="canonical" href="/"> 50 51<!-- GitHub button-->
52<script async defer src="https://buttons.github.io/buttons.js"></script>
52 53 54<!-- JQuery --> 55<!-- jQuery -->
56<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js" integrity="sha256-/xUj+3OJU5yExlq6GSYGSHk7tPXikynS7ogEvDej/m4=" crossorigin="anonymous"></script>
56 57 58 59<!-- Theming--> 60 61 62 63 64 65 <!-- Cronitor RUM --> 66
66<script async src="https://rum.cronitor.io/script.js"></script>
66 67
67<script> 68 window.cronitor = window.cronitor || function() { (window.cronitor.q = window.cronitor.q || []).push(arguments); }; 69 cronitor('config', { clientKey: '28195958db5d6c4a93b052b9befd86f4' }); 70 </script>
70 71 72 </head> 73 74 <body class="fixed-top-nav "> 75 76 <!-- Header --> 77 78 <header> 79 80 <!-- Nav Bar --> 81 <nav id="navbar" class="navbar navbar-light navbar-expand-sm fixed-top"> 82 <div class="container"> 83 84 <!-- Social Icons --> 85 <div class="social"> 86 <span class="contact-icon text-center"> 87 <a href="mailto:%68.%62%61%6E%69%65%63%6B%69@%75%77.%65%64%75.%70%6C"><i class="fas fa-envelope"></i></a> 88 89 <a href="https://scholar.google.com/citations?user=H72DRC0AAAAJ" target="_blank" title="Google Scholar"><i class="ai ai-google-scholar"></i></a> 90 <a href="https://dblp.org/pid/264/5189" target="_blank" title="DBLP"><i class="ai ai-dblp"></i></a> 91 92 93 <a href="https://github.com/hbaniecki" target="_blank" title="GitHub"><i class="fab fa-github"></i></a> 94 95 <a href="https://www.linkedin.com/in/hbaniecki" target="_blank" title="LinkedIn"><i class="fab fa-linkedin"></i></a> 96 <a href="https://x.com/hbaniecki" target="_blank" title="X (Twitter)"><i class="fab fa-x-twitter"></i></a> 97 98 99 100 101 102 103</span> 104 </div> 105 106 <!-- Navbar Toggle --> 107 <button class="navbar-toggler collapsed ml-auto" type="button" data-toggle="collapse" data-target="#navbarNav" aria-controls="navbarNav" aria-expanded="false" aria-label="Toggle navigation"> 108 <span class="sr-only">Toggle navigation</span> 109 <span class="icon-bar top-bar"></span> 110 <span class="icon-bar middle-bar"></span> 111 <span class="icon-bar bottom-bar"></span> 112 </button> 113 <div class="collapse navbar-collapse text-right" id="navbarNav"> 114 <ul class="navbar-nav ml-auto flex-nowrap"> 115 <!-- About --> 116 <li class="nav-item active"> 117 <a class="nav-link" href="/"> 118 bio <!-- home/about name --> 119 120 <span class="sr-only">(current)</span> 121 122 </a> 123 </li> 124 125 <!-- CV --> 126 <li class="nav-item"> 127 <a class="nav-link" href="/assets/pdf/cv_hbaniecki.pdf"> 128 cv 129 </a> 130 </li> 131 132 <!-- Other pages --> 133 134 135 136 137 138 139 140 <li class="nav-item "> 141 <a class="nav-link" href="/news/"> 142 news 143 144 </a> 145 </li> 146 147 148 149 <li class="nav-item "> 150 <a class="nav-link" href="/publications/"> 151 publications 152 153 </a> 154 </li> 155 156 157 158 <li class="nav-item "> 159 <a class="nav-link" href="/software/"> 160 software 161 162 </a> 163 </li> 164 165 166 167 <li class="nav-item "> 168 <a class="nav-link" href="/talks/"> 169 talks 170 171 </a> 172 </li> 173 174 175 176 <li class="nav-item "> 177 <a class="nav-link" href="/teaching/"> 178 teaching 179 180 </a> 181 </li> 182 183 184 185 </ul> 186 </div> 187 </div> 188 </nav> 189 190</header> 191 192 193 <!-- Content --> 194 195 <div class="container mt-4"> 196 <div class="post"> 197 198 <header class="post-header"> 199 <h2 class="post-title"> 200 Hubert <span class="font-weight-bold">Baniecki</span> 201 </h2> 202 203 <p class="post-description"> 204 <a href="https://en.uw.edu.pl" target="_blank">University of Warsaw</a> 205 ⢠206 <a href="https://ai.meta.com" target="_blank">Meta</a> 207 208 <!-- ⢠209 <a href="" target="_blank"></a> --> 210 </p> 211 212 </header> 213 214 <article> 215 216 <div class="profile float-right"> 217 218 <img class="img-fluid z-depth-1 rounded" src="/assets/img/photo.webp"> 219 220 221 <div class="address"> 222 h.baniecki(at)uw.edu.pl 223 </div> 224 225 </div> 226 227 228 <div class="clearfix"> 229 <p>I am a PhD candidate in Computer Science at the University of Warsaw, advised by <a href="https://scholar.google.com/citations?user=Af0O75cAAAAJ" target="_blank">Przemyslaw Biecek</a>. 230During my PhD, I interned at <a href="https://ai.meta.com" target="_blank">Meta</a> in New York (â26), and stayed at LMU Munich, hosted by <a href="https://scholar.google.com/citations?user=usVJeNN3xFAC" target="_blank">Eyke Hüllermeier</a> (â25) and <a href="https://scholar.google.com/citations?user=s34UckkAAAAJ" target="_blank">
230Bernd Bischl</a> (â24).</p> 231 232<p style="margin-bottom: 0pt;">My research focuses on <b>machine learning interpretability</b> (a.k.a. explainable AI):</p> 233<ul> 234 <li>understanding multimodal, visionâlanguage models (<em><a href="https://proceedings.neurips.cc/paper_files/paper/2025/hash/4e7686b7b147fac3352b24ab86ff78fb-Abstract-Conference.html" target="_blank">NeurIPSâ25</a>, <a href="https://proceedings.mlr.press/v267/zaigrajew25a.html" target="_blank">ICMLâ25</a></em>),</li> 235 <li>statistical foundations of interpretability methods (<em><a href="https://openreview.net/forum?id=LiUfN9h0Lx" target="_blank">ICLRâ25 Spotlight</a>, <a href="https://arxiv.org/abs/2602.16505" target="_blank">ICMLâ26</a></em>),</li> 236 <li>open-source software and benchmarks in this domain (<em><a href="https://www.jmlr.org/papers/v22/20-1473.html" target="_blank">JMLRâ21</a>, <a href="https://papers.nips.cc/paper_files/paper/2024/hash/eb3a9313405e2d4175a5a3cfcd49999b-Abstract-Datasets_and_Benchmarks_Track.html" target="_blank">NeurIPSâ24</a></em>),</li> 237 <li>applications of interpretability in medicine (<em><a href="https://openaccess.thecvf.com/content/WACV2025/html/Chrabaszcz_Aggregated_Attributions_for_Explanatory_Analysis_of_3D_Segmentation_Models_WACV_2025_paper.html" target="_blank">WACVâ25</a></em>) and beyond (<em><a href="https://doi.org/10.1073/pnas.2402028121" target="_blank">PNASâ24</a></em>).