Semantic Information Pursuit for Multimodal Data Analysis
Lead Research Organisation:
Imperial College London
Department Name: Mathematics
Abstract
In 1948, Shannon published his famous paper "A Mathematical Theory of Communication" [88], which laid the foundations of information theory and led to a revolution in communication technologies. Shannon's fundamental contribution was to provide a precise way by which information could be represented,
quantified and transmitted. Critical to Shannon's ideas was the notion that the content of a message is irrelevant to its transmission, since any signal can be represented in terms of bits.
However, Shannon's theory has some limitations. In 1953, Weaver argued that there are three levels
of communication problems: the technical problem "How accurately can the symbols of
communication be transmitted?", the semantic problem "How precisely do the transmitted symbols
convey the desired meaning?", and the effectiveness problem "How effectively does the received
meaning affect conduct in the desired way?" Hence, a key limitation of Shannon's theory is that it
is limited to the technical problem.
This was also pointed out by Bar-Hillel and Carnap in 1953, who argued that "The Mathematical Theory of Communication, often referred to also as Theory (of Transmission) of Information, as practised nowadays, is not interested in the content of the symbols whose information it measures. The measures, as defined, for instance, by Shannon, have nothing to do with what these symbols symbolise, but only with the frequency of their occurrence." While Bar-Hillel and Carnap argued that "the fundamental concepts of the theory of semantic information can be defined in a straightforward way on the basis of the theory of inductive probability", their work was based primarily on logic rules that were applicable to a very restricted class of
signals (e.g. text). In the last 60 years there has been extraordinary progress in information theory,
signal, image and video processing, statistics, machine learning and optimization, which have led
to dramatic improvements in speech recognition, machine translation, and computer vision technologies.
However, the fundamental question of how to represent, quantify and transmit semantic is what this programme of research shall address.
quantified and transmitted. Critical to Shannon's ideas was the notion that the content of a message is irrelevant to its transmission, since any signal can be represented in terms of bits.
However, Shannon's theory has some limitations. In 1953, Weaver argued that there are three levels
of communication problems: the technical problem "How accurately can the symbols of
communication be transmitted?", the semantic problem "How precisely do the transmitted symbols
convey the desired meaning?", and the effectiveness problem "How effectively does the received
meaning affect conduct in the desired way?" Hence, a key limitation of Shannon's theory is that it
is limited to the technical problem.
This was also pointed out by Bar-Hillel and Carnap in 1953, who argued that "The Mathematical Theory of Communication, often referred to also as Theory (of Transmission) of Information, as practised nowadays, is not interested in the content of the symbols whose information it measures. The measures, as defined, for instance, by Shannon, have nothing to do with what these symbols symbolise, but only with the frequency of their occurrence." While Bar-Hillel and Carnap argued that "the fundamental concepts of the theory of semantic information can be defined in a straightforward way on the basis of the theory of inductive probability", their work was based primarily on logic rules that were applicable to a very restricted class of
signals (e.g. text). In the last 60 years there has been extraordinary progress in information theory,
signal, image and video processing, statistics, machine learning and optimization, which have led
to dramatic improvements in speech recognition, machine translation, and computer vision technologies.
However, the fundamental question of how to represent, quantify and transmit semantic is what this programme of research shall address.
Planned Impact
The proposed information-theoretic framework for characterizing information
content in multimodal data combines principles from information physics with probabilistic models that capture rich semantic and contextual relationships between data modalities and tasks. These information measures will be used to develop novel statistical methods for deriving minimal sufficient representations of multimodal data that are invariant to some nuisance factors as well as novel domain adaptation techniques that mitigate the impact of data transformations on information content by finding optimal data transformations.
The computation of such optimal representations and transformations for classification and perception tasks will require solving nonconvex optimization problems for which novel optimization algorithms with provable
guarantees of convergence and global optimality will be developed.
The uncertainty of such information representations derived from multimodal data will be characterized via novel statistical sampling methods that are broadly applicable to various representation learning problems.
The information representations obtained from multiple modalities will be integrated by using a novel information theoretic approach to multi-modal data analysis called information pursuit, which uses a Bayesian model of the scene to determine what evidence to acquire from multiple data modalities, scales and locations, and to coherently integrate this evidence.
