skip to content

CambPlants

A networking organisation for plants-related research and impact
 

Mon 04 Jul 17:00: The Problem of Size Generalization in Graph Neural Networks

http://talks.cam.ac.uk/show/rss/49845 - Sun, 03/07/2022 - 00:21
The Problem of Size Generalization in Graph Neural Networks

In the past few years, graph neural networks (GNNs) have become the de facto model of choice for graph classification and other tasks on graph structured data. While, from the theoretical viewpoint, most GNNs can operate on graphs of any size, it is empirically observed that their classification performance degrades when they are applied on graphs with sizes that differ from those in the training data. In this talk we will give an overview of the current approaches to tackle the issue of poor size-generalization in GNNs, and we will introduce our recent work in this area.

Add to your calendar or Include in your list

Mon 11 Jul 14:00: Combining multi-omics and biological knowledge to extract disease mechanisms Please contact Ciara for further details

http://talks.cam.ac.uk/show/rss/49845 - Fri, 01/07/2022 - 17:06
Combining multi-omics and biological knowledge to extract disease mechanisms

Multi-omics technologies, and in particular those with single-cell and spatial resolution, provide unique opportunities to study deregulation of intra- and inter-cellular processes in cancer and other diseases. In this talk I will present recent methods and applications from our group towards this aim, with a focus is on computational approaches that combine data with biological knowledge within statistical and machine learning methods. This combination allows us to increase both the statistical power of our approaches and the mechanistic interpretability of the results. I will also discuss the value to perform perturbation studies, combined with mathematical modeling, to increase our understanding and therapeutic opportunities. Finally, I will show how, using novel microfluidics-based technologies, this approach can also be applied directly to biopsies, allowing to build mechanistic models for individual cancer patients, and use these models to propose new therapies.

Please contact Ciara for further details

Add to your calendar or Include in your list

Thu 07 Jul 16:00: Mycobacterium tuberculosis genome economization for pathogenesis: “More from less for more” Room changed - will be Seminar Rooms 2/3, Pathology Block, Department of Veterinary Medicine.

http://talks.cam.ac.uk/show/rss/29729 - Fri, 01/07/2022 - 10:46
Mycobacterium tuberculosis genome economization for pathogenesis: “More from less for more”

Abstract: Mycobacterium tuberculosis (M.tb) is the deadliest bacterial pathogen known to humanity causing the disease TB, taking the largest toll of human lives globally with a person dying every 15-20 seconds despite the fact that TB is completely curable if diagnosed timely and treated properly. This problem is further compounded by the development of drug resistance, to the extent of total drug resistance, HIV AIDS co-infection and the accompanying TB-IRIS and the impending impact of the emerging diabetes epidemic and of late the COVID -19 pandemic. M.tb has undergone reductive evolution, over millions of years, into a very slim and trim genomic and functional architecture. Not only has it shed much of its genome, but has balanced this genome deficit by resorting to very intelligent survival strategies such as gene co option, moon lighting and molecular mimicry.

Room changed - will be Seminar Rooms 2/3, Pathology Block, Department of Veterinary Medicine.

Add to your calendar or Include in your list

Fri 08 Jul 11:00: Synthetics with Digital Humans

http://talks.cam.ac.uk/show/rss/49845 - Thu, 30/06/2022 - 15:36
Synthetics with Digital Humans

Abstract

Nowadays, collecting the right dataset for machine learning is often more challenging than choosing the model. We address this with photorealistic synthetic training data – labelled images of humans made using computer graphics. With synthetics we can generate clean labels without annotation noise or error, produce labels otherwise impossible to annotate by hand, and easily control variation and diversity in our datasets. I will show you how synthetics underpins our work on understanding humans, including how it enables fast and accurate 3D face reconstruction, in the wild.

Bio

Dr. Erroll Wood is a Staff Software Engineer at Google, working on Digital Humans. Previously, he was a member of Microsoft’s Mixed Reality AI Lab, where he worked on hand tracking for HoloLens 2, avatars for Microsoft Mesh, synthetic data for face tracking, and Holoportation. He did his PhD at the University of Cambridge, working on gaze estimation.

