On incorporating inductive biases into deep neural networks

Date

2022

Authors

Ramasinghe, Sameera

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

A machine learning (ML) algorithm can be interpreted as a system that learns to capture patterns in data distributions. Before the modern \emph{deep learning era}, emulating the human brain, the use of structured representations and strong inductive bias have been prevalent in building ML models, partly due to the expensive computational resources and the limited availability of data. On the contrary, armed with increasingly cheaper hardware and abundant data, deep learning has made unprecedented progress during the past decade, showcasing incredible performance on a diverse set of ML tasks. In contrast to \emph{classical ML} models, the latter seeks to minimize structured representations and inductive bias when learning, implicitly favoring the flexibility of learning over manual intervention. Despite the impressive performance, attention is being drawn towards enhancing the (relatively) weaker areas of deep models such as learning with limited resources, robustness, minimal overhead to realize simple relationships, and ability to generalize the learned representations beyond the training conditions, which were (arguably) the forte of classical ML. Consequently, a recent hybrid trend is surfacing that aims to blend structured representations and substantial inductive bias into deep models, with the hope of improving them. Based on the above motivation, this thesis investigates methods to improve the performance of deep models using inductive bias and structured representations across multiple problem domains. To this end, we inject a priori knowledge into deep models in the form of enhanced feature extraction techniques, geometrical priors, engineered features, and optimization constraints. Especially, we show that by leveraging the prior knowledge about the task in hand and the structure of data, the performance of deep learning models can be significantly elevated. We begin by exploring equivariant representation learning. In general, the real-world observations are prone to fundamental transformations (e.g., translation, rotation), and deep models typically demand expensive data-augmentations and a high number of filters to tackle such variance. In comparison, carefully designed equivariant filters possess this ability by nature. Henceforth, we propose a novel \emph{volumetric convolution} operation that can convolve arbitrary functions in the unit-ball ($\mathbb{B}^3$) while preserving rotational equivariance by projecting the input data onto the Zernike basis. We conduct extensive experiments and show that our formulations can be used to construct significantly cheaper ML models. Next, we study generative modeling of 3D objects and propose a principled approach to synthesize 3D point-clouds in the spectral-domain by obtaining a structured representation of 3D points as functions on the unit sphere ($\mathbb{S}^2$). Using the prior knowledge about the spectral moments and the output data manifold, we design an architecture that can maximally utilize the information in the inputs and generate high-resolution point-clouds with minimal computational overhead. Finally, we propose a framework to build normalizing flows (NF) based on increasing triangular maps and Bernstein-type polynomials. Compared to the existing NF approaches, our framework consists of favorable characteristics for fusing inductive bias within the model i.e., theoretical upper bounds for the approximation error, robustness, higher interpretability, suitability for compactly supported densities, and the ability to employ higher degree polynomials without training instability. Most importantly, we present a constructive universality proof, which permits us to analytically derive the optimal model coefficients for known transformations without training.

Description

Keywords

Citation

Source

Type

Thesis (PhD)

Book Title

Entity type

Access Statement

Open Access

License Rights

Restricted until

Downloads

File
Description