Adam Prügel-Bennett

Ethan Harris, Antonia Marcu, Matthew Painter et al.

Mixed Sample Data Augmentation (MSDA) has received increasing attention in recent years, with many successful variants such as MixUp and CutMix. By studying the mutual information between the function learned by a VAE on the original data and on the augmented data we show that MixUp distorts learned functions in a way that CutMix does not. We further demonstrate this by showing that MixUp acts as a form of adversarial training, increasing robustness to attacks such as Deep Fool and Uniform Noise which produce examples similar to those generated by MixUp. We argue that this distortion prevents models from learning about sample specific features in the data, aiding generalisation performance. In contrast, we suggest that CutMix works more like a traditional augmentation, improving performance by preventing memorisation without distorting the data distribution. However, we argue that an MSDA which builds on CutMix to include masks of arbitrary shape, rather than just square, could further prevent memorisation whilst preserving the data distribution in the same way. To this end, we propose FMix, an MSDA that uses random binary masks obtained by applying a threshold to low frequency images sampled from Fourier space. These random masks can take on a wide range of shapes and can be generated for use with one, two, and three dimensional data. FMix improves performance over MixUp and CutMix, without an increase in training time, for a number of models across a range of data sets and problem settings, obtaining a new single model state-of-the-art result on CIFAR-10 without external data. Finally, we show that a consequence of the difference between interpolating MSDA such as MixUp and masking MSDA such as FMix is that the two can be combined to improve performance even further. Code for all experiments is provided at https://github.com/ecs-vlc/FMix .

Jay Bear, Adam Prügel-Bennett, Jonathon Hare

Iterative algorithms solve problems by taking steps until a solution is reached. Models in the form of Deep Thinking (DT) networks have been demonstrated to learn iterative algorithms in a way that can scale to different sized problems at inference time using recurrent computation and convolutions. However, they are often unstable during training, and have no guarantees of convergence/termination at the solution. This paper addresses the problem of instability by analyzing the growth in intermediate representations, allowing us to build models (referred to as Deep Thinking with Lipschitz Constraints (DT-L)) with many fewer parameters and providing more reliable solutions. Additionally our DT-L formulation provides guarantees of convergence of the learned iterative procedure to a unique solution at inference time. We demonstrate DT-L is capable of robustly learning algorithms which extrapolate to harder problems than in the training set. We benchmark on the traveling salesperson problem to evaluate the capabilities of the modified system in an NP-hard problem where DT fails to learn.

Halil Ibrahim Aysel, Xiaohao Cai, Adam Prügel-Bennett

Semantic segmentation is a critical task in computer vision aiming to identify and classify individual pixels in an image, with numerous applications in for example autonomous driving and medical image analysis. However, semantic segmentation can be highly challenging particularly due to the need for large amounts of annotated data. Annotating images is a time-consuming and costly process, often requiring expert knowledge and significant effort; moreover, saving the annotated images could dramatically increase the storage space. In this paper, we propose a novel approach for semantic segmentation, requiring the rough information of individual semantic class proportions, shortened as semantic proportions, rather than the necessity of ground-truth segmentation maps. This greatly simplifies the data annotation process and thus will significantly reduce the annotation time, cost and storage space, opening up new possibilities for semantic segmentation tasks where obtaining the full ground-truth segmentation maps may not be feasible or practical. Our proposed method of utilising semantic proportions can (i) further be utilised as a booster in the presence of ground-truth segmentation maps to gain performance without extra data and model complexity, and (ii) also be seen as a parameter-free plug-and-play module, which can be attached to existing deep neural networks designed for semantic segmentation. Extensive experimental results demonstrate the good performance of our method compared to benchmark methods that rely on ground-truth segmentation maps. Utilising semantic proportions suggested in this work offers a promising direction for future semantic segmentation research.

Dominic Belcher, Antonia Marcu, Adam Prügel-Bennett

In this paper we show that the expected generalisation performance of a learning machine is determined by the distribution of risks or equivalently its logarithm -- a quantity we term the risk entropy -- and the fluctuations in a quantity we call the training ratio. We show that the risk entropy can be empirically inferred for deep neural network models using Markov Chain Monte Carlo techniques. Results are presented for different deep neural networks on a variety of problems. The asymptotic behaviour of the risk entropy acts in an analogous way to the capacity of the learning machine, but the generalisation performance experienced in practical situations is determined by the behaviour of the risk entropy before the asymptotic regime is reached. This performance is strongly dependent on the distribution of the data (features and targets) and not just on the capacity of the learning machine.

