Amos Storkey

Chenhongyi Yang, Mateusz Ochal, Amos Storkey et al.

Real-world object detection models should be cheap and accurate. Knowledge distillation (KD) can boost the accuracy of a small, cheap detection model by leveraging useful information from a larger teacher model. However, a key challenge is identifying the most informative features produced by the teacher for distillation. In this work, we show that only a very small fraction of features within a ground-truth bounding box are responsible for a teacher's high detection performance. Based on this, we propose Prediction-Guided Distillation (PGD), which focuses distillation on these key predictive regions of the teacher and yields considerable gains in performance over many existing KD baselines. In addition, we propose an adaptive weighting scheme over the key regions to smooth out their influence and achieve even better performance. Our proposed approach outperforms current state-of-the-art KD baselines on a variety of advanced one-stage detection architectures. Specifically, on the COCO dataset, our method achieves between +3.1% and +4.6% AP improvement using ResNet-101 and ResNet-50 as the teacher and student backbones, respectively. On the CrowdHuman dataset, we achieve +3.2% and +2.0% improvements in MR and AP, also using these backbones. Our code is available at https://github.com/ChenhongyiYang/PGD.

Jamie Burke, Justin Engelmann, Charlene Hamid et al.

Purpose: To develop an open-source, fully-automatic deep learning algorithm, DeepGPET, for choroid region segmentation in optical coherence tomography (OCT) data. Methods: We used a dataset of 715 OCT B-scans (82 subjects, 115 eyes) from 3 clinical studies related to systemic disease. Ground truth segmentations were generated using a clinically validated, semi-automatic choroid segmentation method, Gaussian Process Edge Tracing (GPET). We finetuned a UNet with MobileNetV3 backbone pre-trained on ImageNet. Standard segmentation agreement metrics, as well as derived measures of choroidal thickness and area, were used to evaluate DeepGPET, alongside qualitative evaluation from a clinical ophthalmologist. Results: DeepGPET achieves excellent agreement with GPET on data from 3 clinical studies (AUC=0.9994, Dice=0.9664; Pearson correlation of 0.8908 for choroidal thickness and 0.9082 for choroidal area), while reducing the mean processing time per image on a standard laptop CPU from 34.49s ($\pm$15.09) using GPET to 1.25s ($\pm$0.10) using DeepGPET. Both methods performed similarly according to a clinical ophthalmologist, who qualitatively judged a subset of segmentations by GPET and DeepGPET, based on smoothness and accuracy of segmentations. Conclusions: DeepGPET, a fully-automatic, open-source algorithm for choroidal segmentation, will enable researchers to efficiently extract choroidal measurements, even for large datasets. As no manual interventions are required, DeepGPET is less subjective than semi-automatic methods and could be deployed in clinical practice without necessitating a trained operator.

Lukas Schäfer, Filippos Christianos, Amos Storkey et al. · microsoft-research

Successful deployment of multi-agent reinforcement learning often requires agents to adapt their behaviour. In this work, we discuss the problem of teamwork adaptation in which a team of agents needs to adapt their policies to solve novel tasks with limited fine-tuning. Motivated by the intuition that agents need to be able to identify and distinguish tasks in order to adapt their behaviour to the current task, we propose to learn multi-agent task embeddings (MATE). These task embeddings are trained using an encoder-decoder architecture optimised for reconstruction of the transition and reward functions which uniquely identify tasks. We show that a team of agents is able to adapt to novel tasks when provided with task embeddings. We propose three MATE training paradigms: independent MATE, centralised MATE, and mixed MATE which vary in the information used for the task encoding. We show that the embeddings learned by MATE identify tasks and provide useful information which agents leverage during adaptation to novel tasks.

Justin Engelmann, Amos Storkey, Miguel O. Bernabeu

Image quality remains a key problem for both traditional and deep learning (DL)-based approaches to retinal image analysis, but identifying poor quality images can be time consuming and subjective. Thus, automated methods for retinal image quality scoring (RIQS) are needed. The current state-of-the-art is MCFNet, composed of three Densenet121 backbones each operating in a different colour space. MCFNet, and the EyeQ dataset released by the same authors, was a huge step forward for RIQS. We present QuickQual, a simple approach to RIQS, consisting of a single off-the-shelf ImageNet-pretrained Densenet121 backbone plus a Support Vector Machine (SVM). QuickQual performs very well, setting a new state-of-the-art for EyeQ (Accuracy: 88.50% vs 88.00% for MCFNet; AUC: 0.9687 vs 0.9588). This suggests that RIQS can be solved with generic perceptual features learned on natural images, as opposed to requiring DL models trained on large amounts of fundus images. Additionally, we propose a Fixed Prior linearisation scheme, that converts EyeQ from a 3-way classification to a continuous logistic regression task. For this task, we present a second model, QuickQual MEga Minified Estimator (QuickQual-MEME), that consists of only 10 parameters on top of an off-the-shelf Densenet121 and can distinguish between gradable and ungradable images with an accuracy of 89.18% (AUC: 0.9537). Code and model are available on GitHub: https://github.com/justinengelmann/QuickQual . QuickQual is so lightweight, that the entire inference code (and even the parameters for QuickQual-MEME) is already contained in this paper.