</li> 238</ul> 239 240<p>I actively serve as a reviewer for conferences like <em>NeurICMLR</em>, along with their workshops on interpretability, and journals like <em>JMLR</em>, <em>Machine Learning</em>, <em>Nature Communications</em>.</p> 241 242 </div> 243 244 <hr> 245 246 247 <div class="news"> 248 <h3>recent news <text class="archive">[<a href="/news">previous</a>]</text> </h3> 249 250 <div class="table-responsive"> 251 <table class="table table-sm table-borderless"> 252 <colgroup> 253 <col style="width:10.1%"> 254 <col style="width:89.1%"> 255 </colgroup> 256 257 258 <tr> 259 <th scope="row"> 260 2026 Sep 261 262 </th> 263 <td> 264 265 Two papers are accepted at <text class="venue">NeurIPS 2026</text>: <text class="title">The metagame of interpretability and meta-attributions</text> <a href="https://arxiv.org/abs/2605.06295" target="_blank"><i class="fa-solid fa-arrow-up-right-from-square"></i></a>, and <text class="title">Proxy-based approximation of Shapley and Banzhaf interactions 266</text> <a href="https://arxiv.org/abs/2605.22738" target="_blank"><i class="fa-solid fa-arrow-up-right-from-square"></i></a>. 267<span class="icon-mobile" aria-label="Final version to appear in conference proceedings." data-balloon-pos="left" data-balloon-length="medium"> <i class="fa-solid fa-circle-info"></i> </span> 268 269 270 </td> 271 </tr> 272 273 <tr> 274 <th scope="row"> 275 2026 Jun 276 277 </th> 278 <td> 279 280 Iâve joined <a href="https://ai.meta.com" target="_blank">Meta Superintelligence Labs</a> as a research scientist intern in New York for the summer of 2026. 281 282 283 </td> 284 </tr> 285 286 <tr> 287 <th scope="row"> 288 2026 Apr 289 290 </th> 291 <td> 292 293 A paper <text class="title">Functional decomposition and Shapley interactions for interpreting survival models</text> is accepted at <text class="venue">ICML 2026</text>. 294<a href="https://arxiv.org/abs/2602.16505" target="_blank"><i class="fa-solid fa-arrow-up-right-from-square"></i></a> 295<span class="icon-mobile" aria-label="Final version to appear in conference proceedings." data-balloon-pos="left" data-balloon-length="medium"> <i class="fa-solid fa-circle-info"></i> </span> 296 297 298 </td> 299 </tr> 300 301 <tr> 302 <th scope="row"> 303 2025 Nov 304 305 </th> 306 <td> 307 308 I stay in Italy until December for a 1-month research visit at the <a href="https://www.unipi.it/en" target="_blank">University of Pisa</a> hosted by <a href="https://scholar.google.com/citations?user=KZUaK6YAAAAJ" target="_blank">Riccardo Guidotti</a>. 309 310 311 </td> 312 </tr> 313 314 <tr> 315 <th scope="row"> 316 2025 Sep 317 318 </th> 319 <td> 320 321 A paper <text class="title">Explaining similarity in vision-language encoders with weighted Banzhaf interactions</text> is accepted at <text class="venue">NeurIPS 2025</text>. 322<a href="https://proceedings.neurips.cc/paper_files/paper/2025/hash/4e7686b7b147fac3352b24ab86ff78fb-Abstract-Conference.html" target="_blank"><i class="fa-solid fa-arrow-up-right-from-square"></i></a> 323<!-- <span class="icon-mobile" aria-label="Final version to appear in conference proceedings." data-balloon-pos="left" data-balloon-length="medium"> <i class="fa-solid fa-circle-info"></i> </span> --> 324 325 326 </td> 327 </tr> 328 329 <tr> 330 <th scope="row"> 331 2025 Sep 332 333 </th> 334 <td> 335 336 A paper <text class="title">Birds look like cars: Adversarial analysis of intrinsically interpretable deep learning</text> is accepted for publication in the <text class="venue">Machine Learning</text> journal. <a href="https://doi.org/10.1007/s10994-025-06896-w" target="_blank"><i class="fa-solid fa-arrow-up-right-from-square"></i></a> 337<!-- <span class="icon-mobile" aria-label="Final version to appear in the journal." data-balloon-pos="left" data-balloon-length="medium"> <i class="fa-solid fa-circle-info"></i> </span> --> 338 339 340 </td> 341 </tr> 342 343 <tr> 344 <th scope="row"> 345 2025 May 346 347 </th> 348 <td> 349 350 Foundation for Polish Science awarded me the START scholarship for young scientists. <a href="https://www.fnp.org.pl/aktualnosci/znamy-laureatow-konkursu-start-2025" target="_blank"><i class="fa-solid fa-arrow-up-right-from-square"></i></a> <span aria-label="News in the Polish language." data-balloon-pos="left" data-balloon-length="medium"><i class="fa-solid fa-flag-checkered flag-pl"></i></span> 351 352 353 </td> 354 </tr> 355 356 <tr> 357 <th scope="row"> 358 2025 May 359 360 </th> 361 <td> 362 363 A paper <text class="title">Interpreting CLIP with hierarchical sparse autoencoders</text> is accepted at <text class="venue">ICML 2025</text>. 364<a href="https://proceedings.mlr.press/v267/zaigrajew25a.html" target="_blank"><i class="fa-solid fa-arrow-up-right-from-square"></i></a> 365<!-- <span class="icon-mobile" aria-label="Final version to appear in conference proceedings." data-balloon-pos="left" data-balloon-length="medium"> <i class="fa-solid fa-circle-info"></i> </span> --> 366 367 368 </td> 369 </tr> 370 371 <tr> 372 <th scope="row"> 373 2025 Mar 374 375 </th> 376 <td> 377 378 I stay in Germany until April for a 1-month research visit at <a href="https://www.lmu.de/en" target="_blank">LMU Munich</a> hosted by <a href="https://scholar.google.com/citations?user=usVJeNN3xFAC" target="_blank">Eyke Hüllermeier</a>. 