A5056268210- Advanced Multi-Objective Optimization Algorithms
- Optimal Experimental Design Methods
- Gaussian Processes and Bayesian Inference
- Probabilistic and Robust Engineering Design
- Atmospheric aerosols and clouds
- Domain Adaptation and Few-Shot Learning
- Atmospheric and Environmental Gas Dynamics
- Complex Network Analysis Techniques
- Generative Adversarial Networks and Image Synthesis
- Topic Modeling
- Machine Learning in Materials Science
- Machine Learning and Algorithms
- Stochastic processes and statistical mechanics
- Calibration and Measurement Techniques
- Machine Learning and Data Classification
- Cell Image Analysis Techniques
- Advanced Text Analysis Techniques
- Analytical Chemistry and Chromatography
- Spectroscopy and Quantum Chemical Studies
- Seismic Waves and Analysis
- Anomaly Detection Techniques and Applications
- Multimodal Machine Learning Applications
- Mathematical Approximation and Integration
- Single-cell and spatial transcriptomics
- demographic modeling and climate adaptation
Microsoft Research (United Kingdom)
2024-2025
University of Oxford
2018-2022
Birkbeck, University of London
2018
Variational ab initio methods in quantum chemistry stand out among other providing direct access to the wave function. This allows, principle, straightforward extraction of any observable interest, besides energy, but, practice, this is often technically difficult and computationally impractical. Here, we consider electron density as a central introduce novel method obtain accurate densities from real-space many-electron functions by representing with neural network that captures known...
Bayesian optimal experimental design (BOED) is a principled framework for making efficient use of limited resources. Unfortunately, its applicability hampered by the difficulty obtaining accurate estimates expected information gain (EIG) an experiment. To address this, we introduce several classes fast EIG estimators building on ideas from amortized variational inference. We show theoretically and empirically that these can provide significant gains in speed accuracy over previous...
Related DatabasesWeb of Science You must be logged in with an active subscription to view this.Article DataHistorySubmitted: 18 May 2020Accepted: 11 October 2021Published online: 31 January 2022KeywordsBayesian experimental design, expected information gain, multilevel Monte Carlo, nested expectation, stochastic gradient descentAMS Subject Headings62K05, 62L20, 65C05, 92C45, 94A17Publication DataISSN (print): 1064-8275ISSN (online): 1095-7197Publisher: Society for Industrial and Applied...
We introduce a fully stochastic gradient based approach to Bayesian optimal experimental design (BOED). Our utilizes variational lower bounds on the expected information gain (EIG) of an experiment that can be simultaneously optimized with respect both and parameters. This allows process carried out through single unified ascent procedure, in contrast existing approaches typically construct pointwise EIG estimator, before passing this estimator separate optimizer. provide number different...
We propose methods to strengthen the invariance properties of representations obtained by contrastive learning. While existing approaches implicitly induce a degree as are learned, we look more directly enforce in encoding process. To this end, first introduce training objective for learning that uses novel regularizer control how representation changes under transformation. show trained with perform better on downstream tasks and robust introduction nuisance transformations at test time....
In data-to-text Natural Language Generation (NLG) systems, computers need to find the right words describe phenomena seen in data. This paper focuses on problem of choosing appropriate verbs express direction and magnitude a percentage change (e.g., stock prices). Rather than simply using same again again, we present principled data-driven approach this based Shannon’s noisy-channel model so as bring variation naturalness into generated text. Our experiments three large-scale real-world news...
Fully supervised models are predominant in Bayesian active learning. We argue that their neglect of the information present unlabelled data harms not just predictive performance but also decisions about what to acquire. Our proposed solution is a simple framework for semi-supervised find it produces better-performing than either conventional learning or with randomly acquired data. It easier scale up approach. As well as supporting shift towards models, our findings highlight importance...
We present Causal Amortized Active Structure Learning (CAASL), an active intervention design policy that can select interventions are adaptive, real-time and does not require access to the likelihood. This policy, amortized network based on transformer, is trained with reinforcement learning a simulator of environment, reward function measures how close true causal graph posterior inferred from gathered data. On synthetic data single-cell gene expression simulator, we demonstrate empirically...
Variational ab-initio methods in quantum chemistry stand out among other providing direct access to the wave function. This allows principle straightforward extraction of any observable interest, besides energy, but practice this is often technically difficult and computationally impractical. Here, we consider electron density as a central introduce novel method obtain accurate densities from real-space many-electron functions by representing with neural network that captures known...
Learning meaningful representations of data that can address challenges such as batch effect correction and counterfactual inference is a central problem in many domains including computational biology. Adopting Conditional VAE framework, we show marginal independence between the representation condition variable plays key role both these challenges. We propose Contrastive Mixture Posteriors (CoMP) method uses novel misalignment penalty defined terms mixtures variational posteriors to...