Mark Tuddenham, Adam Prügel-Bennett, Jonathan Hare

The optimisation of neural networks can be sped up by orthogonalising the gradients before the optimisation step, ensuring the diversification of the learned representations. We orthogonalise the gradients of the layer's components/filters with respect to each other to separate out the intermediate representations. Our method of orthogonalisation allows the weights to be used more flexibly, in contrast to restricting the weights to an orthogonalised sub-space. We tested this method on ImageNet and CIFAR-10 resulting in a large decrease in learning time, and also obtain a speed-up on the semi-supervised learning BarlowTwins. We obtain similar accuracy to SGD without fine-tuning and better accuracy for naïvely chosen hyper-parameters.

Antonia Marcu, Adam Prügel-Bennett

The community lacks theory-informed guidelines for building good data sets. We analyse theoretical directions relating to what aspects of the data matter and conclude that the intuitions derived from the existing literature are incorrect and misleading. Using empirical counter-examples, we show that 1) data dimension should not necessarily be minimised and 2) when manipulating data, preserving the distribution is inessential. This calls for a more data-aware theoretical understanding. Although not explored in this work, we propose the study of the impact of data modification on learned representations as a promising research direction.

Antonia Marcu, Adam Prügel-Bennett

Data distortion is commonly applied in vision models during both training (e.g methods like MixUp and CutMix) and evaluation (e.g. shape-texture bias and robustness). This data modification can introduce artificial information. It is often assumed that the resulting artefacts are detrimental to training, whilst being negligible when analysing models. We investigate these assumptions and conclude that in some cases they are unfounded and lead to incorrect results. Specifically, we show current shape bias identification methods and occlusion robustness measures are biased and propose a fairer alternative for the latter. Subsequently, through a series of experiments we seek to correct and strengthen the community's perception of how augmenting affects learning of vision models. Based on our empirical results we argue that the impact of the artefacts must be understood and exploited rather than eliminated.

Takaki Yamada, Adam Prügel-Bennett, Stefan B. Williams et al.

This paper describes Georeference Contrastive Learning of visual Representation (GeoCLR) for efficient training of deep-learning Convolutional Neural Networks (CNNs). The method leverages georeference information by generating a similar image pair using images taken of nearby locations, and contrasting these with an image pair that is far apart. The underlying assumption is that images gathered within a close distance are more likely to have similar visual appearance, where this can be reasonably satisfied in seafloor robotic imaging applications where image footprints are limited to edge lengths of a few metres and are taken so that they overlap along a vehicle's trajectory, whereas seafloor substrates and habitats have patch sizes that are far larger. A key advantage of this method is that it is self-supervised and does not require any human input for CNN training. The method is computationally efficient, where results can be generated between dives during multi-day AUV missions using computational resources that would be accessible during most oceanic field trials. We apply GeoCLR to habitat classification on a dataset that consists of ~86k images gathered using an Autonomous Underwater Vehicle (AUV). We demonstrate how the latent representations generated by GeoCLR can be used to efficiently guide human annotation efforts, where the semi-supervised framework improves classification accuracy by an average of 10.2% compared to the state-of-the-art SimCLR using the same CNN and equivalent number of human annotations for training.

Yue Jiao, Jonathon Hare, Adam Prügel-Bennett

Visual Semantic Embedding (VSE) models, which map images into a rich semantic embedding space, have been a milestone in object recognition and zero-shot learning. Current approaches to VSE heavily rely on static word em-bedding techniques. In this work, we propose a Visual Se-mantic Embedding Probe (VSEP) designed to probe the semantic information of contextualized word embeddings in visual semantic understanding tasks. We show that the knowledge encoded in transformer language models can be exploited for tasks requiring visual semantic understanding.The VSEP with contextual representations can distinguish word-level object representations in complicated scenes as a compositional zero-shot learner. We further introduce a zero-shot setting with VSEPs to evaluate a model's ability to associate a novel word with a novel visual category. We find that contextual representations in language mod-els outperform static word embeddings, when the compositional chain of object is short. We notice that current visual semantic embedding models lack a mutual exclusivity bias which limits their performance.