Kejiang Qian, Amos Storkey, Fengxiang He

This paper proposes a suite of rationality measures and associated theory for reinforcement learning agents, a property increasingly critical yet rarely explored. We define an action in deployment to be perfectly rational if it maximises the hidden true value function in the steepest direction. The expected value discrepancy of a policy's actions against their rational counterparts, culminating over the trajectory in deployment, is defined to be expected rational risk; an empirical average version in training is also defined. Their difference, termed as rational risk gap, is decomposed into (1) an extrinsic component caused by environment shifts between training and deployment, and (2) an intrinsic one due to the algorithm's generalisability in a dynamic environment. They are upper bounded by, respectively, (1) the $1$-Wasserstein distance between transition kernels and initial state distributions in training and deployment, and (2) the empirical Rademacher complexity of the value function class. Our theory suggests hypotheses on the benefits from regularisers (including layer normalisation, $\ell_2$ regularisation, and weight normalisation) and domain randomisation, as well as the harm from environment shifts. Experiments are in full agreement with these hypotheses. The code is available at https://github.com/EVIEHub/Rationality.

Alessandro Fontanella, Grant Mair, Joanna Wardlaw et al.

Segmentation masks of pathological areas are useful in many medical applications, such as brain tumour and stroke management. Moreover, healthy counterfactuals of diseased images can be used to enhance radiologists' training files and to improve the interpretability of segmentation models. In this work, we present a weakly supervised method to generate a healthy version of a diseased image and then use it to obtain a pixel-wise anomaly map. To do so, we start by considering a saliency map that approximately covers the pathological areas, obtained with ACAT. Then, we propose a technique that allows to perform targeted modifications to these regions, while preserving the rest of the image. In particular, we employ a diffusion model trained on healthy samples and combine Denoising Diffusion Probabilistic Model (DDPM) and Denoising Diffusion Implicit Model (DDIM) at each step of the sampling process. DDPM is used to modify the areas affected by a lesion within the saliency map, while DDIM guarantees reconstruction of the normal anatomy outside of it. The two parts are also fused at each timestep, to guarantee the generation of a sample with a coherent appearance and a seamless transition between edited and unedited parts. We verify that when our method is applied to healthy samples, the input images are reconstructed without significant modifications. We compare our approach with alternative weakly supervised methods on the task of brain lesion segmentation, achieving the highest mean Dice and IoU scores among the models considered.

Alessandro Fontanella, Antreas Antoniou, Wenwen Li et al.

In some medical imaging tasks and other settings where only small parts of the image are informative for the classification task, traditional CNNs can sometimes struggle to generalise. Manually annotated Regions of Interest (ROI) are sometimes used to isolate the most informative parts of the image. However, these are expensive to collect and may vary significantly across annotators. To overcome these issues, we propose a framework that employs saliency maps to obtain soft spatial attention masks that modulate the image features at different scales. We refer to our method as Adversarial Counterfactual Attention (ACAT). ACAT increases the baseline classification accuracy of lesions in brain CT scans from 71.39% to 72.55% and of COVID-19 related findings in lung CT scans from 67.71% to 70.84% and exceeds the performance of competing methods. We investigate the best way to generate the saliency maps employed in our architecture and propose a way to obtain them from adversarially generated counterfactual images. They are able to isolate the area of interest in brain and lung CT scans without using any manual annotations. In the task of localising the lesion location out of 6 possible regions, they obtain a score of 65.05% on brain CT scans, improving the score of 61.29% obtained with the best competing method.

Justin Engelmann, Ana Villaplana-Velasco, Amos Storkey et al.

A retinal trait, or phenotype, summarises a specific aspect of a retinal image in a single number. This can then be used for further analyses, e.g. with statistical methods. However, reducing an aspect of a complex image to a single, meaningful number is challenging. Thus, methods for calculating retinal traits tend to be complex, multi-step pipelines that can only be applied to high quality images. This means that researchers often have to discard substantial portions of the available data. We hypothesise that such pipelines can be approximated with a single, simpler step that can be made robust to common quality issues. We propose Deep Approximation of Retinal Traits (DART) where a deep neural network is used predict the output of an existing pipeline on high quality images from synthetically degraded versions of these images. We demonstrate DART on retinal Fractal Dimension (FD) calculated by VAMPIRE, using retinal images from UK Biobank that previous work identified as high quality. Our method shows very high agreement with FD VAMPIRE on unseen test images (Pearson r=0.9572). Even when those images are severely degraded, DART can still recover an FD estimate that shows good agreement with FD VAMPIRE obtained from the original images (Pearson r=0.8817). This suggests that our method could enable researchers to discard fewer images in the future. Our method can compute FD for over 1,000img/s using a single GPU. We consider these to be very encouraging initial results and hope to develop this approach into a useful tool for retinal analysis.

Alessandro Fontanella, Wenwen Li, Grant Mair et al.