379 380 381 </td> 382 </tr> 383 384 <tr> 385 <th scope="row"> 386 2025 Jan 387 388 </th> 389 <td> 390 391 A paper <text class="title">Efficient and accurate explanation estimation with distribution compression</text> is accepted as a Spotlight at <text class="venue">ICLR 2025</text> (notable 5% of submissions). 392<a href="https://openreview.net/forum?id=LiUfN9h0Lx" target="_blank"><i class="fa-solid fa-arrow-up-right-from-square"></i></a> 393 394 395 </td> 396 </tr> 397 398 </table> 399 </div> 400 401</div> 402 403 404 405 <hr> 406 407 408 <div class="publications"> 409 <h3>selected publications <text class="archive">[<a href="/publications">full list</a>]</text> </h3> 410 <ol class="bibliography"></ol> 411 <ol class="bibliography"><li><!-- https://shopify.github.io/liquid/tags --> 412 413<div class="row"> 414 <div class="col-sm-2 abbr"> 415 416 417 418 <abbr class="badge"><a href="https://neurips.cc" target="_blank">NeurIPS</a></abbr> 419 420 421 422 423 </div> 424 425 <div id="baniecki2026metagame" class="col-sm-10"> 426 427 <div class="title">The metagame of interpretability and meta-attributions</div> 428 <!-- --> 429 <div class="author"> 430 431 432 433 434 435 436 437 438 <!-- convert to int --> 439 440 <em>H. Baniecki</em>, 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 <!-- convert to int --> 457 458 P. Biecek, 459 460 461 462 463 464 465 466 467 468 469 <!-- convert to int --> 470 471 F. Fumagalli 472 473 474 475 </div> 476 477 <div class="periodical"> 478 479 480 <text class="venue">Advances in Neural Information Processing Systems (<b>NeurIPS</b>)</text>, 2026 481 482 483 484 485 <br><text class="note">We introduce meta-attributions, which decompose any feature attribution into directional interactions via Shapley values, and apply them to interpret language and multimodal transformers.</text> 486 487 </div> 488 489 490 <div class="links"> 491 492 <a class="abstract btn btn-sm z-depth-0" role="button">Abstract</a> 493 494 495 <a href="https://arxiv.org/abs/2605.06295" class="btn btn-sm z-depth-0" role="button" target="_blank">Paper</a> 496 497 498 499 <a href="http://arxiv.org/abs/2605.06295" class="btn btn-sm z-depth-0" role="button" target="_blank">arXiv</a> 500 501 502 503 504 <a href="https://github.com/credibleai/metagame" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a> 505 506 507 508 509 </div> 510 511 <!-- Hidden abstract block --> 512 513 <div class="abstract hidden"> 514 <p>How can an arbitrary attribution method be generalized from first principles to capture interactions? We answer this with the metagame, a conceptual framework for quantifying second-order interaction effects of model explanations. We cast the attribution value Ïáµ¢ of feature i as a cooperative game among the other features and compute its Shapley value, which measures how much feature j influences the attribution of i, yielding the directional meta-attribution Ïâ±¼âáµ¢. By decomposing attribution itself rather than the model directly, meta-attributions extend any gradient- or attention-based method to interactions, uniting removal-based perturbations with model internals. Theoretically, we prove that meta-attributions sum to the first-order attribution they explain, a hierarchical decomposition that Shapley interactions and integrated Hessians turn out to perform implicitly. Empirically, we demonstrate that meta-attributions deliver insights across diverse interpretability applications: (i) quantifying token interactions in instruction-tuned language models, (ii) explaining cross-modal similarity in vision-language encoders, and (iii) interpreting text-to-image concepts in multimodal diffusion transformers.</p> 515 </div> 516 517 </div> 518</div> 519</li> 520<li><!-- https://shopify.github.io/liquid/tags --> 521 522<div class="row"> 523 <div class="col-sm-2 abbr"> 524 525 526 527 <abbr class="badge"><a href="https://icml.cc" target="_blank">ICML</a></abbr> 528 529 530 531 532 </div> 533 534 <div id="langbein2026functional" class="col-sm-10"> 535 536 <div class="title">Functional decomposition and Shapley interactions for interpreting survival models</div> 537 <!-- --> 538 <div class="author"> 539 540 541 542 543 544 545 546 547 <!-- convert to int --> 548 549 S. H. Langbein, 550 551 552 553 554 555 556 557 558 559 560 <!-- convert to int --> 561 562 <em>H. Baniecki</em>, 563 564 565 566 567 568 569 570 571 572 573 <!-- convert to int --> 574 575 F. Fumagalli, 576 577 578 579 580 581 582 583 584 585 586 <!-- convert to int --> 587 588 N. Koenen, 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 <!-- convert to int --> 605 606 M. N. Wright, 607 608 609 610 611 612 613 614 615 616 617 <!-- convert to int --> 618 619 J. Herbinger 620 621 622 623 </div> 624 625 <div class="periodical"> 626 627 628 <text class="venue">International Conference on Machine Learning (<b>ICML</b>)</text>, 2026 629 630 631 632 633 <br><text class="note">
