Richard Klein

Devon Jarvis, Richard Klein, Benjamin Rosman et al.

A number of machine learning models have been proposed with the goal of achieving systematic generalization: the ability to reason about new situations by combining aspects of previous experiences. These models leverage compositional architectures which aim to learn specialized modules dedicated to structures in a task that can be composed to solve novel problems with similar structures. While the compositionality of these architectures is guaranteed by design, the modules specializing is not. Here we theoretically study the ability of network modules to specialize to useful structures in a dataset and achieve systematic generalization. To this end we introduce a minimal space of datasets motivated by practical systematic generalization benchmarks. From this space of datasets we present a mathematical definition of systematicity and study the learning dynamics of linear neural modules when solving components of the task. Our results shed light on the difficulty of module specialization, what is required for modules to successfully specialize, and the necessity of modular architectures to achieve systematicity. Finally, we confirm that the theoretical results in our tractable setting generalize to more complex datasets and non-linear architectures.

Devon Jarvis, Richard Klein, Benjamin Rosman et al.

Model collapse, the degradation in performance that arises when generative models are trained on the outputs of prior models, is an increasing concern as artificially generated content proliferates. Related critiques of large language models have highlighted their tendency to reproduce frequent patterns in training data, their reliance on vast datasets, and their substantial environmental cost. Together, these factors contribute to data degradation, the reinforcement of cultural biases, and inefficient resource use. In this position paper we aim to combine these views and argue that model collapse threatens current efforts to democratize AI. By reducing training efficiency and skewing data distributions away from the tails of their support, model collapse disproportionately impacts low-resource and marginalized communities. We examine both the environmental and cultural implications of this phenomenon, situate our position within recent position papers on model collapse, and conclude with a call to action. Finally, we outline initial directions for mitigating these effects.

Michael Beukman, Devon Jarvis, Richard Klein et al.

While reinforcement learning has achieved remarkable successes in several domains, its real-world application is limited due to many methods failing to generalise to unfamiliar conditions. In this work, we consider the problem of generalising to new transition dynamics, corresponding to cases in which the environment's response to the agent's actions differs. For example, the gravitational force exerted on a robot depends on its mass and changes the robot's mobility. Consequently, in such cases, it is necessary to condition an agent's actions on extrinsic state information and pertinent contextual information reflecting how the environment responds. While the need for context-sensitive policies has been established, the manner in which context is incorporated architecturally has received less attention. Thus, in this work, we present an investigation into how context information should be incorporated into behaviour learning to improve generalisation. To this end, we introduce a neural network architecture, the Decision Adapter, which generates the weights of an adapter module and conditions the behaviour of an agent on the context information. We show that the Decision Adapter is a useful generalisation of a previously proposed architecture and empirically demonstrate that it results in superior generalisation performance compared to previous approaches in several environments. Beyond this, the Decision Adapter is more robust to irrelevant distractor variables than several alternative methods.

David Torpey, Richard Klein

It is known that representations from self-supervised pre-training can perform on par, and often better, on various downstream tasks than representations from fully-supervised pre-training. This has been shown in a host of settings such as generic object classification and detection, semantic segmentation, and image retrieval. However, some issues have recently come to the fore that demonstrate some of the failure modes of self-supervised representations, such as performance on non-ImageNet-like data, or complex scenes. In this paper, we show that self-supervised representations based on the instance discrimination objective lead to better representations of objects that are more robust to changes in the viewpoint and perspective of the object. We perform experiments of modern self-supervised methods against multiple supervised baselines to demonstrate this, including approximating object viewpoint variation through homographies, and real-world tests based on several multi-view datasets. We find that self-supervised representations are more robust to object viewpoint and appear to encode more pertinent information about objects that facilitate the recognition of objects from novel views.

Nathan Michlo, Devon Jarvis, Richard Klein et al.

In this work, we investigate the properties of data that cause popular representation learning approaches to fail. In particular, we find that in environments where states do not significantly overlap, variational autoencoders (VAEs) fail to learn useful features. We demonstrate this failure in a simple gridworld domain, and then provide a solution in the form of metric learning. However, metric learning requires supervision in the form of a distance function, which is absent in reinforcement learning. To overcome this, we leverage the sequential nature of states in a replay buffer to approximate a distance metric and provide a weak supervision signal, under the assumption that temporally close states are also semantically similar. We modify a VAE with triplet loss and demonstrate that this approach is able to learn useful features for downstream tasks, without additional supervision, in environments where standard VAEs fail.