Computed Tomography (CT) is commonly used to image acute ischemic stroke (AIS) patients, but its interpretation by radiologists is time-consuming and subject to inter-observer variability. Deep learning (DL) techniques can provide automated CT brain scan assessment, but usually require annotated images. Aiming to develop a DL method for AIS using labelled but not annotated CT brain scans from patients with AIS, we designed a convolutional neural network-based DL algorithm using routinely-collected CT brain scans from the Third International Stroke Trial (IST-3), which were not acquired using strict research protocols. The DL model aimed to detect AIS lesions and classify the side of the brain affected. We explored the impact of AIS lesion features, background brain appearances, and timing on DL performance. From 5772 unique CT scans of 2347 AIS patients (median age 82), 54% had visible AIS lesions according to expert labelling. Our best-performing DL method achieved 72% accuracy for lesion presence and side. Lesions that were larger (80% accuracy) or multiple (87% accuracy for two lesions, 100% for three or more), were better detected. Follow-up scans had 76% accuracy, while baseline scans 67% accuracy. Chronic brain conditions reduced accuracy, particularly non-stroke lesions and old stroke lesions (32% and 31% error rates respectively). DL methods can be designed for AIS lesion detection on CT using the vast quantities of routinely-collected CT brain scan data. Ultimately, this should lead to more robust and widely-applicable methods.

Adam Jelley, Amos Storkey, Antreas Antoniou et al.

Many contrastive and meta-learning approaches learn representations by identifying common features in multiple views. However, the formalism for these approaches generally assumes features to be shared across views to be captured coherently. We consider the problem of learning a unified representation from partial observations, where useful features may be present in only some of the views. We approach this through a probabilistic formalism enabling views to map to representations with different levels of uncertainty in different components; these views can then be integrated with one another through marginalisation over that uncertainty. Our approach, Partial Observation Experts Modelling (POEM), then enables us to meta-learn consistent representations from partial observations. We evaluate our approach on an adaptation of a comprehensive few-shot learning benchmark, Meta-Dataset, and demonstrate the benefits of POEM over other meta-learning methods at representation learning from partial observations. We further demonstrate the utility of POEM by meta-learning to represent an environment from partial views observed by an agent exploring the environment.

Luke Darlow, Qiwen Deng, Ahmed Hassan et al.

It is challenging to scale time series forecasting models such that they forecast accurately for multiple distinct domains and datasets, all with potentially different underlying collection procedures (e.g., sample resolution), patterns (e.g., periodicity), and prediction requirements (e.g., reconstruction vs. forecasting). We call this general task universal forecasting. Existing methods usually assume that input data is regularly sampled, and they forecast to pre-determined horizons, resulting in failure to generalise outside of the scope of their training. We propose the DAM - a neural model that takes randomly sampled histories and outputs an adjustable basis composition as a continuous function of time for forecasting to non-fixed horizons. It involves three key components: (1) a flexible approach for using randomly sampled histories from a long-tail distribution, that enables an efficient global perspective of the underlying temporal dynamics while retaining focus on the recent history; (2) a transformer backbone that is trained on these actively sampled histories to produce, as representational output, (3) the basis coefficients of a continuous function of time. We show that a single univariate DAM, trained on 25 time series datasets, either outperformed or closely matched existing SoTA models at multivariate long-term forecasting across 18 datasets, including 8 held-out for zero-shot transfer, even though these models were trained to specialise for each dataset-horizon combination. This single DAM excels at zero-shot transfer and very-long-term forecasting, performs well at imputation, is interpretable via basis function composition and attention, can be tuned for different inference-cost requirements, is robust to missing and irregularly sampled data {by design}.

Trevor McInroe, Adam Jelley, Stefano V. Albrecht et al.

Offline pretraining with a static dataset followed by online fine-tuning (offline-to-online, or OtO) is a paradigm well matched to a real-world RL deployment process. In this scenario, we aim to find the best-performing policy within a limited budget of online interactions. Previous work in the OtO setting has focused on correcting for bias introduced by the policy-constraint mechanisms of offline RL algorithms. Such constraints keep the learned policy close to the behavior policy that collected the dataset, but we show this can unnecessarily limit policy performance if the behavior policy is far from optimal. Instead, we forgo constraints and frame OtO RL as an exploration problem that aims to maximize the benefit of online data-collection. We first study the major online RL exploration methods based on intrinsic rewards and UCB in the OtO setting, showing that intrinsic rewards add training instability through reward-function modification, and UCB methods are myopic and it is unclear which learned-component's ensemble to use for action selection. We then introduce an algorithm for planning to go out-of-distribution (PTGOOD) that avoids these issues. PTGOOD uses a non-myopic planning procedure that targets exploration in relatively high-reward regions of the state-action space unlikely to be visited by the behavior policy. By leveraging concepts from the Conditional Entropy Bottleneck, PTGOOD encourages data collected online to provide new information relevant to improving the final deployment policy without altering rewards. We show empirically in several continuous control tasks that PTGOOD significantly improves agent returns during online fine-tuning and avoids the suboptimal policy convergence that many of our baselines exhibit in several environments.

Asif Khan, Amos Storkey

In an unsupervised attack on variational autoencoders (VAEs), an adversary finds a small perturbation in an input sample that significantly changes its latent space encoding, thereby compromising the reconstruction for a fixed decoder. A known reason for such vulnerability is the distortions in the latent space resulting from a mismatch between approximated latent posterior and a prior distribution. Consequently, a slight change in an input sample can move its encoding to a low/zero density region in the latent space resulting in an unconstrained generation. This paper demonstrates that an optimal way for an adversary to attack VAEs is to exploit a directional bias of a stochastic pullback metric tensor induced by the encoder and decoder networks. The pullback metric tensor of an encoder measures the change in infinitesimal latent volume from an input to a latent space. Thus, it can be viewed as a lens to analyse the effect of input perturbations leading to latent space distortions. We propose robustness evaluation scores using the eigenspectrum of a pullback metric tensor. Moreover, we empirically show that the scores correlate with the robustness parameter $β$ of the $β-$VAE. Since increasing $β$ also degrades reconstruction quality, we demonstrate a simple alternative using \textit{mixup} training to fill the empty regions in the latent space, thus improving robustness with improved reconstruction.

