Example 100-word statements

Examples of 100-word statements from colleagues in Informatics.

With the kind permission of the people concerned, we are sharing a number of 100-word "Reason for Submission" statements for REF outputs selected by people in the school. The aim of sharing these examples is that they will hopefully give everyone ideas as to how they may improve their own 100-word statements.

Probably none of these 100-word statements are perfect but they give some very good examples of what makes a good 100 word statement including:

looking at how a paper has been cited, to give specific ways in which the paper has been impactful or followed up by other researchers
explaining how the ideas in the paper have led to further research, for example by underpinning a future research grant
giving good ways to express the fact that the submitted paper is one of a series of related papers [but only one can be submitted to REF]
explaining how the contributions in the paper have been used beyond the original area of the paper
clearly explaining in a single sentence why the paper is original and interesting - clearly linking to related software

Example statements

The paper proposes a novel solution to address the problem of probabilistic inference with mixed discrete-continuous data in statistical machine learning, which has many important applications in areas such as high-energy physics and computational biology. Its key contribution is to relax the strong assumptions of previous approaches without sacrificing robustness. The paper was accepted at the highest-profile conference in uncertainty in AI, and was awarded the Microsoft best paper award.

It formed the basis for a continuing collaboration with UCLA, Leuven and Trento, in terms of joint 3 follow-up papers and co-supervision of PhD students.

Introduced the idea of using pairs of word-like units extracted in an unsupervised way to provide a noisy top-down signal for representation learning from raw (untranscribed) speech. The learned representations capture phonetic distinctions better than standard (un-learned) features or those learned purely bottom-up. Others later applied this idea cross-lingually (Yuan et al., Interspeech 2016) and used it as a baseline for other approaches (He, Wang, and Livescu, ICLR 2017). This paper focussed on engineering applications, but led to later funding from NSF and ESRC to explore the idea introduced here as a model of perceptual learning in infants.

Presents a novel method to jointly segment unlabelled speech and cluster the segments into hypothesized word groupings. Uses a mathematically rigorous model and inference - unlike previous approaches, which required heuristics. Was a top five finalist for the annual Commonwealth Scholar Best Journal Article (for an article in any discipline by a Commonwealth Scholar within five years of PhD). We later scaled up this proof-of-concept system and showed it outperformed competing systems on large-vocabulary datasets (Kamper et al., 2017); the scaled up version was used by others to help with learning unsupervised semantic embeddings from speech (Chung et al., NIPS 2018).

Distributed training is crucial to training neural networks on big data sets, network communication is a bottleneck. This well-cited paper showed how to reduce network traffic by 50x without impacting the quality of the final model or convergence, by exchanging only the largest gradients. While the paper showed experiments on neural machine translation and vision, it has been adopted for speech recognition and language modeling. Impact was quick: 8 months after publication, Google, Stanford, NVIDIA, and Tsinghua posted a paper tuning our work to achieve even greater compression: "Deep Gradient Compression: Reducing the Communication Bandwidth for Distributed Training" (ICLR 2018).

This paper is the journal version of the first highly cited Sketch-a-Net paper that won Best Paper Prize at BMVC2015 conference. The Sketch-a-Net deep neural network was the first to show super-human performance at the Sketch recognition task in computer vision. This surprising outcome was widely reported in BBC News, Business Insider, New Scientist, etc. This paper underpinned many subsequent applications by the author’s group (CVPR’16x2, CVPR’18x3, CVPR’19) and the wide publicity of this result re-popularized the sketch analysis area of computer vision leading to takeup by other groups (USTC, GeorgiaTech, Surrey, Inception Institute, Surrey, ETHZ, Fudan, IIT-Madras, Trento)

The paper proposes a novel architecture where the neural network is factorised by the locomotion phase, which allows learning a wide range of motions. The system is highly scalable - can learn from 1.5GB of data while only requiring 10MB runtime memory.The paper is presented/published in SIGGRAPH 2017/ACM Transactions on Graphics, which is one of the best conferences/journals in computer science. The paper is widely shared in social media/news (NVIDIA News, TechCrunch etc). The university now has multiple commercial contracts with companies (AXYZ Design, NVidia) for this technology. It also led to a 780K project with Facebook.

This was the first deep neural network for fitting high-dimensional probability distributions where normalized test probabilities can be evaluated with one GPU-friendly network pass. It became a component in multiple probabilistic models and systems from other groups at top-tier venues, including "Inverse autoregressive flow" (NeurIPS 2016), "Inference networks" (ICML 2016). Similar masking was adopted in application-specific architectures, e.g., Pixel RNNs (Google DeepMind 2016, >300 citations). Published code made state-of-the-art results reproducible. Theory of masking mechanism and notation checked with separate code published on author's website, confirming the presentation in paper was correct.

This new method for "likelihood-free inference" (estimating parameters of simulation-based models) exploited recent advances in machine learning to reduce the cost of inferring model parameters. This work was directly built on by: Lueckmann (NeurIPS 2017) for modelling neural systems; Alsing (MNRAS 2018) for cosmology; Trippe & Turner (NeurIPS BDLW, 2018); Chen & Gutmann (AISTATS, 2019). The paper was the topic of extensive 2-part critical discussion by Dennis Prangle (Lancaster): "This paper certainly seems like a big step forwards for likelihood-free inference, making substantial progress on most of the limitations listed in my first post."

