Matching Items (222)
Filtering by
- Creators: Arizona State University
- Member of: ASU Electronic Theses and Dissertations
- Status: Published
Description
In image classification tasks, images are often corrupted by spatial transformationslike translations and rotations. In this work, I utilize an existing method that uses
the Fourier series expansion to generate a rotation and translation invariant representation of closed contours found in sketches, aiming to attenuate the effects of distribution shift caused by the aforementioned transformations. I use this technique to
transform input images into one of two different invariant representations, a Fourier
series representation and a corrected raster image representation, prior to passing
them to a neural network for classification. The architectures used include convolutional neutral networks (CNNs), multi-layer perceptrons (MLPs), and graph neural
networks (GNNs). I compare the performance of this method to using data augmentation during training, the standard approach for addressing distribution shift, to see
which strategy yields the best performance when evaluated against a test set with
rotations and translations applied. I include experiments where the augmentations
applied during training both do and do not accurately reflect the transformations encountered at test time. Additionally, I investigate the robustness of both approaches
to high-frequency noise. In each experiment, I also compare training efficiency across
models. I conduct experiments on three data sets, the MNIST handwritten digit
dataset, a custom dataset (QD-3) consisting of three classes of geometric figures from
the Quick, Draw! hand-drawn sketch dataset, and another custom dataset (QD-345)
featuring sketches from all 345 classes found in Quick, Draw!. On the smaller problem space of MNIST and QD-3, the networks utilizing the Fourier-based technique to
attenuate distribution shift perform competitively with the standard data augmentation strategy. On the more complex problem space of QD-345, the networks using the
Fourier technique do not achieve the same test performance as correctly-applied data
augmentation. However, they still outperform instances where train-time augmentations mis-predict test-time transformations, and outperform a naive baseline model
where no strategy is used to attenuate distribution shift. Overall, this work provides
evidence that strategies which attempt to directly mitigate distribution shift, rather
than simply increasing the diversity of the training data, can be successful when
certain conditions hold.
ContributorsWatson, Matthew (Author) / Yang, Yezhou YY (Thesis advisor) / Kerner, Hannah HK (Committee member) / Yang, Yingzhen YY (Committee member) / Arizona State University (Publisher)
Created2023
Description
Scientific research encompasses a variety of objectives, including measurement, making predictions, identifying laws, and more. The advent of advanced measurement technologies and computational methods has largely automated the processes of big data collection and prediction. However, the discovery of laws, particularly universal ones, still heavily relies on human intellect. Even with human intelligence, complex systems present a unique challenge in discerning the laws that govern them. Even the preliminary step, system description, poses a substantial challenge. Numerous metrics have been developed, but universally applicable laws remain elusive. Due to the cognitive limitations of human comprehension, a direct understanding of big data derived from complex systems is impractical. Therefore, simplification becomes essential for identifying hidden regularities, enabling scientists to abstract observations or draw connections with existing knowledge. As a result, the concept of macrostates -- simplified, lower-dimensional representations of high-dimensional systems -- proves to be indispensable. Macrostates serve a role beyond simplification. They are integral in deciphering reusable laws for complex systems. In physics, macrostates form the foundation for constructing laws and provide building blocks for studying relationships between quantities, rather than pursuing case-by-case analysis. Therefore, the concept of macrostates facilitates the discovery of regularities across various systems. Recognizing the importance of macrostates, I propose the relational macrostate theory and a machine learning framework, MacroNet, to identify macrostates and design microstates. The relational macrostate theory defines a macrostate based on the relationships between observations, enabling the abstraction from microscopic details. In MacroNet, I propose an architecture to encode microstates into macrostates, allowing for the sampling of microstates associated with a specific macrostate. My experiments on simulated systems demonstrate the effectiveness of this theory and method in identifying macrostates such as energy. Furthermore, I apply this theory and method to a complex chemical system, analyzing oil droplets with intricate movement patterns in a Petri dish, to answer the question, ``which combinations of parameters control which behavior?'' The macrostate theory allows me to identify a two-dimensional macrostate, establish a mapping between the chemical compound and the macrostate, and decipher the relationship between oil droplet patterns and the macrostate.
ContributorsZhang, Yanbo (Author) / Walker, Sara I (Thesis advisor) / Anbar, Ariel (Committee member) / Daniels, Bryan (Committee member) / Das, Jnaneshwar (Committee member) / Davies, Paul (Committee member) / Arizona State University (Publisher)
Created2023