Justin Engelmann, Alice D. McTrusty, Ian J. C. MacCormick et al.

Ultra-widefield (UWF) imaging is a promising modality that captures a larger retinal field of view compared to traditional fundus photography. Previous studies showed that deep learning (DL) models are effective for detecting retinal disease in UWF images, but primarily considered individual diseases under less-than-realistic conditions (excluding images with other diseases, artefacts, comorbidities, or borderline cases; and balancing healthy and diseased images) and did not systematically investigate which regions of the UWF images are relevant for disease detection. We first improve on the state of the field by proposing a DL model that can recognise multiple retinal diseases under more realistic conditions. We then use global explainability methods to identify which regions of the UWF images the model generally attends to. Our model performs very well, separating between healthy and diseased retinas with an area under the curve (AUC) of 0.9206 on an internal test set, and an AUC of 0.9841 on a challenging, external test set. When diagnosing specific diseases, the model attends to regions where we would expect those diseases to occur. We further identify the posterior pole as the most important region in a purely data-driven fashion. Surprisingly, 10% of the image around the posterior pole is sufficient for achieving comparable performance to having the full images available.

William Toner, Amos Storkey

Training neural network classifiers on datasets contaminated with noisy labels significantly increases the risk of overfitting. Thus, effectively implementing Early Stopping in noisy label environments is crucial. Under ideal circumstances, Early Stopping utilises a validation set uncorrupted by label noise to effectively monitor generalisation during training. However, obtaining a noise-free validation dataset can be costly and challenging to obtain. This study establishes that, in many typical learning environments, a noise-free validation set is not necessary for effective Early Stopping. Instead, near-optimal results can be achieved by monitoring accuracy on a noisy dataset - drawn from the same distribution as the noisy training set. Referred to as `Noisy Early Stopping' (NES), this method simplifies and reduces the cost of implementing Early Stopping. We provide theoretical insights into the conditions under which this method is effective and empirically demonstrate its robust performance across standard benchmarks using common loss functions.

Thomas L. Lee, Amos Storkey

Work on continual learning (CL) has thus far largely focused on the problems arising from shifts in the data distribution. However, CL can be decomposed into two sub-problems: (a) shifts in the data distribution, and (b) dealing with the fact that the data is split into chunks and so only a part of the data is available to be trained on at any point in time. In this work, we look at the latter sub-problem, the chunking of data. We show that chunking is an important part of CL, accounting for around half of the performance drop from offline learning in our experiments. Furthermore, our results reveal that current CL algorithms do not address the chunking sub-problem, only performing as well as plain SGD training when there is no shift in the data distribution. Therefore, we show that chunking is both an important and currently unaddressed sub-problem and until it is addressed CL methods will be capped in performance. Additionally, we analyse why performance drops when learning occurs on identically distributed chunks of data, and find that forgetting, which is often seen to be a problem due to distribution shift, still arises and is a significant problem. We also show that performance on the chunking sub-problem can be increased and that this performance transfers to the full CL setting, where there is distribution shift. Hence, we argue that work on chunking can help advance CL in general.

Asif Khan, Amos Storkey

Learning node-level representations of heterophilic graphs is crucial for various applications, including fraudster detection and protein function prediction. In such graphs, nodes share structural similarity identified by the equivalence of their connectivity which is implicitly encoded in the form of higher-order hierarchical information in the graphs. The contrastive methods are popular choices for learning the representation of nodes in a graph. However, existing contrastive methods struggle to capture higher-order graph structures. To address this limitation, we propose a novel multiview contrastive learning approach that integrates diffusion filters on graphs. By incorporating multiple graph views as augmentations, our method captures the structural equivalence in heterophilic graphs, enabling the discovery of hidden relationships and similarities not apparent in traditional node representations. Our approach outperforms baselines on synthetic and real structural datasets, surpassing the best baseline by $16.06\%$ on Cornell, $3.27\%$ on Texas, and $8.04\%$ on Wisconsin. Additionally, it consistently achieves superior performance on proximal tasks, demonstrating its effectiveness in uncovering structural information and improving downstream applications.

Perry Gibson, José Cano, Elliot J. Crowley et al.

Deep Neural Networks (DNNs) are extremely computationally demanding, which presents a large barrier to their deployment on resource-constrained devices. Since such devices are where many emerging deep learning applications lie (e.g., drones, vision-based medical technology), significant bodies of work from both the machine learning and systems communities have attempted to provide optimizations to accelerate DNNs. To help unify these two perspectives, in this paper we combine machine learning and systems techniques within the Deep Learning Acceleration Stack (DLAS), and demonstrate how these layers can be tightly dependent on each other with an across-stack perturbation study. We evaluate the impact on accuracy and inference time when varying different parameters of DLAS across two datasets, seven popular DNN architectures, four DNN compression techniques, three algorithmic primitives with sparse and dense variants, untuned and auto-scheduled code generation, and four hardware platforms. Our evaluation highlights how perturbations across DLAS parameters can cause significant variation and across-stack interactions. The highest level observation from our evaluation is that the model size, accuracy, and inference time are not guaranteed to be correlated. Overall we make 13 key observations, including that speedups provided by compression techniques are very hardware dependent, and that compiler auto-tuning can significantly alter what the best algorithm to use for a given configuration is. With DLAS, we aim to provide a reference framework to aid machine learning and systems practitioners in reasoning about the context in which their respective DNN acceleration solutions exist in. With our evaluation strongly motivating the need for co-design, we believe that DLAS can be a valuable concept for exploring the next generation of co-designed accelerated deep learning solutions.