Add to your calendar or Include in your list

Wed 06 Jul 11:00: The unreasonable effectiveness of mathematics in large scale deep learning

http://talks.cam.ac.uk/show/rss/49845 - Wed, 29/06/2022 - 13:01
The unreasonable effectiveness of mathematics in large scale deep learning

Recently, the theory of infinite-width neural networks led to the first technology, muTransfer, for tuning enormous neural networks that are too expensive to train more than once. For example, this allowed us to tune the 6.7 billion parameter version of GPT -3 using only 7% of its pretraining compute budget, and with some asterisks, we get a performance comparable to the original GPT -3 model with twice the parameter count. In this talk, I will explain the core insight behind this theory. In fact, this is an instance of what I call the Optimal Scaling Thesis, which connects infinite-size limits for general notions of “size” to the optimal design of large models in practice, illustrating a way for theory to reliably guide the future of AI. I’ll end with several concrete key mathematical research questions whose resolutions will have incredible impact on how practitioners scale up their NNs.

There’s no required reading for the talk but folks can look at my homepage for an overview of Tensor Programs.

Add to your calendar or Include in your list

Understanding the development, evolution, and function of bullseye pigmentation patterns in Hibiscus trionum

Plant Sciences at DSpace - Wed, 29/06/2022 - 10:22
Understanding the development, evolution, and function of bullseye pigmentation patterns in Hibiscus trionum Fairnie, Alice Colourful spot, stripe and ring patterns decorate the corolla of many flowering plants and fulfil important biotic and abiotic functions. These petal patterns are created by spatial differences in pigmentation, cell shape and texture of the adaxial petal epidermis. The mechanisms controlling formation and evolution of these patterns, and their exact role in plant-animal communication are not well understood. My PhD adds to current understanding of petal patterns by investigating the development, evolution, and function of the bullseye pigmentation pattern on Hibiscus trionum petals. The bi-coloured bullseye pattern of H.trionum is created from contrast in cell shape, cuticle texture, and pigmentation between the bottom and top of the petal. The bottom petal appears shiny and purple because cells in this region are elongated with a striated cuticle, and are pigmented with anthocyanins. A minimal regulatory network restricting anthocyanin pigment production to the bottom of Hibiscus trionum petal was identified. The network relies on two MYB regulators: HtCREAM1 which represses anthocyanin production in the top of the petal; HtBERRY1 which promotes anthocyanin biosynthesis in the bottom of the petal. This minimal network is a starting point to understand the molecular mechanisms both creating, and creating diversity, in petal patterns. Natural variation in the bullseye pattern in close relatives of H.trionum from Australia and New Zealand is in part due to four independent restrictions of anthocyanin pigmentation in the petal bottom. Preliminary results suggest restriction of pigmentation in H.richardsonii, sister-species to H.trionum, results from mutation in the regulatory region and the coding sequence of the BERRY1 homolog of H. richardsonii. Natural variation in the bullseye pigmentation pattern could reflect a function in plant-pollinator communication. Buff-tailed bumblebees were found to discriminate between, and prefer, artificial flowers with H.trionum-like bullseye patterns to H.richardsonii-like bullseye patterns.

Thu 30 Jun 14:00: A Comparative Study on the Loss Functions for Image Enhancement Networks

http://talks.cam.ac.uk/show/rss/49845 - Tue, 28/06/2022 - 21:52
A Comparative Study on the Loss Functions for Image Enhancement Networks

Image enhancement and image retouching processes are often dominated by global (shift-invariant) change of colour and tones. Most “deep learning” based methods proposed for image enhancement are trained to enforce similarity in pixel values and/or in the high-level feature space. We hypothesise that for tasks, such as image enhancement and retouching, which involve a significant shift in colour statistics, training the model to restore the overall colour distribution can be of vital importance. To address this, we study the effect of a Histogram Matching loss function on a state-of-the art colour enhancement network – HDR Net. The loss enforces similarity of the RGB histograms of the predicted and the target images. By providing detailed qualitative and quantitative comparison of different loss functions on varied datasets, we conclude that enforcing similarity in the colour distribution achieves substantial improvement in performance and can play a significant role while choosing loss functions for image enhancement networks.