David Torpey, Richard Klein

Often, applications of self-supervised learning to 3D medical data opt to use 3D variants of successful 2D network architectures. Although promising approaches, they are significantly more computationally demanding to train, and thus reduce the widespread applicability of these methods away from those with modest computational resources. Thus, in this paper, we aim to improve standard 2D SSL algorithms by modelling the inherent 3D nature of these datasets implicitly. We propose two variants that build upon a strong baseline model and show that both of these variants often outperform the baseline in a variety of downstream tasks. Importantly, in contrast to previous works in both 2D and 3D approaches for 3D medical data, both of our proposals introduce negligible additional overhead over the baseline, improving the democratisation of these approaches for medical applications.

Devon Jarvis, Richard Klein, Benjamin Rosman et al.

In spite of finite dimension ReLU neural networks being a consistent factor behind recent deep learning successes, a theory of feature learning in these models remains elusive. Currently, insightful theories still rely on assumptions including the linearity of the network computations, unstructured input data and architectural constraints such as infinite width or a single hidden layer. To begin to address this gap we establish an equivalence between ReLU networks and Gated Deep Linear Networks, and use their greater tractability to derive dynamics of learning. We then consider multiple variants of a core task reminiscent of multi-task learning or contextual control which requires both feature learning and nonlinearity. We make explicit that, for these tasks, the ReLU networks possess an inductive bias towards latent representations which are not strictly modular or disentangled but are still highly structured and reusable between contexts. This effect is amplified with the addition of more contexts and hidden layers. Thus, we take a step towards a theory of feature learning in finite ReLU networks and shed light on how structured mixed-selective latent representations can emerge due to a bias for node-reuse and learning speed.

David Torpey, Lawrence Pratt, Richard Klein

Pretraining has been shown to improve performance in many domains, including semantic segmentation, especially in domains with limited labelled data. In this work, we perform a large-scale evaluation and benchmarking of various pretraining methods for Solar Cell Defect Detection (SCDD) in electroluminescence images, a field with limited labelled datasets. We cover supervised training with semantic segmentation, semi-supervised learning, and two self-supervised techniques. We also experiment with both in-distribution and out-of-distribution (OOD) pretraining and observe how this affects downstream performance. The results suggest that supervised training on a large OOD dataset (COCO), self-supervised pretraining on a large OOD dataset (ImageNet), and semi-supervised pretraining (CCT) all yield statistically equivalent performance for mean Intersection over Union (mIoU). We achieve a new state-of-the-art for SCDD and demonstrate that certain pretraining schemes result in superior performance on underrepresented classes. Additionally, we provide a large-scale unlabelled EL image dataset of $22000$ images, and a $642$-image labelled semantic segmentation EL dataset, for further research in developing self- and semi-supervised training techniques in this domain.

David Torpey, Richard Klein

The standard approach to modern self-supervised learning is to generate random views through data augmentations and minimise a loss computed from the representations of these views. This inherently encourages invariance to the transformations that comprise the data augmentation function. In this work, we show that adding a module to constrain the representations to be predictive of an affine transformation improves the performance and efficiency of the learning process. The module is agnostic to the base self-supervised model and manifests in the form of an additional loss term that encourages an aggregation of the encoder representations to be predictive of an affine transformation applied to the input images. We perform experiments in various modern self-supervised models and see a performance improvement in all cases. Further, we perform an ablation study on the components of the affine transformation to understand which of them is affecting performance the most, as well as on key architectural design decisions.