William Toner, Amos Storkey

Training neural network classifiers on datasets with label noise poses a risk of overfitting them to the noisy labels. To address this issue, researchers have explored alternative loss functions that aim to be more robust. The `forward-correction' is a popular approach wherein the model outputs are noised before being evaluated against noisy data. When the true noise model is known, applying the forward-correction guarantees consistency of the learning algorithm. While providing some benefit, the correction is insufficient to prevent overfitting to finite noisy datasets. In this work, we propose an approach to tackling overfitting caused by label noise. We observe that the presence of label noise implies a lower bound on the noisy generalised risk. Motivated by this observation, we propose imposing a lower bound on the training loss to mitigate overfitting. Our main contribution is providing theoretical insights that allow us to approximate the lower bound given only an estimate of the average noise rate. We empirically demonstrate that using this bound significantly enhances robustness in various settings, with virtually no additional computational cost.

Vaisakh Shaj, Cameron Barker, Aidan Scannell et al.

State-space language models such as Mamba and gated linear attention (GLA) offer efficient alternatives to transformers due to their linear complexity and parallel training, but often lack the expressivity and robust state-tracking needed for complex reasoning. We address these limitations by reframing sequence modelling through a probabilistic lens, using Bayesian filters as a core primitive. While classical filters such as Kalman filters provide principled state estimation and uncertainty tracking, they are typically viewed as inherently sequential. We show that reparameterising the Kalman filter in information form enables its updates to be computed via an associative scan, allowing efficient parallel training. Building on this insight, we introduce the Kalman Linear Attention (KLA) layer, a neural sequence-modelling primitive that performs time-parallel probabilistic inference while maintaining explicit belief-state uncertainty. KLA offers strictly more expressive nonlinear updates and gating than GLA variants while retaining their computational advantages. On language modelling tasks, KLA matches or outperforms modern SSMs and GLAs across representative discrete token-manipulation and state-tracking benchmarks.

Kale-ab Tessera, Leonard Hinckeldey, Riccardo Zamboni et al.

Cooperative multi-agent reinforcement learning (MARL) is typically framed as a decentralised partially observable Markov decision process (Dec-POMDP), a setting whose hardness stems from two key challenges: partial observability and decentralised coordination. Genuinely solving such tasks requires Dec-POMDP reasoning, where agents use history to infer hidden states and coordinate based on local information. Yet it remains unclear whether popular benchmarks actually demand this reasoning or permit success via simpler strategies. We introduce a diagnostic suite combining statistically grounded performance comparisons and information-theoretic probes to audit the behavioural complexity of baseline policies (IPPO and MAPPO) across 37 scenarios spanning MPE, SMAX, Overcooked, Hanabi, and MaBrax. Our diagnostics reveal that success on these benchmarks rarely requires genuine Dec-POMDP reasoning. Reactive policies match the performance of memory-based agents in over half the scenarios, and emergent coordination frequently relies on brittle, synchronous action coupling rather than robust temporal influence. These findings suggest that some widely used benchmarks may not adequately test core Dec-POMDP assumptions under current training paradigms, potentially leading to over-optimistic assessments of progress. We release our diagnostic tooling to support more rigorous environment design and evaluation in cooperative MARL.

Thomas L. Lee, Edoardo M. Ponti, Amos Storkey

Time series foundation models (TSFMs) have become increasingly popular for zero-shot forecasting. However, for a new time series domain not fully covered by the pretraining set, performance can suffer. Therefore, when a practitioner cares about a new domain and has access to a set of related datasets, the question arises: how best to fine-tune a TSFM to improve zero-shot forecasting? A typical approach to this type of problem is to fine-tune a LoRA module on all datasets or separately on each dataset. Tuning a separate module on each dataset allows for the specialisation of the TSFM to different types of data distribution, by selecting differing combinations of per-dataset modules for different time series contexts. However, we find that, using per-dataset modules might not be optimal, since a time series dataset can contain data from several types of distributions, i.e. sub-domains. This can be due to the distribution shifting or having differing distributions for different dimensions of the time series. Hence, we propose MixFT which re-divides the data using Bayesian mixtures into sets that best represent the sub-domains present in the data, and fine-tunes separately on each of these sets. This re-division of the data ensures that each set is more homogeneous, leading to fine-tuned modules focused on specific sub-domains. Our experiments show that MixFT performs better than per-dataset methods and when fine-tuning a single module on all the data. This suggests that by re-partitioning the data to represent sub-domains we can better specialise TSFMs to improve zero-shot forecasting.

Ben Sanati, Thomas L. Lee, Trevor McInroe et al.