Add to your calendar or Include in your list

Tue 28 Jun 11:00: Noise-Aware Differentially Private Synthetic Data

http://talks.cam.ac.uk/show/rss/49845 - Tue, 28/06/2022 - 09:32
Noise-Aware Differentially Private Synthetic Data

Synthetic data generated under differential privacy (DP) promises to significantly simplify analysis of sensitive personal data. Existing work has shown that simply analysing DP synthetic data as if it were real does not produce valid inferences of population-level quantities, leading to too narrow confidence intervals and thereby risking false discoveries. We propose using multiple imputation techniques to avoid these problems. This requires simulating multiple synthetic data sets from the Bayesian posterior predictive distribution over data sets. We propose a novel noise-aware Bayesian DP synthetic data generation mechanism for discrete data that enables generating such a distribution of data sets. Our experiments demonstrate that the method is able to produce accurate confidence intervals from DP synthetic data.

Add to your calendar or Include in your list

Wed 29 Jun 11:00: The role of meta-learning for few-shot classification

http://talks.cam.ac.uk/show/rss/49845 - Tue, 28/06/2022 - 08:59
The role of meta-learning for few-shot classification

While deep learning has driven impressive progress, one of the toughest remaining challenges is generalization beyond the training distribution. Few-shot learning is an area of research that aims to address this, by striving to build models that can learn new concepts rapidly in a more “human-like” way. While many influential few-shot learning methods were based on meta-learning, recently progress has been made by simpler transfer learning algorithms, and it has been suggested in fact that few-shot learning might be an emergent property of large-scale models. In this talk, I will give an overview of the evolution of few-shot learning methods and benchmarks, with an emphasis on the role of meta-learning on few-shot classification. I will discuss lessons learned from using larger and more diverse benchmarks for evaluation and trade-offs between different approaches, closing with an open discussion about remaining challenges.

Add to your calendar or Include in your list

Thu 30 Jun 15:00: Trustworthy Digital Identity - Systems Architecture

http://talks.cam.ac.uk/show/rss/49845 - Mon, 27/06/2022 - 09:25
Trustworthy Digital Identity - Systems Architecture

Governments around the world are committed to supporting the roll out of national digital IDs, but there are privacy and security implications associated with scaling these systems at a national level.

Responsible implementation of ID services is a critical enabler for financial inclusion; it enables access to services and enactment of civil rights. According to the World Bank, more than 1 billion people are currently living without an official digital identity.

Questions of trust are based around the complex interplay of socio-technical considerations, requiring multi-disciplinary expertise. The ‘trustworthiness’ of digital IDs is characterised by multiple inter-related dimensions that include security, privacy, ethics, resilience, robustness and reliability. These dimensions are required to provide the knowledge, tools and guidance needed to implement privacy-preserving, secure identification systems

The project aims to enhance the privacy and security of national digital identity systems, with the ultimate goal to maximise the value to beneficiaries, whilst limiting known and unknown risks to these constituents and maintaining the integrity of the overall system.

Add to your calendar or Include in your list

Wed 29 Jun 16:00: Critical weaknesses in shielding strategies for COVID-19

http://talks.cam.ac.uk/show/rss/29729 - Fri, 24/06/2022 - 09:35
Critical weaknesses in shielding strategies for COVID-19

The COVID -19 pandemic, caused by the coronavirus SARS -CoV-2, has led to a wide range of non-pharmaceutical interventions being implemented around the world to curb transmission. However, the economic and social costs of some of these measures, especially lockdowns, has been high. An alternative and widely discussed public health strategy for the COVID -19 pandemic would have been to `shield’ those most vulnerable to COVID -19 (minimising their contacts with others), while allowing infection to spread among lower risk individuals with the aim of reaching herd immunity. In this talk we will retrospectively explore the effectiveness of such a strategy using a stochastic SEIR framework, showing that even under the unrealistic assumption of perfect shielding, hospitals would have been rapidly overwhelmed with many deaths among lower risk individuals. Crucially, even a small (20%) reduction in the effectiveness of shielding would have likely led to a large increase (>150%) in the number of deaths compared to perfect shielding. Our findings demonstrate that shielding the vulnerable while allowing infections to spread among the wider population would not have been a viable public health strategy for COVID -19 and is unlikely to be effective for future pandemics.

Add to your calendar or Include in your list