633We propose a principled approach based on functional decomposition and Shapley values to explain time-dependent feature interactions in machine learning survival models.</text> 634 635 </div> 636 637 638 <div class="links"> 639 640 <a class="abstract btn btn-sm z-depth-0" role="button">Abstract</a> 641 642 643 <a href="https://openreview.net/forum?id=SldP4LGjdz" class="btn btn-sm z-depth-0" role="button" target="_blank">Paper</a> 644 645 646 647 <a href="http://arxiv.org/abs/2602.16505" class="btn btn-sm z-depth-0" role="button" target="_blank">arXiv</a> 648 649 650 651 652 <a href="https://github.com/sophhan/survshapiq" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a> 653 654 655 656 657 </div> 658 659 <!-- Hidden abstract block --> 660 661 <div class="abstract hidden"> 662 <p>Hazard and survival functions are natural, interpretable targets in time-to-event prediction, but their inherent non-additivity fundamentally limits standard additive explanation methods. We introduce Survival Functional Decomposition (SurvFD), a principled approach for analyzing feature interactions in machine learning survival models. By separating higher-order effects into time-dependent and time-independent components, SurvFD offers a previously unrecognized perspective on survival explanations, explicitly characterizing when and why additive explanations fail. Building on this theoretical decomposition, we propose SurvSHAP-IQ, which extends Shapley interactions to time-indexed functions, providing a practical estimator for higher-order, time-dependent interactions. Together, SurvFD and SurvSHAP-IQ establish a interaction- and time-aware interpretability framework for survival modeling, with broad applicability across time-to-event prediction tasks.</p> 663 </div> 664 665 </div> 666</div> 667</li> 668<li><!-- https://shopify.github.io/liquid/tags --> 669 670<div class="row"> 671 <div class="col-sm-2 abbr"> 672 673 674 675 <abbr class="badge"><a href="https://neurips.cc" target="_blank">NeurIPS</a></abbr> 676 677 678 679 680 </div> 681 682 <div id="baniecki2025explaining" class="col-sm-10"> 683 684 <div class="title">Explaining similarity in vision-language encoders with weighted Banzhaf interactions</div> 685 <!-- --> 686 <div class="author"> 687 688 689 690 691 692 693 694 695 <!-- convert to int --> 696 697 <em>H. Baniecki</em>, 698 699 700 701 702 703 704 705 706 707 708 <!-- convert to int --> 709 710 M. Muschalik, 711 712 713 714 715 716 717 718 719 720 721 <!-- convert to int --> 722 723 F. Fumagalli, 724 725 726 727 728 729 730 731 732 733 734 <!-- convert to int --> 735 736 B. Hammer, 737 738 739 740 741 742 743 744 745 746 747 <!-- convert to int --> 748 749 E. Hüllermeier, 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 <!-- convert to int --> 766 767 P. Biecek 768 769 770 771 </div> 772 773 <div class="periodical"> 774 775 776 <text class="venue">Advances in Neural Information Processing Systems (<b>NeurIPS</b>)</text>, 2025 777 778 779 780 781 <br><text class="note">We introduce faithful interaction explanations of CLIP and SigLIP models (FIxLIP), offering a unique, game-theoretic perspective on interpreting imageâtext similarity predictions.</text> 782 783 </div> 784 785 786 <div class="links"> 787 788 <a class="abstract btn btn-sm z-depth-0" role="button">Abstract</a> 789 790 791 <a href="https://proceedings.neurips.cc/paper_files/paper/2025/hash/4e7686b7b147fac3352b24ab86ff78fb-Abstract-Conference.html" class="btn btn-sm z-depth-0" role="button" target="_blank">Paper</a> 792 793 794 795 <a href="http://arxiv.org/abs/2508.05430" class="btn btn-sm z-depth-0" role="button" target="_blank">arXiv</a> 796 797 798 799 800 <a href="https://github.com/hbaniecki/fixlip" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a> 801 802 803 804 805 </div> 806 807 <!-- Hidden abstract block --> 808 809 <div class="abstract hidden"> 810 <p>Language-image pre-training (LIP) enables the development of vision-language models capable of zero-shot classification, localization, multimodal retrieval, and semantic understanding. Various explanation methods have been proposed to visualize the importance of input image-text pairs on the model's similarity outputs. However, popular saliency maps are limited by capturing only first-order attributions, overlooking the complex cross-modal interactions intrinsic to such encoders. We introduce faithful interaction explanations of LIP models (FIxLIP) as a unified approach to decomposing the similarity in vision-language encoders. FIxLIP is rooted in game theory, where we analyze how using the weighted Banzhaf interaction index offers greater flexibility and improves computational efficiency over the Shapley interaction quantification framework. From a practical perspective, we propose how to naturally extend explanation evaluation metrics, such as the pointing game and area between the insertion/deletion curves, to second-order interaction explanations. Experiments on the MS COCO and ImageNet-1k benchmarks validate that second-order methods, such as FIxLIP, outperform first-order attribution methods. Beyond delivering high-quality explanations, we demonstrate the utility of FIxLIP in comparing different models, e.g. CLIP vs. SigLIP-2.</p> 811 </div> 812 813 </div> 814</div> 815</li> 816<li><!-- https://shopify.github.io/liquid/tags --> 817 818<div class="row"> 819 <div class="col-sm-2 abbr"> 820 821 822 823 <abbr class="badge"><a href="https://icml.cc" target="_blank">ICML</a></abbr> 824 825 826 827 828 </div> 829 830 <div id="zaigrajew2025interpreting" class="col-sm-10"> 831 832 <div class="title">Interpreting CLIP with hierarchical sparse autoencoders</div> 833 <!-- --> 834 <div class="author"> 835 836 837 838 839 840 841 842 843 <!-- convert to int --> 844 845 V. Zaigrajew, 846 847 848 849 850 851 852 853 854 855 856 <!-- convert to int --> 857 858 <em>H. Baniecki</em>, 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 <!-- convert to int --> 875 876 P. Biecek 877 878 879 880 </div> 881 882 <div class="periodical"> 883 884 885 <text class="venue">International Conference on Machine Learning (<b>ICML</b>)</text>, 2025 886 887 888 889 890 <br><text class="note">