A fundamental challenge in developing general learning algorithms is their tendency to forget past knowledge when adapting to new data. Addressing this problem requires a principled understanding of forgetting; yet, despite decades of study, no unified definition has emerged that provides insights into the underlying dynamics of learning. We propose an algorithm- and task-agnostic theory that characterises forgetting as a lack of self-consistency in a learner's predictive distribution over future experiences, manifesting as a loss of predictive information. Our theory naturally yields a general measure of an algorithm's propensity to forget. To validate the theory, we design a comprehensive set of experiments that span classification, regression, generative modelling, and reinforcement learning. We empirically demonstrate how forgetting is present across all learning settings and plays a significant role in determining learning efficiency. Together, these results establish a principled understanding of forgetting and lay the foundation for analysing and improving the information retention capabilities of general learning algorithms.

Justin Engelmann, Jamie Burke, Charlene Hamid et al.

Purpose: To develop Choroidalyzer, an open-source, end-to-end pipeline for segmenting the choroid region, vessels, and fovea, and deriving choroidal thickness, area, and vascular index. Methods: We used 5,600 OCT B-scans (233 subjects, 6 systemic disease cohorts, 3 device types, 2 manufacturers). To generate region and vessel ground-truths, we used state-of-the-art automatic methods following manual correction of inaccurate segmentations, with foveal positions manually annotated. We trained a U-Net deep-learning model to detect the region, vessels, and fovea to calculate choroid thickness, area, and vascular index in a fovea-centred region of interest. We analysed segmentation agreement (AUC, Dice) and choroid metrics agreement (Pearson, Spearman, mean absolute error (MAE)) in internal and external test sets. We compared Choroidalyzer to two manual graders on a small subset of external test images and examined cases of high error. Results: Choroidalyzer took 0.299 seconds per image on a standard laptop and achieved excellent region (Dice: internal 0.9789, external 0.9749), very good vessel segmentation performance (Dice: internal 0.8817, external 0.8703) and excellent fovea location prediction (MAE: internal 3.9 pixels, external 3.4 pixels). For thickness, area, and vascular index, Pearson correlations were 0.9754, 0.9815, and 0.8285 (internal) / 0.9831, 0.9779, 0.7948 (external), respectively (all p<0.0001). Choroidalyzer's agreement with graders was comparable to the inter-grader agreement across all metrics. Conclusions: Choroidalyzer is an open-source, end-to-end pipeline that accurately segments the choroid and reliably extracts thickness, area, and vascular index. Especially choroidal vessel segmentation is a difficult and subjective task, and fully-automatic methods like Choroidalyzer could provide objectivity and standardisation.

Kale-ab Abebe Tessera, Arrasy Rahman, Amos Storkey et al.

Adaptive cooperation in multi-agent reinforcement learning (MARL) requires policies to express homogeneous, specialised, or mixed behaviours, yet achieving this adaptivity remains a critical challenge. While parameter sharing (PS) is standard for efficient learning, it notoriously suppresses the behavioural diversity required for specialisation. This failure is largely due to cross-agent gradient interference, a problem we find is surprisingly exacerbated by the common practice of coupling agent IDs with observations. Existing remedies typically add complexity through altered objectives, manual preset diversity levels, or sequential updates -- raising a fundamental question: can shared policies adapt without these intricacies? We propose a solution built on a key insight: an agent-conditioned hypernetwork can generate agent-specific parameters and decouple observation- and agent-conditioned gradients, directly countering the interference from coupling agent IDs with observations. Our resulting method, HyperMARL, avoids the complexities of prior work and empirically reduces policy gradient variance. Across diverse MARL benchmarks (22 scenarios, up to 30 agents), HyperMARL achieves performance competitive with six key baselines while preserving behavioural diversity comparable to non-parameter sharing methods, establishing it as a versatile and principled approach for adaptive MARL. The code is publicly available at https://github.com/KaleabTessera/HyperMARL.

Eloi Alonso, Adam Jelley, Vincent Micheli et al.

World models constitute a promising approach for training reinforcement learning agents in a safe and sample-efficient manner. Recent world models predominantly operate on sequences of discrete latent variables to model environment dynamics. However, this compression into a compact discrete representation may ignore visual details that are important for reinforcement learning. Concurrently, diffusion models have become a dominant approach for image generation, challenging well-established methods modeling discrete latents. Motivated by this paradigm shift, we introduce DIAMOND (DIffusion As a Model Of eNvironment Dreams), a reinforcement learning agent trained in a diffusion world model. We analyze the key design choices that are required to make diffusion suitable for world modeling, and demonstrate how improved visual details can lead to improved agent performance. DIAMOND achieves a mean human normalized score of 1.46 on the competitive Atari 100k benchmark; a new best for agents trained entirely within a world model. We further demonstrate that DIAMOND's diffusion world model can stand alone as an interactive neural game engine by training on static Counter-Strike: Global Offensive gameplay. To foster future research on diffusion for world modeling, we release our code, agents, videos and playable world models at https://diamond-wm.github.io.

Riccardo Mereu, Aidan Scannell, Yuxin Hou et al.

World models are a powerful paradigm in AI and robotics, enabling agents to reason about the future by predicting visual observations or compact latent states. The 1X World Model Challenge introduces an open-source benchmark of real-world humanoid interaction, with two complementary tracks: sampling, focused on forecasting future image frames, and compression, focused on predicting future discrete latent codes. For the sampling track, we adapt the video generation foundation model Wan-2.2 TI2V-5B to video-state-conditioned future frame prediction. We condition the video generation on robot states using AdaLN-Zero, and further post-train the model using LoRA. For the compression track, we train a Spatio-Temporal Transformer model from scratch. Our models achieve 23.0 dB PSNR in the sampling task and a Top-500 CE of 6.6386 in the compression task, securing 1st place in both challenges.