The proposed methods will be evaluated in various complex multimodal datasets, including text,
images, video, cellphone data, and body-worn cameras.
content in multimodal data combines principles from information physics with probabilistic models that capture rich semantic and contextual relationships between data modalities and tasks. These information measures will be used to develop novel statistical methods for deriving minimal sufficient representations of multimodal data that are invariant to some nuisance factors as well as novel domain adaptation techniques that mitigate the impact of data transformations on information content by finding optimal data transformations.
The computation of such optimal representations and transformations for classification and perception tasks will require solving nonconvex optimization problems for which novel optimization algorithms with provable
guarantees of convergence and global optimality will be developed.
The uncertainty of such information representations derived from multimodal data will be characterized via novel statistical sampling methods that are broadly applicable to various representation learning problems.
The information representations obtained from multiple modalities will be integrated by using a novel information theoretic approach to multi-modal data analysis called information pursuit, which uses a Bayesian model of the scene to determine what evidence to acquire from multiple data modalities, scales and locations, and to coherently integrate this evidence.
The proposed methods will be evaluated in various complex multimodal datasets, including text,
images, video, cellphone data, and body-worn cameras.
People |
ORCID iD |
Mark Girolami (Principal Investigator) |
Publications
Barp A
(2022)
A Riemann-Stein kernel method
in Bernoulli
Barp A
(2019)
Minimum Stein Discrepancy Estimators
Barp A.
(2019)
Minimum Stein discrepancy estimators
in Advances in Neural Information Processing Systems
Bartels Simon
(2019)
Probabilistic linear solvers: a unifying view
in STATISTICS AND COMPUTING
Briol F
(2019)
Rejoinder: Probabilistic Integration: A Role in Statistical Computation?
in Statistical Science
Briol F
(2019)
Probabilistic Integration: A Role in Statistical Computation?
in Statistical Science
Chen W
(2018)
Stein Points
Chen W
(2019)
Stein Point Markov Chain Monte Carlo
Chen W.Y.
(2019)
Stein point Markov chain Monte Carlo
in 36th International Conference on Machine Learning, ICML 2019
Cockayne J
(2019)
Bayesian Probabilistic Numerical Methods
in SIAM Review
Cockayne J
(2018)
A Bayesian Conjugate Gradient Method
Cockayne J
(2019)
A Bayesian Conjugate Gradient Method (with Discussion)
in Bayesian Analysis
Duffin C
(2021)
Statistical finite elements for misspecified models.
in Proceedings of the National Academy of Sciences of the United States of America
Dunlop M.M.
(2018)
How deep are deep Gaussian processes?
in Journal of Machine Learning Research
Dunlop Matthew M.
(2018)
How Deep Are Deep Gaussian Processes?
in JOURNAL OF MACHINE LEARNING RESEARCH
Ellam L
(2018)
Stochastic Modelling of Urban Structure
Ellam L
(2018)
Stochastic modelling of urban structure.
in Proceedings. Mathematical, physical, and engineering sciences
Girolami M
(2021)
The statistical finite element method (statFEM) for coherent synthesis of observation data and model predictions
in Computer Methods in Applied Mechanics and Engineering
Glyn-Davies A
(2022)
Anomaly detection in streaming data with gaussian process based stochastic differential equations
in Pattern Recognition Letters
Gregory A
(2019)
The synthesis of data from instrumented structures and physics-based models via Gaussian processes
in Journal of Computational Physics
Hartmann M
(2022)
Lagrangian Manifold Monte Carlo on Monge Patches
Hartmann M
(2022)
Lagrangian Manifold Monte Carlo on Monge Patches
Manderson A
(2019)
Uncertainty Quantification of Density and Stratification Estimates with Implications for Predicting Ocean Dynamics
in Journal of Atmospheric and Oceanic Technology
Oates C
(2019)
Optimality Criteria for Probabilistic Numerical Methods
Oates C
(2019)
Bayesian Probabilistic Numerical Methods in Time-Dependent State Estimation for Industrial Hydrocyclone Equipment
in Journal of the American Statistical Association
Oates C
(2019)
Convergence rates for a class of estimators based on Stein's method
in Bernoulli