Yue Jiao, Jonathon Hare, Adam Prügel-Bennett

Zero shot learning (ZSL) has seen a surge in interest over the decade for its tight links with the mechanism making young children recognize novel objects. Although different paradigms of visual semantic embedding models are designed to align visual features and distributed word representations, it is unclear to what extent current ZSL models encode semantic information from distributed word representations. In this work, we introduce the split of tiered-ImageNet to the ZSL task, in order to avoid the structural flaws in the standard ImageNet benchmark. We build a unified framework for ZSL with contrastive learning as pre-training, which guarantees no semantic information leakage and encourages linearly separable visual features. Our work makes it fair for evaluating visual semantic embedding models on a ZSL setting in which semantic inference is decisive. With this framework, we show that current ZSL models struggle with encoding semantic relationships from word analogy and word hierarchy. Our analyses provide motivation for exploring the role of context language representations in ZSL tasks.

Vlad Velici, Adam Prügel-Bennett

This report presents the application of object detection on a database of underwater images of different species of crabs, as well as aerial images of sea lions and finally the Pascal VOC dataset. The model is an end-to-end object detection neural network based on a convolutional network base and a Long Short-Term Memory detector.

Vlad Velici, Adam Prügel-Bennett

Long Short-Term Memory (LSTM) units have the ability to memorise and use long-term dependencies between inputs to generate predictions on time series data. We introduce the concept of modifying the cell state (memory) of LSTMs using rotation matrices parametrised by a new set of trainable weights. This addition shows significant increases of performance on some of the tasks from the bAbI dataset.

Mark Tuddenham, Adam Prügel-Bennett, Jonathan Hare

Classification problems using deep learning have been shown to have a high-curvature subspace in the loss landscape equal in dimension to the number of classes. Moreover, this subspace corresponds to the subspace spanned by the logit gradients for each class. An obvious strategy to speed up optimisation would be to use Newton's method in the high-curvature subspace and stochastic gradient descent in the co-space. We show that a naive implementation actually slows down convergence and we speculate why this might be.

Antonia Marcu, Adam Prügel-Bennett

In this paper, a new approach to computing the generalisation performance is presented that assumes the distribution of risks, $ρ(r)$, for a learning scenario is known. From this, the expected error of a learning machine using empirical risk minimisation is computed for both classification and regression problems. A critical quantity in determining the generalisation performance is the power-law behaviour of $ρ(r)$ around its minimum value---a quantity we call attunement. The distribution $ρ(r)$ is computed for the case of all Boolean functions and for the perceptron used in two different problem settings. Initially a simplified analysis is presented where an independence assumption about the losses is made. A more accurate analysis is carried out taking into account chance correlations in the training set. This leads to corrections in the typical behaviour that is observed.

Yan Zhang, Jonathon Hare, Adam Prügel-Bennett

Current approaches for predicting sets from feature vectors ignore the unordered nature of sets and suffer from discontinuity issues as a result. We propose a general model for predicting sets that properly respects the structure of sets and avoids this problem. With a single feature vector as input, we show that our model is able to auto-encode point sets, predict the set of bounding boxes of objects in an image, and predict the set of attributes of these objects.

Yan Zhang, Jonathon Hare, Adam Prügel-Bennett

Traditional set prediction models can struggle with simple datasets due to an issue we call the responsibility problem. We introduce a pooling method for sets of feature vectors based on sorting features across elements of the set. This can be used to construct a permutation-equivariant auto-encoder that avoids this responsibility problem. On a toy dataset of polygons and a set version of MNIST, we show that such an auto-encoder produces considerably better reconstructions and representations. Replacing the pooling function in existing set encoders with FSPool improves accuracy and convergence speed on a variety of datasets.

Yan Zhang, Jonathon Hare, Adam Prügel-Bennett

Representations of sets are challenging to learn because operations on sets should be permutation-invariant. To this end, we propose a Permutation-Optimisation module that learns how to permute a set end-to-end. The permuted set can be further processed to learn a permutation-invariant representation of that set, avoiding a bottleneck in traditional set models. We demonstrate our model's ability to learn permutations and set representations with either explicit or implicit supervision on four datasets, on which we achieve state-of-the-art results: number sorting, image mosaics, classification from image mosaics, and visual question answering.

Yan Zhang, Jonathon Hare, Adam Prügel-Bennett

Visual Question Answering (VQA) models have struggled with counting objects in natural images so far. We identify a fundamental problem due to soft attention in these models as a cause. To circumvent this problem, we propose a neural network component that allows robust counting from object proposals. Experiments on a toy task show the effectiveness of this component and we obtain state-of-the-art accuracy on the number category of the VQA v2 dataset without negatively affecting other categories, even outperforming ensemble models with our single model. On a difficult balanced pair metric, the component gives a substantial improvement in counting over a strong baseline by 6.6%.