Adam Jelley, Trevor McInroe, Sam Devlin et al.

Supervised imitation-based approaches are often favored over off-policy reinforcement learning approaches for learning policies offline, since their straightforward optimization objective makes them computationally efficient and stable to train. However, their performance is fundamentally limited by the behavior policy that collected the dataset. Off-policy reinforcement learning provides a promising approach for improving on the behavior policy, but training is often computationally inefficient and unstable due to temporal-difference bootstrapping. In this paper, we propose a best-of-both approach by pre-training with supervised learning before improving performance with off-policy reinforcement learning. Specifically, we demonstrate improved efficiency by pre-training an actor with behavior cloning and a critic with a supervised Monte-Carlo value error. We find that we are able to substantially improve the training time of popular off-policy algorithms on standard benchmarks, and also achieve greater stability. Code is available at: https://github.com/AdamJelley/EfficientOfflineRL

Jamie Burke, Justin Engelmann, Charlene Hamid et al.

The choroid is a key vascular layer of the eye, supplying oxygen to the retinal photoreceptors. Non-invasive enhanced depth imaging optical coherence tomography (EDI-OCT) has recently improved access and visualisation of the choroid, making it an exciting frontier for discovering novel vascular biomarkers in ophthalmology and wider systemic health. However, current methods to measure the choroid often require use of multiple, independent semi-automatic and deep learning-based algorithms which are not made open-source. Previously, Choroidalyzer -- an open-source, fully automatic deep learning method trained on 5,600 OCT B-scans from 385 eyes -- was developed to fully segment and quantify the choroid in EDI-OCT images, thus addressing these issues. Using the same dataset, we propose a Robust, Resolution-agnostic and Efficient Attention-based network for CHoroid segmentation (REACH). REACHNet leverages multi-resolution training with domain-specific data augmentation to promote generalisation, and uses a lightweight architecture with resolution-agnostic self-attention which is not only faster than Choroidalyzer's previous network (4 images/s vs. 2.75 images/s on a standard laptop CPU), but has greater performance for segmenting the choroid region, vessels and fovea (Dice coefficient for region 0.9769 vs. 0.9749, vessels 0.8612 vs. 0.8192 and fovea 0.8243 vs. 0.3783) due to its improved hyperparameter configuration and model training pipeline. REACHNet can be used with Choroidalyzer as a drop-in replacement for the original model and will be made available upon publication.

Mateusz Ochal, Massimiliano Patacchiola, Amos Storkey et al.

Meta-Learning (ML) has proven to be a useful tool for training Few-Shot Learning (FSL) algorithms by exposure to batches of tasks sampled from a meta-dataset. However, the standard training procedure overlooks the dynamic nature of the real-world where object classes are likely to occur at different frequencies. While it is generally understood that imbalanced tasks harm the performance of supervised methods, there is no significant research examining the impact of imbalanced meta-datasets on the FSL evaluation task. This study exposes the magnitude and extent of this problem. Our results show that ML methods are more robust against meta-dataset imbalance than imbalance at the task-level with a similar imbalance ratio ($ρ<20$), with the effect holding even in long-tail datasets under a larger imbalance ($ρ=65$). Overall, these results highlight an implicit strength of ML algorithms, capable of learning generalizable features under dataset imbalance and domain-shift. The code to reproduce the experiments is released under an open-source license.

Paul Micaelli, Amos Storkey

Gradient-based hyperparameter optimization has earned a widespread popularity in the context of few-shot meta-learning, but remains broadly impractical for tasks with long horizons (many gradient steps), due to memory scaling and gradient degradation issues. A common workaround is to learn hyperparameters online, but this introduces greediness which comes with a significant performance drop. We propose forward-mode differentiation with sharing (FDS), a simple and efficient algorithm which tackles memory scaling issues with forward-mode differentiation, and gradient degradation issues by sharing hyperparameters that are contiguous in time. We provide theoretical guarantees about the noise reduction properties of our algorithm, and demonstrate its efficiency empirically by differentiating through $\sim 10^4$ gradient steps of unrolled optimization. We consider large hyperparameter search ranges on CIFAR-10 where we significantly outperform greedy gradient-based alternatives, while achieving $\times 20$ speedups compared to the state-of-the-art black-box methods. Code is available at: \url{https://github.com/polo5/FDS}

Perry Gibson, José Cano, Jack Turner et al.

When deploying a deep neural network on constrained hardware, it is possible to replace the network's standard convolutions with grouped convolutions. This allows for substantial memory savings with minimal loss of accuracy. However, current implementations of grouped convolutions in modern deep learning frameworks are far from performing optimally in terms of speed. In this paper we propose Grouped Spatial Pack Convolutions (GSPC), a new implementation of grouped convolutions that outperforms existing solutions. We implement GSPC in TVM, which provides state-of-the-art performance on edge devices. We analyze a set of networks utilizing different types of grouped convolutions and evaluate their performance in terms of inference time on several edge devices. We observe that our new implementation scales well with the number of groups and provides the best inference times in all settings, improving the existing implementations of grouped convolutions in TVM, PyTorch and TensorFlow Lite by 3.4x, 8x and 4x on average respectively. Code is available at https://github.com/gecLAB/tvm-GSPC/

Joseph Mellor, Jack Turner, Amos Storkey et al.