This was a new way to construct deep neural density estimators that gave state-of-the-art results, while also presenting connections betwen other methods and their tradeoffs. It was an oral at NeurIPS 2017 (~1% submissions top machine learning venue). It was reimplemented by TensorFlow Probability, now a part of TensorFlow. It's been used within variational autoencoders by several groups, in reinforcement learning (Schroecker et al., 2019), and in likelihood-free inference (Brehmer et al., 2018; Papamakarios et al., 2019). Data were released, which have been used as benchmarks in papers from several other groups.

This work integrates two seemingly disparate types of processing -- relational and signal processing -- in a manner which is performance competitive with the best data management products and the best signal processing products. This novel integration introduces elegant abstractions that broaden the capability of data management systems with domain-specific operations. This work is the result of collaboration with Microsoft Research in Redmond and is now part of the commercial engine used in Microsoft Azure Stream Analytics.

This paper proposes a 'whole-cell' model built from first principles. It has become a landmark in systems biology and is extremely well adopted by theoretical and experimental labs around the world. It led to the organization of an international workshop (2015, Lorentz Centre, Leiden) plus at least 5 keynote talks in prestigious conferences by Weisse and Oyarzun. The model, widely known now as the 'Weisse model', unifies several approaches from the literature into a single predictive model for how bacteria growth and replicate. It explains various datasets and provides intriguing predictions for synthetic biology.

This simple and effective adaptation technique (LHUC) for neural network acoustic models offers consistent increases in speech recognition accuracy, and is widely used. An initial version presented at the IEEE SLT-2014 workshop was awarded best paper and is very well-cited. LHUC is a general method that has been successfully applied in quite different areas such as domain-adaptive machine translation (Amazon - Vilar, NAACL-HLT, 2018), and has provided the basis for further models such as Subspace LHUC (Samarakoon, Interspeech-2016) and Bayesian LHUC (Xie et al, ICASSP-2019). This work led to a project funded by Samsung about adapting end-to-end neural speech recognition.

The culmination of a series of papers where the voice source is modelled using a parametric model, enabling increased controllability of text-to-speech synthesis. This paper significantly extends the previous work in terms of the detailed formulation of the approach, and a novel application to voice transformation. This work has been widely recognised as the major approach to applying parametric models of glottal flow to speech synthesis, for instance by Raitio et al (Speech Communication, 2016, "Phase perception of the glottal excitation") and Abrol et al (Speech Communication, 2016, "Greedy double sparse dictionary learning for sparse representation of speech signals").

This paper introduced a novel machine learning approach to extract features from DNA methylation data and showed that this approach could help elucidate the relationship between DNA methylation and gene expression. Its originality led to the award of the Best Paper Award at the premier computational biology conference ECCB (European Conference on Computational Biology) in 2016. It also led to a new collaboration with researchers at Cambridge and EBI, where it was one of the main computational tools used in a Nature Communications paper. The associated software BPRMeth was downloaded over 2000 times from over 1000 IP addresses.

This paper introduces SLAMBench, the first publicly-available software framework that enables quantitative and reproducible experimental research to investigate trade-offs in performance, accuracy and energy consumption of dense RGB-D SLAM systems. SLAMBench provides a KinectFusion implementation in C++, OpenMP, OpenCL and CUDA, and harnesses the ICL-NUIM dataset of synthetic RGB-D sequences with trajectory and scene ground truth for reliable accuracy comparison of different implementations. We present an analysis and breakdown of the constituent elements of KinectFusion, and investigate their execution time on a variety of multicore and GPU-accelerated platforms. SLAMBench is now being used by researchers around the world.

One of the first large studies examining why people choose to not update software. The work has since been cited in a wide variety of areas including: electronic cars, augmented reality, operating systems, and right to repair; showing the breadth of impact of the topic. It was also quoted in news publications such as the Financial Times several times after the WannaCry attacks in 2017. The work was published in Computer Human Interaction which is premier conference that had 2435 submitted papers and a 23% acceptance rate in 2016.

Often in security research what users say they do and what they actually do are not the same. This study crosses that divide by asking users about their software update behavior and comparing it to extracted logs on their computer. Such studies are rare as they require a wide range of expertise to execute. The work has also seen exposure outside of academics with a CyberUK talk and being cited by the Financial Times. SOUPS had 79 papers submitted in 2014 and an acceptance rate of 27%.

This paper introduces the Merlin open-source toolkit which has become a widely-used baseline for research in deep neural network based speech synthesis. It now has a considerable user group: e.g. the repository at https://github.com/CSTR-Edinburgh/merlin has been starred over 800 times and forked over 300 times. There are presumably also many anonymous users of the toolkit. At a single recent conference, 8 of the text-to-speech papers used default Merlin voices as reference systems, including 6 from other institutions.

This output automated the process of testing compilers for errors by using a deep learned language model to generate test code. It won a distinguished paper award in ISSTA 2018, an ACM SIGSOFT conference with 23% acceptance rate. The fuzzing toolchain is currently used by Codeplay, one of the leading companies for heterogeneous development tools, to find errors in their own compilers.

First work to quantify the issues that arise when evaluating text-to-speech with not enough listeners and sentences. This study was also the first to point out that most peer reviewed papers in TTS contain poorly designed evaluation and therefore potentially unreliable results.

After this study was published there was a shift in the quality of peer reviewed publications, noticed among other things, by how often this paper is cited to justify design decisions.

The critique presented was supported by a quantitative analysis of real data obtained from the Blizzard challenge, a large scale rigorously designed evaluation of state-of-the-art TTS.

This article was published on 23 Jul, 2019