BEGIN:VCALENDAR VERSION:2.0 PRODID:-//jEvents 2.0 for Joomla//EN CALSCALE:GREGORIAN METHOD:PUBLISH BEGIN:VTIMEZONE TZID:America/New_York BEGIN:STANDARD DTSTART:20181104T010000 RDATE:20190310T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20191103T010000 RDATE:20200308T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20201101T010000 RDATE:20210314T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20211107T010000 RDATE:20220313T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20221106T010000 RDATE:20230312T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20231105T010000 RDATE:20240310T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20241103T010000 RDATE:20250309T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20251102T010000 RDATE:20260308T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20261101T010000 RDATE:20270314T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20271107T010000 RDATE:20280312T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20281105T010000 RDATE:20290311T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20291104T010000 RDATE:20300310T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20301103T010000 RDATE:20310309T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20311102T010000 RDATE:20320314T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20321107T010000 RDATE:20330313T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20331106T010000 RDATE:20340312T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:STANDARD DTSTART:20341105T010000 RDATE:20350311T030000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 TZNAME:America/New_York EST END:STANDARD BEGIN:DAYLIGHT DTSTART:20180421T153000 RDATE:20181104T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20190310T030000 RDATE:20191103T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20200308T030000 RDATE:20201101T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20210314T030000 RDATE:20211107T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20220313T030000 RDATE:20221106T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20230312T030000 RDATE:20231105T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20240310T030000 RDATE:20241103T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20250309T030000 RDATE:20251102T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20260308T030000 RDATE:20261101T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20270314T030000 RDATE:20271107T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20280312T030000 RDATE:20281105T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20290311T030000 RDATE:20291104T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20300310T030000 RDATE:20301103T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20310309T030000 RDATE:20311102T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20320314T030000 RDATE:20321107T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20330313T030000 RDATE:20331106T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT BEGIN:DAYLIGHT DTSTART:20340312T030000 RDATE:20341105T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 TZNAME:America/New_York EDT END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT UID:6de693ccae75c5fa4d2720523dcd94e5 CATEGORIES:Qualifying Exam CREATED:20190823T084022 SUMMARY:Grounded and Interpretable Learning-Based Models for Visual Recognition Tasks LOCATION:CBIM 22 DESCRIPTION;ENCODING=QUOTED-PRINTABLE:
Abstract:
Safety-critical applications (e.g., aut onomous vehicles, human-machine teaming, and automated medical diagnosis) o ften require the use of computational agents that are capable of understand ing and reasoning about the high-level content of real world scene images i n order to make rational and grounded decisions that can be trusted by huma ns. Many of these agents rely on machine learning-based models which are in creasingly being treated as black-boxes. One way to increase model interpre tability is to make explainability a core principle of the model, e.g., by forcing deep neural networks (DNNs) to explicitly learn grounded and interp retable features. I introduce two novel approaches for making convolutional neural networks (CNNs) more interpretable by utilizing explainability as a guiding principle when designing the model architecture.
I) I propos e a CNN architecture that utilizes "visual attributes" as a form of explana tion. Visual attributes are semantic properties that are shared across diff erent categories and are able to be recognized from visual data, e.g., has_ horns::true, fur_color::brown, beak_shape::pointed, etc. Existing visual at tribute-based models assume all attributes are 1) necessary to achieve high accuracy on the target task and 2) able to be easily recognized from visua l data. These assumptions are rarely true. Not all attributes are discrimin ative w.r.t. the target task because of redundancy and irrelevancy, and not all attributes can be accurately extracted from data because of limited tr aining data, noisy labels, visual subtlety, cluttered/complex images, etc. To solve these problems, I propose a novel neural network-based framework t hat can jointly and simultaneously 1) select a subset of k visual attribute s from a much larger set of initial attributes, 2) map low-level features t o the selected attributes, and 3) learn a classifier that uses the selected attributes as features for some target task. By identifying a high-quality , smaller set of attributes, we can produce more compact and less noisy exp lanations which are easier for humans to understand.
II) I propose a CNN architecture that utilizes "scenarios" as a form of explanation. The sc enario is a data-driven representation based on sets of frequently co-occur ring objects (and/or visual attributes). Scenes can be decomposed as combin ations of scenarios, e.g., a bathroom scene might consist of: {toilet, toil et paper} + {shower, towel, shampoo, soap} + {sink, mirror, toothbrush, too thpaste}. Scenarios capture the high-level context that exists between obje cts/attributes, and this information is useful for better understanding the content of a scene, e.g., the "screen" object plays different roles in {sc reen, remote control, cable box} and {screen, keyboard, mouse}. By exploiti ng context, the semantic information contained in a scene image can be effi ciently compressed. Instead of having to recognize hundreds of objects and determine the role each object plays w.r.t. a target recognition task; inst ead, we can recognize and analyze a very low-dimensional set of scenarios. When used as features for some visual recognition task, scenarios result in more compact explanations. I propose a novel CNN which consists of three p arts: 1) global pooling layers that identify the parts of an image the netw ork attends to when recognizing whether each scenario is present in an imag e, 2) layers that use a matrix factorization-based loss function to learn a dictionary of scenarios and predict the presence of each scenario for a gi ven image, and 3) layers equivalent to a multinomial logistic regression mo del that use scenarios as low-dimensional features for prediction on the ta rget task.
DTSTAMP:20240329T070722Z DTSTART;TZID=America/New_York:20190422T153000 SEQUENCE:0 TRANSP:OPAQUE END:VEVENT END:VCALENDAR