890We introduce the Matryoshka sparse autoencoder (MSAE) that establishes a state-of-the-art Pareto frontier between reconstruction quality and sparsity for interpreting CLIP models.</text> 891 892 </div> 893 894 895 <div class="links"> 896 897 <a class="abstract btn btn-sm z-depth-0" role="button">Abstract</a> 898 899 900 <a href="https://proceedings.mlr.press/v267/zaigrajew25a.html" class="btn btn-sm z-depth-0" role="button" target="_blank">Paper</a> 901 902 903 904 <a href="http://arxiv.org/abs/2502.20578" class="btn btn-sm z-depth-0" role="button" target="_blank">arXiv</a> 905 906 907 908 909 <a href="https://github.com/WolodjaZ/MSAE" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a> 910 911 912 913 914 </div> 915 916 <!-- Hidden abstract block --> 917 918 <div class="abstract hidden"> 919 <p>Sparse autoencoders (SAEs) are useful for detecting and steering interpretable features in neural networks, with particular potential for understanding complex multimodal representations. Given their ability to uncover interpretable features, SAEs are particularly valuable for analyzing vision-language models (e.g., CLIP and SigLIP), which are fundamental building blocks in modern large-scale systems yet remain challenging to interpret and control. However, current SAE methods are limited by optimizing both reconstruction quality and sparsity simultaneously, as they rely on either activation suppression or rigid sparsity constraints. To this end, we introduce Matryoshka SAE (MSAE), a new architecture that learns hierarchical representations at multiple granularities simultaneously, enabling a direct optimization of both metrics without compromise. MSAE establishes a state-of-the-art Pareto frontier between reconstruction quality and sparsity for CLIP, achieving 0.99 cosine similarity and less than 0.1 fraction of variance unexplained while maintaining 80% sparsity. Finally, we demonstrate the utility of MSAE as a tool for interpreting and controlling CLIP by extracting over 120 semantic concepts from its representation to perform concept-based similarity search and bias analysis in downstream tasks like CelebA. We make the codebase available at https://github.com/WolodjaZ/MSAE.</p> 920 </div> 921 922 </div> 923</div> 924</li> 925<li><!-- https://shopify.github.io/liquid/tags --> 926 927<div class="row"> 928 <div class="col-sm-2 abbr"> 929 930 931 932 <abbr class="badge"><a href="https://iclr.cc" target="_blank">ICLR</a></abbr> 933 934 935 936 937 938 <comment class="badge" style="cursor: default;">Spotlight</comment> 939 940 941 </div> 942 943 <div id="baniecki2025efficient" class="col-sm-10"> 944 945 <div class="title">Efficient and accurate explanation estimation with distribution compression</div> 946 <!-- --> 947 <div class="author"> 948 949 950 951 952 953 954 955 956 <!-- convert to int --> 957 958 <em>H. Baniecki</em>, 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 <!-- convert to int --> 975 976 G. Casalicchio, 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 <!-- convert to int --> 993 994 B. Bischl, 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 <!-- convert to int --> 1011 1012 P. Biecek 1013 1014 1015 1016 </div> 1017 1018 <div class="periodical"> 1019 1020 1021 <text class="venue">International Conference on Learning Representations (<b>ICLR</b>)</text>, 2025 (<b>Spotlight</b>) 1022 1023 1024 1025 1026 <br><text class="note">We introduce compress then explain (CTE) as a new paradigm for sample-efficient estimation of post-hoc explanations, including feature attributions, importance, and effects.</text> 1027 1028 </div> 1029 1030 1031 <div class="links"> 1032 1033 <a class="abstract btn btn-sm z-depth-0" role="button">Abstract</a> 1034 1035 1036 <a href="https://openreview.net/forum?id=LiUfN9h0Lx" class="btn btn-sm z-depth-0" role="button" target="_blank">Paper</a> 1037 1038 1039 1040 <a href="http://arxiv.org/abs/2406.18334" class="btn btn-sm z-depth-0" role="button" target="_blank">arXiv</a> 1041 1042 1043 1044 1045 <a href="https://github.com/hbaniecki/compress-then-explain" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a> 1046 1047 1048 1049 1050 </div> 1051 1052 <!-- Hidden abstract block --> 1053 1054 <div class="abstract hidden"> 1055 <p>We discover a theoretical connection between explanation estimation and distribution compression that significantly improves the approximation of feature attributions, importance, and effects. While the exact computation of various machine learning explanations requires numerous model inferences and becomes impractical, the computational cost of approximation increases with an ever-increasing size of data and model parameters. We show that the standard i.i.d. sampling used in a broad spectrum of algorithms for post-hoc explanation leads to an approximation error worthy of improvement. To this end, we introduce Compress Then Explain (CTE), a new paradigm of sample-efficient explainability. It relies on distribution compression through kernel thinning to obtain a data sample that best approximates its marginal distribution. CTE significantly improves the accuracy and stability of explanation estimation with negligible computational overhead. It often achieves an on-par explanation approximation error 2-3x faster by using fewer samples, i.e. requiring 2-3x fewer model evaluations. CTE is a simple, yet powerful, plug-in for any explanation method that now relies on i.i.d. sampling.</p> 1056 </div> 1057 1058 </div> 1059</div> 1060</li> 1061<li><!