This work proposes a new computational framework for learning a structured generative model for real-world datasets. In particular, we propose to learn a Closed-loop Transcriptionbetween a multi-class, multi-dimensional data distribution and a Linear discriminative representation (CTRL) in the feature space that consists of multiple independent multi-dimensional linear subspaces. In particular, we argue that the optimal encoding and decoding mappings sought can be formulated as a two-player minimax game between the encoder and decoderfor the learned representation. A natural utility function for this game is the so-called rate reduction, a simple information-theoretic measure for distances between mixtures of subspace-like Gaussians in the feature space. Our formulation draws inspiration from closed-loop error feedback from control systems and avoids expensive evaluating and minimizing of approximated distances between arbitrary distributions in either the data space or the feature space. To a large extent, this new formulation unifies the concepts and benefits of Auto-Encoding and GAN and naturally extends them to the settings of learning a both discriminative and generative representation for multi-class and multi-dimensional real-world data. Our extensive experiments on many benchmark imagery datasets demonstrate tremendous potential of this new closed-loop formulation: under fair comparison, visual quality of the learned decoder and classification performance of the encoder is competitive and arguably better than existing methods based on GAN, VAE, or a combination of both. Unlike existing generative models, the so-learned features of the multiple classes are structured instead of hidden: different classes are explicitly mapped onto corresponding independent principal subspaces in the feature space, and diverse visual attributes within each class are modeled by the independent principal components within each subspace.
more »
« less
This content will become publicly available on October 1, 2024
On the complexity of Bayesian generalization
We consider concept generalization at a large scale in the diverse and natural visual spectrum. Established computational modes (i.e., rule-based or similarity-based) are primarily studied isolated and focus on confined and abstract problem spaces. In this work, we study these two modes when the problem space scales up, and the complexity of concepts becomes diverse. Specifically, at the representational level, we seek to answer how the complexity varies when a visual concept is mapped to the representation space. Prior psychology literature has shown that two types of complexities (i.e., subjective complexity and visual complexity) build an inverted-U relation. Leveraging the Representativeness of Attribute (RoA), we computationally confirm the following observation: Models use attributes with high RoA to describe visual concepts, and the description length falls in an inverted-U relation with the increment in visual complexity. At the computational level, we aim to answer how the complexity of representation affects the shift between the rule- and similarity-based generalization. We hypothesize that category-conditioned visual modeling estimates the co-occurrence frequency between visual and categorical attributes, thus potentially serving as the prior for the natural visual world. Experimental results show that representations with relatively high subjective complexity out-perform those with relatively low subjective complexity in the rule-based generalization, while the trend is the opposite in the similarity-based generalization.
more »
« less
- Award ID(s):
- 2015577
- NSF-PAR ID:
- 10469433
- Publisher / Repository:
- International Conference on Machine Learning (ICML 2023)
- Date Published:
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Semantic oppositeness is the natural counterpart of the much popular natural language processing concept, semantic similarity. Much like how semantic similarity is a measure of the degree to which two concepts are similar, semantic oppositeness yields the degree to which two concepts would oppose each other. This complementary nature has resulted in most applications and studies incorrectly assuming semantic oppositeness to be the inverse of semantic similarity. In other trivializations, “semantic oppositeness” is used interchangeably with “antonymy”, which is as inaccurate as replacing semantic similarity with simple synonymy. These erroneous assumptions and over-simplifications exist due, mainly, to either lack of information, or the computational complexity of calculation of semantic oppositeness. The objective of this research is to prove that it is possible to extend the idea of word vector embedding to incorporate semantic oppositeness, so that an effective mapping of semantic oppositeness can be obtained in a given vector space. In the experiments we present in this paper, we show that our proposed method achieves a training accuracy of 97.91% and a test accuracy of 97.82%, proving the applicability of this method even in potentially highly sensitive applications and dispelling doubts of over-fitting. Further, this work also introduces a novel, unanchored vector embedding method and a novel, inductive transfer learning process.more » « less
-
