Consistent and coherent relational representation learning
File(s)
Author(s)
Stromfelt, Harry
Type
Thesis
Abstract
Achieving human-like general intelligence is a long-standing goal of Artificial Intelligence (AI) research. A hallmark of this intelligence is the freeing of learned concepts from the particular experiences that led to their acquisition. In contrast, Deep Learning (DL) methods are typically unable to generalise in this manner. Inspired by this capacity, this thesis proposes and formalises a new approach to learning at the intersection of symbolic AI and representation learning. The contributions are as follows.
Firstly, in the context of Lifelong Reinforcement Learning, trade-o↵s are explored between traditional deep Q-learning and architectures that explicitly incorporate relational abstractions. Results suggest that including such inductive biases can improve the lifelong learning performance of an agent.
Next, disentangled representation learners, namely Variational AutoEncoders (VAEs), are combined with relation-decoders, towards obtaining stronger representations for downstream learning. The second contribution is to show that vanilla VAEs are susceptible to dataset ambiguities and may learn representations that fail to capture abstract properties of certain objects, negatively impacting downstream performance. Including an explicit relational inductive bias, however, mitigates this issue.
Whilst the above approaches resemble the learning of abstract relational concepts, their semantics are unclear. The third contribution is the introduction of a formal framework for relation learning, taking inspiration from formal semantics in symbolic AI. In particular, formal definitions and conditions for the consistency and coherence of relational systems are presented, underpinned by soft-structures, a DL counterpart to the model-theoretic structure.
As the final contribution, this thesis introduces a novel consistent and coherent relation-decoder model, Dynamic Comparator (DC), and a Partial Relation Transfer experimental setting that exposes the semantics of learned relations. Using a proposed practical consistency and coherence measurement, results show that DC is able to learn a more generalisable set of relations and suggest a strong correlation between soft-structure coherence and generalisation performance.
Firstly, in the context of Lifelong Reinforcement Learning, trade-o↵s are explored between traditional deep Q-learning and architectures that explicitly incorporate relational abstractions. Results suggest that including such inductive biases can improve the lifelong learning performance of an agent.
Next, disentangled representation learners, namely Variational AutoEncoders (VAEs), are combined with relation-decoders, towards obtaining stronger representations for downstream learning. The second contribution is to show that vanilla VAEs are susceptible to dataset ambiguities and may learn representations that fail to capture abstract properties of certain objects, negatively impacting downstream performance. Including an explicit relational inductive bias, however, mitigates this issue.
Whilst the above approaches resemble the learning of abstract relational concepts, their semantics are unclear. The third contribution is the introduction of a formal framework for relation learning, taking inspiration from formal semantics in symbolic AI. In particular, formal definitions and conditions for the consistency and coherence of relational systems are presented, underpinned by soft-structures, a DL counterpart to the model-theoretic structure.
As the final contribution, this thesis introduces a novel consistent and coherent relation-decoder model, Dynamic Comparator (DC), and a Partial Relation Transfer experimental setting that exposes the semantics of learned relations. Using a proposed practical consistency and coherence measurement, results show that DC is able to learn a more generalisable set of relations and suggest a strong correlation between soft-structure coherence and generalisation performance.
Version
Open Access
Date Issued
2023-11
Date Awarded
2024-03
Copyright Statement
Creative Commons Attribution NonCommercial ShareAlike Licence
Advisor
Russo, Alessandra
Dickens, Luke
d'Avila Garcez, Artur
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/L504786/1
Publisher Department
Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)