John M. Statheros, Hairong Wang, Richard Klein

The interpretation of chest X-rays (CXRs) poses significant challenges, particularly in achieving accurate multi-label pathology classification and spatial localization. These tasks demand different levels of annotation granularity but are frequently constrained by the scarcity of region-level (dense) annotations. We introduce CLARiTy (Class Localizing and Attention Refining Image Transformer), a vision transformer-based model for joint multi-label classification and weakly-supervised localization of thoracic pathologies. CLARiTy employs multiple class-specific tokens to generate discriminative attention maps, and a SegmentCAM module for foreground segmentation and background suppression using explicit anatomical priors. Trained on image-level labels from the NIH ChestX-ray14 dataset, it leverages distillation from a ConvNeXtV2 teacher for efficiency. Evaluated on the official NIH split, the CLARiTy-S-16-512 (a configuration of CLARiTy), achieves competitive classification performance across 14 pathologies, and state-of-the-art weakly-supervised localization performance on 8 pathologies, outperforming prior methods by 50.7%. In particular, pronounced gains occur for small pathologies like nodules and masses. The lower-resolution variant of CLARiTy, CLARiTy-S-16-224, offers high efficiency while decisively surpassing baselines, thereby having the potential for use in low-resource settings. An ablation study confirms contributions of SegmentCAM, DINO pretraining, orthogonal class token loss, and attention pooling. CLARiTy advances beyond CNN-ViT hybrids by harnessing ViT self-attention for global context and class-specific localization, refined through convolutional background suppression for precise, noise-reduced heatmaps.

Devon Jarvis, Verena Klar, Richard Klein et al.

Patients with semantic dementia (SD) present with remarkably consistent atrophy of neurons in the anterior temporal lobe and behavioural impairments, such as graded loss of category knowledge. While relearning of lost knowledge has been shown in acute brain injuries such as stroke, it has not been widely supported in chronic cognitive diseases such as SD. Previous research has shown that deep linear artificial neural networks exhibit stages of semantic learning akin to humans. Here, we use a deep linear network to test the hypothesis that relearning during disease progression rather than particular atrophy cause the specific behavioural patterns associated with SD. After training the network to generate the common semantic features of various hierarchically organised objects, neurons are successively deleted to mimic atrophy while retraining the model. The model with relearning and deleted neurons reproduced errors specific to SD, including prototyping errors and cross-category confusions. This suggests that relearning is necessary for artificial neural networks to reproduce the behavioural patterns associated with SD in the absence of \textit{output} non-linearities. Our results support a theory of SD progression that results from continuous relearning of lost information. Future research should revisit the role of relearning as a contributing factor to cognitive diseases.

Nathan Michlo, Richard Klein, Steven James

Learning disentangled representations with variational autoencoders (VAEs) is often attributed to the regularisation component of the loss. In this work, we highlight the interaction between data and the reconstruction term of the loss as the main contributor to disentanglement in VAEs. We show that standard benchmark datasets have unintended correlations between their subjective ground-truth factors and perceived axes in the data according to typical VAE reconstruction losses. Our work exploits this relationship to provide a theory for what constitutes an adversarial dataset under a given reconstruction loss. We verify this by constructing an example dataset that prevents disentanglement in state-of-the-art frameworks while maintaining human-intuitive ground-truth factors. Finally, we re-enable disentanglement by designing an example reconstruction loss that is once again able to perceive the ground-truth factors. Our findings demonstrate the subjective nature of disentanglement and the importance of considering the interaction between the ground-truth factors, data and notably, the reconstruction loss, which is under-recognised in the literature.

David Poulton, Richard Klein

This research presents the idea of activity fusion into existing Pose Estimation architectures to enhance their predictive ability. This is motivated by the rise in higher level concepts found in modern machine learning architectures, and the belief that activity context is a useful piece of information for the problem of pose estimation. To analyse this concept we take an existing deep learning architecture and augment it with an additional 1x1 convolution to fuse activity information into the model. We perform evaluation and comparison on a common pose estimation dataset, and show a performance improvement over our baseline model, especially in uncommon poses and on typically difficult joints. Additionally, we perform an ablative analysis to indicate that the performance improvement does in fact draw from the activity information.

Jared Harris-Dewey, Richard Klein

We give an overview of the different rendering methods and we demonstrate that the use of a Generative Adversarial Networks (GAN) for Global Illumination (GI) gives a superior quality rendered image to that of a rasterisations image. We utilise the Pix2Pix architecture and specify the hyper-parameters and methodology used to mimic ray-traced images from a set of input features. We also demonstrate that the GANs quality is comparable to the quality of the ray-traced images, but is able to produce the image, at a fraction of the time.