Abstract Gene Ontology (GO) is widely used in the biological domain. It is the most comprehensive ontology providing formal representation of gene functions (GO concepts) and relations between them. However, unintentional quality defects (e.g. missing or erroneous relations) in GO may exist due to the large size of GO concepts and complexity of GO structures. Such quality defects would impact the results of GO-based analyses and applications. In this work, we introduce a novel evidence-based lexical pattern approach for quality assurance of GO relations. We leverage two layers of evidence to suggest potentially missing relations in GO as follows. We first utilize related concept pairs (i.e. existing relations) in GO to extract relationship-specific lexical patterns, which serve as the first layer evidence to automatically suggest potentially missing relations between unrelated concept pairs. For each suggested missing relation, we further identify two other existing relations as the second layer of evidence that resemble the difference between the missing relation and the existing relation based on which the missing relation is suggested. Applied to the 15 December 2021 release of GO, this approach suggested a total of 866 potentially missing relations. Local domain experts evaluated the entire set of potentially missing relations, and identified 821 as missing relations and 45 indicate erroneous existing relations. We submitted these findings to the GO consortium for further validation and received encouraging feedback. These indicate that our evidence-based approach can be utilized to uncover missing relations and erroneous existing relations in GO.more » « less
-
The landscapes of many elementary, middle, and high school math classrooms have undergone major transformations over the last half-century, moving from drill-and-skill work to more conceptual reasoning and hands-on manipulative work. However, if you look at a college level calculus class you are likely to find the main difference is the professor now has a whiteboard marker in hand rather than a piece of chalk. It is possible that some student work may be done on the computer, but much of it contains the same type of repetitive skill building problems. This should seem strange given the advancements in technology that allow more freedom than ever to build connections between different representations of a concept. Several class activities have been developed using a combination of approaches, depending on the topic. Topics covered in the activities include Riemann Sums, Accumulation, Center of Mass, Volumes of Revolution (Discs, Washers, and Shells), and Volumes of Similar Cross-section. All activities use student note outlines that are either done in a whole group interactive-lecture approach, or in a group work inquiry-based approach. Some of the activities use interactive graphs designed on desmos.com and others use physical models that have been designed in OpenSCAD and 3D-printed for students to use in class. Tactile objects were developed because they should provide an advantage to students by enabling them to physically interact with the concepts being taught, deepening their involvement with the material, and providing more stimuli for the brain to encode the learning experience. Web-based activities were developed because the topics involved needed substantial changes in graphical representations (i.e. limits with Riemann Sums). Assessment techniques for each topic include online homework, exams, and online concept questions with an explanation response area. These concept questions are intended to measure students’ ability to use multiple representations in order to answer the question, and are not generally computational in nature. Students are also given surveys to rate the overall activities as well as finer grained survey questions to try and elicit student thoughts on certain aspects of the models, websites, and activity sheets. We will report on student responses to the activity surveys, looking for common themes in students’ thoughts toward specific attributes of the activities. We will also compare relevant exam question responses and online concept question results, including common themes present or absent in student reasoning.more » « less
-
more » « less
Abstract Most existing unsteady flow visualization techniques concentrate on the depiction of geometric patterns in flow, assuming the geometry information provides sufficient representation of the underlying physical characteristics, which is not always the case. To address this challenge, this work proposes to analyse the time‐dependent characteristics of the physical attributes measured along pathlines which can be represented as a series of time activity curves (TAC). We demonstrate that the temporal trends of these TACs can convey the relation between pathlines and certain well‐known flow features (e.g. vortices and shearing layers), which enables us to select pathlines that can effectively represent the physical characteristics of interest and their temporal behaviour in the unsteady flow. Inspired by this observation, a new TAC‐based unsteady flow visualization and analysis framework is proposed. The centre of this framework is a new similarity measure that compares the similarity of two TACs, from which a new spatio‐temporal, hierarchical clustering that classifies pathlines based on their physical attributes, and a TAC‐based pathline exploration and selection strategy are proposed. A visual analytic system incorporating the TAC‐based pathline clustering and exploration is developed, which also provides new visualizations to support the user exploration of unsteady flow using TACs. This visual analytic system is applied to a number of unsteady flow in 2D and 3D to demonstrate its utility. The new system successfully reveals the detailed structure of vortices, the relation between shear layer and vortex formation, and vortex breakdown, which are difficult to convey with conventional methods.