-- https://shopify.github.io/liquid/tags --> 1062 1063<div class="row"> 1064 <div class="col-sm-2 abbr"> 1065 1066 1067 1068 <abbr class="badge"><a href="https://neurips.cc" target="_blank">NeurIPS</a></abbr> 1069 1070 1071 1072 1073 </div> 1074 1075 <div id="muschalik2024shapiq" class="col-sm-10"> 1076 1077 <div class="title">shapiq: Shapley interactions for machine learning</div> 1078 <!-- --> 1079 <div class="author"> 1080 1081 1082 1083 1084 1085 1086 1087 1088 <!-- convert to int --> 1089 1090 M. Muschalik, 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 <!-- convert to int --> 1102 1103 <em>H. Baniecki</em>, 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 <!-- convert to int --> 1115 1116 F. Fumagalli, 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 <!-- convert to int --> 1128 1129 P. Kolpaczki, 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 <!-- convert to int --> 1141 1142 B. Hammer, 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 <!-- convert to int --> 1154 1155 E. Hüllermeier 1156 1157 1158 1159 </div> 1160 1161 <div class="periodical"> 1162 1163 1164 <text class="venue">Advances in Neural Information Processing Systems (<b>NeurIPS</b>)</text>, 2024 1165 1166 1167 1168 1169 <br><text class="note">We develop {shapiq}, an open-source Python package that implements several algorithms and benchmarks for efficiently approximating game-theoretic attribution and interaction indices.</text> 1170 1171 </div> 1172 1173 1174 <div class="links"> 1175 1176 <a class="abstract btn btn-sm z-depth-0" role="button">Abstract</a> 1177 1178 1179 <a href="https://papers.nips.cc/paper_files/paper/2024/hash/eb3a9313405e2d4175a5a3cfcd49999b
1179-Abstract-Datasets_and_Benchmarks_Track.html" class="btn btn-sm z-depth-0" role="button" target="_blank">Paper</a> 1180 1181 1182 1183 <a href="http://arxiv.org/abs/2410.01649" class="btn btn-sm z-depth-0" role="button" target="_blank">arXiv</a> 1184 1185 1186 1187 1188 <a href="https://github.com/mmschlk/shapiq" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a> 1189 1190 1191 1192 1193 </div> 1194 1195 <!-- Hidden abstract block --> 1196 1197 <div class="abstract hidden"> 1198 <p>Originally rooted in game theory, the Shapley Value (SV) has recently become an important tool in machine learning research. Perhaps most notably, it is used for feature attribution and data valuation in explainable artificial intelligence. Shapley Interactions (SIs) naturally extend the SV and address its limitations by assigning joint contributions to groups of entities, which enhance understanding of black box machine learning models. Due to the exponential complexity of computing SVs and SIs, various methods have been proposed that exploit structural assumptions or yield probabilistic estimates given limited resources. In this work, we introduce shapiq, an open-source Python package that unifies state-of-the-art algorithms to efficiently compute SVs and any-order SIs in an application-agnostic framework. Moreover, it includes a benchmarking suite containing 11 machine learning applications of SIs with pre-computed games and ground-truth values to systematically assess computational performance across domains. For practitioners, shapiq is able to explain and visualize any-order feature interactions in predictions of models, including vision transformers, language models, as well as XGBoost and LightGBM with TreeSHAP-IQ. With shapiq, we extend shap beyond feature attributions and consolidate the application of SVs and SIs in machine learning that facilitates future research. The source code and documentation are available at https://github.com/mmschlk/shapiq.</p> 1199 </div> 1200 1201 </div> 1202</div> 1203</li> 1204<li><!-- https://shopify.github.io/liquid/tags --> 1205 1206<div class="row"> 1207 <div class="col-sm-2 abbr"> 1208 1209 1210 1211 <abbr class="badge"><a href="https://www.pnas.org" target="_blank">PNAS</a></abbr> 1212 1213 1214 1215 1216 </div> 1217 1218 <div id="zhi2024increasing" class="col-sm-10"> 1219 1220 <div class="title">Increasing phosphorus loss despite widespread concentration decline in US rivers</div> 1221 <!-- --> 1222 <div class="author"> 1223 1224 1225 1226 1227 1228 1229 1230 1231 <!-- convert to int --> 1232 1233 W. Zhi, 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 <!-- convert to int --> 1245 1246 <em>H. Baniecki</em>, 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 <!-- convert to int --> 1258 1259 J. Liu, 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 <!-- convert to int --> 1271 1272 E. Boyer, 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 <!-- convert to int --> 1284 1285 C. Shen, 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 <!-- convert to int --> 1297 1298 G. Shenk, 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 <!-- convert to int --> 1310 1311 X. Liu, 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 <!-- convert to int --> 1323 1324 L. Li 1325 1326 1327 1328 </div> 1329 1330 <div class="periodical"> 1331 1332 <text class="venue">Proceedings of the National Academy of Sciences</text>, 2024 1333 1334 1335 1336 <br><text class="note">We reveal a paradox in US rivers with deep learning: phosphorus concentration is down over the last 40 years, particularly in urban areas, but total phosphorus loss is up due to climate change.</text> 1337 1338 </div> 1339 1340 1341 <div class="links"> 1342 1343 <a class="abstract btn btn-sm z-depth-0" role="button">Abstract</a> 1344 1345 1346 <a href="https://doi.org/10.1073/pnas.2402028121" class="btn btn-sm z-depth-0" role="button" target="_blank">Paper</a> 1347 1348 1349 1350 1351 1352 1353 <a href="https://github.com/LiReactiveWater/WT-DO-US-CE-LSTM" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a> 1354 1355 1356 1357 1358 </div> 1359 1360 <!-- Hidden abstract block --> 1361 1362 <div class="abstract hidden"> 1363 <p>The loss of phosphorous (P) from the land to aquatic systems has polluted waters and threatened food production worldwide. Systematic trend anal