Julien Nyambal, Richard Klein

Finding a parking space nowadays becomes an issue that is not to be neglected, it consumes time and energy. We have used computer vision techniques to infer the state of the parking lot given the data collected from the University of The Witwatersrand. This paper presents an approach for a real-time parking space classification based on Convolutional Neural Networks (CNN) using Caffe and Nvidia DiGITS framework. The training process has been done using DiGITS and the output is a caffemodel used for predictions to detect vacant and occupied parking spots. The system checks a defined area whether a parking spot (bounding boxes defined at initialization of the system) is containing a car or not (occupied or vacant). Those bounding box coordinates are saved from a frame of the video of the parking lot in a JSON format, to be later used by the system for sequential prediction on each parking spot. The system has been trained using the LeNet network with the Nesterov Accelerated Gradient as solver and the AlexNet network with the Stochastic Gradient Descent as solver. We were able to get an accuracy on the validation set of 99\% for both networks. The accuracy on a foreign dataset(PKLot) returned as well 99\%. Those are experimental results based on the training set shows how robust the system can be when the prediction has to take place in a different parking space.

Pierce Burke, Richard Klein

With the increased interest in machine learning and big data problems, the need for large amounts of labelled data has also grown. However, it is often infeasible to get experts to label all of this data, which leads many practitioners to crowdsourcing solutions. In this paper, we present new techniques to improve the quality of the labels while attempting to reduce the cost. The naive approach to assigning labels is to adopt a majority vote method, however, in the context of data labelling, this is not always ideal as data labellers are not equally reliable. One might, instead, give higher priority to certain labellers through some kind of weighted vote based on past performance. This paper investigates the use of more sophisticated methods, such as Bayesian inference, to measure the performance of the labellers as well as the confidence of each label. The methods we propose follow an iterative improvement algorithm which attempts to use the least amount of workers necessary to achieve the desired confidence in the inferred label. This paper explores simulated binary classification problems with simulated workers and questions to test the proposed methods. Our methods outperform the standard voting methods in both cost and accuracy while maintaining higher reliability when there is disagreement within the crowd.

Richard Klein, Turgay Celik

To perform contingent teaching and be responsive to students' needs during class, lecturers must be able to quickly assess the state of their audience. While effective teachers are able to gauge easily the affective state of the students, as class sizes grow this becomes increasingly difficult and less precise. The Wits Intelligent Teaching System (WITS) aims to assist lecturers with real-time feedback regarding student affect. The focus is primarily on recognising engagement or lack thereof. Student engagement is labelled based on behaviour and postures that are common to classroom settings. These proxies are then used in an observational checklist to construct a dataset of engagement upon which a CNN based on AlexNet is successfully trained and which significantly outperforms a Support Vector Machine approach. The deep learning approach provides satisfactory results on a challenging, real-world dataset with significant occlusion, lighting and resolution constraints.

David Torpey, Richard Klein

The typical contrastive self-supervised algorithm uses a similarity measure in latent space as the supervision signal by contrasting positive and negative images directly or indirectly. Although the utility of self-supervised algorithms has improved recently, there are still bottlenecks hindering their widespread use, such as the compute needed. In this paper, we propose a module that serves as an additional objective in the self-supervised contrastive learning paradigm. We show how the inclusion of this module to regress the parameters of an affine transformation or homography, in addition to the original contrastive objective, improves both performance and learning speed. Importantly, we ensure that this module does not enforce invariance to the various components of the affine transform, as this is not always ideal. We demonstrate the effectiveness of the additional objective on two recent, popular self-supervised algorithms. We perform an extensive experimental analysis of the proposed method and show an improvement in performance for all considered datasets. Further, we find that although both the general homography and affine transformation are sufficient to improve performance and convergence, the affine transformation performs better in all cases.

Kimessha Paupamah, Steven James, Richard Klein

Deep neural networks are typically too computationally expensive to run in real-time on consumer-grade hardware and low-powered devices. In this paper, we investigate reducing the computational and memory requirements of neural networks through network pruning and quantisation. We examine their efficacy on large networks like AlexNet compared to recent compact architectures: ShuffleNet and MobileNet. Our results show that pruning and quantisation compresses these networks to less than half their original size and improves their efficiency, particularly on MobileNet with a 7x speedup. We also demonstrate that pruning, in addition to reducing the number of parameters in a network, can aid in the correction of overfitting.