1363ysis of P, a nonrenewable resource, has been challenging, primarily due to sparse and inconsistent historical data. Here, we leveraged intensive hydrometeorological data and the recent renaissance of deep learning approaches to fill data gaps and reconstruct temporal trends. We trained a multitask long short-term memory model for total P (TP) using data from 430 rivers across the contiguous United States (CONUS). Trend analysis of reconstructed daily records (1980â2019) shows widespread decline in concentrations, with declining, increasing, and insignificantly changing trends in 60%, 28%, and 12% of the rivers, respectively. Concentrations in urban rivers have declined the most despite rising urban population in the past decades; concentrations in agricultural rivers however have mostly increased, suggesting not-as-effective controls of nonpoint sources in agriculture lands compared to point sources in cities. TP loss, calculated as fluxes by multiplying concentration and discharge, however exhibited an overall increasing rate of 6.5% per decade at the CONUS scale over the past 40 y, largely due to increasing river discharge. Results highlight the challenge of reducing TP loss that is complicated by changing river discharge in a warming climate.</p> 1364 </div> 1365 1366 </div> 1367</div> 1368</li> 1369<li><!-- https://shopify.github.io/liquid/tags --> 1370 1371<div class="row"> 1372 <div class="col-sm-2 abbr"> 1373 1374 1375 1376 <abbr class="badge"><a href="https://www.springer.com/journal/10618" target="_blank">DAMI</a></abbr> 1377 1378 1379 1380 1381 </div> 1382 1383 <div id="baniecki2023iema" class="col-sm-10"> 1384 1385 <div class="title">The grammar of interactive explanatory model analysis</div> 1386 <!-- --> 1387 <div class="author"> 1388 1389 1390 1391 1392 1393 1394 1395 1396 <!-- convert to int --> 1397 1398 <em>H. Baniecki</em>, 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 <!-- convert to int --> 1415 1416 D. Parzych, 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 <!-- convert to int --> 1433 1434 P. Biecek 1435 1436 1437 1438 </div> 1439 1440 <div class="periodical"> 1441 1442 <text class="venue">Data Mining and Knowledge Discovery</text>, 2023 1443 1444 1445 1446 <br><text class="note">We propose to juxtapose multiple complementary explanations, and show that an interactive sequential analysis of a model improves the accuracy and confidence of human decision-making.</text> 1447 1448 </div> 1449 1450 1451 <div class="links"> 1452 1453 <a class="abstract btn btn-sm z-depth-0" role="button">Abstract</a> 1454 1455 1456 <a href="https://doi.org/10.1007/s10618-023-00924-w" class="btn btn-sm z-depth-0" role="button" target="_blank">Paper</a> 1457 1458 1459 1460 <a href="http://arxiv.org/abs/2005.00497v4" class="btn btn-sm z-depth-0" role="button" target="_blank">arXiv</a> 1461 1462 1463 1464 1465 <a href="https://github.com/ModelOriented/modelStudio" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a> 1466 1467 1468 1469 1470 <a href="https://iema.drwhy.ai" class="btn btn-sm z-depth-0" role="button" target="_blank">Website</a> 1471 1472 </div> 1473 1474 <!-- Hidden abstract block --> 1475 1476 <div class="abstract hidden"> 1477 <p>The growing need for in-depth analysis of predictive models leads to a series of new methods for explaining their local and global properties. Which of these methods is the best? It turns out that this is an ill-posed question. One cannot sufficiently explain a black-box machine learning model using a single method that gives only one perspective. Isolated explanations are prone to misunderstanding, leading to wrong or simplistic reasoning. This problem is known as the Rashomon effect and refers to diverse, even contradictory, interpretations of the same phenomenon. Surprisingly, most methods developed for explainable and responsible machine learning focus on a
1477single-aspect of the model behavior. In contrast, we showcase the problem of explainability as an interactive and sequential analysis of a model. This paper proposes how different Explanatory Model Analysis (EMA) methods complement each other and discusses why it is essential to juxtapose them. The introduced process of Interactive EMA (IEMA) derives from the algorithmic side of explainable machine learning and aims to embrace ideas developed in cognitive sciences. We formalize the grammar of IEMA to describe human-model interaction. It is implemented in a widely used human-centered open-source software framework that adopts interactivity, customizability and automation as its main traits. We conduct a user study to evaluate the usefulness of IEMA, which indicates that an interactive sequential analysis of a model may increase the accuracy and confidence of human decision making.</p> 1478 </div> 1479 1480 </div> 1481</div> 1482</li> 1483<li><!-- https://shopify.github.io/liquid/tags --> 1484 1485<div class="row"> 1486 <div class="col-sm-2 abbr"> 1487 1488 1489 1490 <abbr class="badge"><a href="https://www.jmlr.org" target="_blank">JMLR</a></abbr> 1491 1492 1493 1494 1495 1496 <comment class="badge"><a href="https://community.amstat.org/jointscsg-section/awards/john-m-chambers" target="_blank">Award</a></comment> 1497 1498 1499 </div> 1500 1501 <div id="baniecki2021dalex" class="col-sm-10"> 1502 1503 <div class="title">dalex: Responsible machine learning with interactive explainability and fairness in Python</div> 1504 <!-- --> 1505 <div class="author"> 1506 1507 1508 1509 1510 1511 1512 1513 1514 <!-- convert to int --> 1515 1516 <em>H. Baniecki</em>, 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 <!-- convert to int --> 1533 1534 W. Kretowicz, 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 <!-- convert to int --> 1546 1547 P. Piatyszek, 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 <!-- convert to int --> 1564 1565 J. Wisniewski, 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 <!-- convert to int --> 1582 1583 P. Biecek 1584 1585 1586 1587 </div> 1588 1589 <div class="periodical"> 1590 1591 <text class="venue">Journal of Machine Learning Research</text>, 2021 1592 1593 1594 1595 <br><text class="note">Software implementing the grammar of interactive explanatory model analysis; having 30K+ monthly downloads, 1.5K+ GitHub stars, and 30+ real-world applications in its citation network.</text> 1596 1597 </div> 1598 1599 1600 <div class="links"> 1601 1602 <a class="abstract btn btn-sm z-depth-0" role="button">Abstract</a> 1603 1604 1605 <a href="https://www.jmlr.org/papers/v22/20-1473.html" class="btn btn-sm z-depth-0" role="button" target="_blank">Paper</a> 1606 1607 1608 1609 <a href="http://arxiv.org/abs/2012.14406" class="btn btn-sm z-depth-0" role="button" target="_blank">arXiv</a> 1610 1611 1612 1613 1614 <a href="https://github.com/modeloriented/dalex" class="btn btn-sm z-depth-0" role="button" target="_blank">Code</a> 1615 1616 1617 1618 1619 <a href="https://dalex.drwhy.ai/python/" class="btn btn-sm z-depth-0" role="button" target="_blank">Website</a> 1620 1621 </div> 1622 1623 <!-- Hidden abstract block --> 1624 1625 <div class="abstract hidden"> 1626 <p>In modern machine learning, we observe the phenomenon of opaqueness debt, which manifests itself by an increased risk of discrimination, lack of reproducibility, and deflated performance due to data drift. An increasing amount of available data and computing power results in the growing complexity of black-box predictive models. To manage these issues, good MLOps practice asks for better validation of model performance and fairness, higher explainability, and continuous monitoring. The necessity for deeper model transparency comes from both
1626scientific and social domains and is also caused by emerging laws and regulations on artificial intelligence. To facilitate the responsible development of machine learning models, we introduce dalex, a Python package which implements a model-agnostic interface for interactive explainability and fairness. It adopts the design crafted through the development of various tools for explainable machine learning; thus, it aims at the unification of existing solutions. This library's source code and documentation are available under open license at https://python.drwhy.ai.</p> 1627 </div> 1628 1629 </div> 1630</div> 1631</li></ol> 1632</div> 1633 1634 1635 </article> 1636 1637</div> 1638 1639 </div> 1640 1641 <!-- Footer --> 1642 1643 1644<footer class="fixed-bottom"> 1645 <div class="container mt-0"> 1646 © Copyright 2020â2026 Hubert Baniecki. 1647 Made with <a href="https://github.com/alshedivat/al-folio">al-folio</a> theme and <a href="https://fontawesome.com/icons">fontawesome</a> icons. 1648 1649 1650 1651 </div> 1652</footer> 1653 1654 1655 1656 <!-- JavaScripts --> 1657 <!-- jQuery -->
1658<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js" integrity="sha256-/xUj+3OJU5yExlq6GSYGSHk7tPXikynS7ogEvDej/m4=" crossorigin="anonymous"></script>
1658 1659 1660 <!-- Bootsrap & MDB scripts -->
1661<script src="https://cdn.jsdelivr.net/npm/@popperjs/[email protected]/dist/umd/popper.min.js" integrity="" crossorigin="anonymous"></script>
vendor: 1 bytes, line 1661
1661
1662<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/js/bootstrap.min.js" integrity="sha256-SyTu6CwrfOhaznYZPoolVw2rxoY7lKYKQvqbtqN93HI=" crossorigin="anonymous"></script>
vendor: 1 bytes, line 1662
1662
1663<script src="https://cdn.jsdelivr.net/npm/[email protected]/js/mdb.min.js" integrity="sha256-NdbiivsvWt7VYCt6hYNT3h/th9vSTL4EDWeGs5SN3DA=" crossorigin="anonymous"></script>
1663 1664 1665 <!-- Mansory & imagesLoaded -->
1666<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/masonry.pkgd.min.js" integrity="sha256-Nn1q/fx0H7SNLZMQ5Hw5JLaTRZp0yILA/FRexe19VdI=" crossorigin="anonymous"></script>
vendor: 1 bytes, line 1666
1666
1667<script defer src="https://cdn.jsdelivr.net/npm/imagesloaded@4/imagesloaded.pkgd.min.js"></script>
vendor: 1 bytes, line 1667
1667
1668<script defer src="/assets/js/mansory.js" type="text/javascript"></script>
1668 1669 1670 1671 1672<!-- Load Common JS -->
1673<script src="/assets/js/common.js"></script>
1673 1674 1675 1676<!-- MathJax -->
1677<script type="text/javascript"> 1678 window.MathJax = { 1679 tex: { 1680 tags: 'ams' 1681 } 1682 }; 1683</script>
vendor: 1 bytes, line 1683
1683
1684<script defer type="text/javascript" id="MathJax-script" src="https://cdn.jsdelivr.net/npm/[email protected]/es5/tex-mml-chtml.js"></script>
1684 1685<!--
1685<script defer src="https://polyfill.io/v3/polyfill.min.js?features=es6"></script>
1685 -->
1686<script src="https://cdn.jsdelivr.net/npm/[email protected]/polyfill.min.js" integrity="sha256-UHghvpiWJZumUbnyJ9MJk6hs9O0U1PgI0GhZ4U1v6Is=" crossorigin="anonymous"></script>
1686 1687 1688 1689 1690 </body> 1691</html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.