The time and effort involved in hand-designing deep neural networks is immense. This has prompted the development of Neural Architecture Search (NAS) techniques to automate this design. However, NAS algorithms tend to be slow and expensive; they need to train vast numbers of candidate networks to inform the search process. This could be alleviated if we could partially predict a network's trained accuracy from its initial state. In this work, we examine the overlap of activations between datapoints in untrained networks and motivate how this can give a measure which is usefully indicative of a network's trained performance. We incorporate this measure into a simple algorithm that allows us to search for powerful networks without any training in a matter of seconds on a single GPU, and verify its effectiveness on NAS-Bench-101, NAS-Bench-201, NATS-Bench, and Network Design Spaces. Our approach can be readily combined with more expensive search methods; we examine a simple adaptation of regularised evolutionary search. Code for reproducing our experiments is available at https://github.com/BayesWatch/nas-without-training.

Jack Turner, Elliot J. Crowley, Michael O'Boyle et al.

The desire to map neural networks to varying-capacity devices has led to the development of a wealth of compression techniques, many of which involve replacing standard convolutional blocks in a large network with cheap alternative blocks. However, not all blocks are created equally; for a required compute budget there may exist a potent combination of many different cheap blocks, though exhaustively searching for such a combination is prohibitively expensive. In this work, we develop BlockSwap: a fast algorithm for choosing networks with interleaved block types by passing a single minibatch of training data through randomly initialised networks and gauging their Fisher potential. These networks can then be used as students and distilled with the original large network as a teacher. We demonstrate the effectiveness of the chosen networks across CIFAR-10 and ImageNet for classification, and COCO for detection, and provide a comprehensive ablation study of our approach. BlockSwap quickly explores possible block configurations using a simple architecture ranking system, yielding highly competitive networks in orders of magnitude less time than most architecture search techniques (e.g. under 5 minutes on a single GPU for CIFAR-10). Code is available at https://github.com/BayesWatch/pytorch-blockswap.

Gavin Gray, Elliot J. Crowley, Amos Storkey

In response to the development of recent efficient dense layers, this paper shows that something as simple as replacing linear components in pointwise convolutions with structured linear decompositions also produces substantial gains in the efficiency/accuracy tradeoff. Pointwise convolutions are fully connected layers and are thus prepared for replacement by structured transforms. Networks using such layers are able to learn the same tasks as those using standard convolutions, and provide Pareto-optimal benefits in efficiency/accuracy, both in terms of computation (mult-adds) and parameter count (and hence memory). Code is available at https://github.com/BayesWatch/deficient-efficient.

Paul Micaelli, Amos Storkey

Performing knowledge transfer from a large teacher network to a smaller student is a popular task in modern deep learning applications. However, due to growing dataset sizes and stricter privacy regulations, it is increasingly common not to have access to the data that was used to train the teacher. We propose a novel method which trains a student to match the predictions of its teacher without using any data or metadata. We achieve this by training an adversarial generator to search for images on which the student poorly matches the teacher, and then using them to train the student. Our resulting student closely approximates its teacher for simple datasets like SVHN, and on CIFAR10 we improve on the state-of-the-art for few-shot distillation (with 100 images per class), despite using no data. Finally, we also propose a metric to quantify the degree of belief matching between teacher and student in the vicinity of decision boundaries, and observe a significantly higher match between our zero-shot student and the teacher, than between a student distilled with real data and the teacher. Code available at: https://github.com/polo5/ZeroShotKnowledgeTransfer

Elliot J. Crowley, Jack Turner, Amos Storkey et al.

Structured pruning is a popular method for compressing a neural network: given a large trained network, one alternates between removing channel connections and fine-tuning; reducing the overall width of the network. However, the efficacy of structured pruning has largely evaded scrutiny. In this paper, we examine ResNets and DenseNets obtained through structured pruning-and-tuning and make two interesting observations: (i) reduced networks---smaller versions of the original network trained from scratch---consistently outperform pruned networks; (ii) if one takes the architecture of a pruned network and then trains it from scratch it is significantly more competitive. Furthermore, these architectures are easy to approximate: we can prune once and obtain a family of new, scalable network architectures that can simply be trained from scratch. Finally, we compare the inference speed of reduced and pruned networks on hardware, and show that reduced networks are significantly faster. Code is available at https://github.com/BayesWatch/pytorch-prunes.

Dongge Han, Trevor McInroe, Adam Jelley et al.

Large language models (LLMs) have shown significant potential for robotics applications, particularly task planning, by harnessing their language comprehension and text generation capabilities. However, in applications such as household robotics, a critical gap remains in the personalization of these models to individual user preferences. We introduce LLM-Personalize, a novel framework with an optimization pipeline designed to personalize LLM planners for household robotics. Our LLM-Personalize framework features an LLM planner that performs iterative planning in multi-room, partially-observable household scenarios, making use of a scene graph constructed with local observations. The generated plan consists of a sequence of high-level actions which are subsequently executed by a controller. Central to our approach is the optimization pipeline, which combines imitation learning and iterative self-training to personalize the LLM planner. In particular, the imitation learning phase performs initial LLM alignment from demonstrations, and bootstraps the model to facilitate effective iterative self-training, which further explores and aligns the model to user preferences. We evaluate LLM-Personalize on Housekeep, a challenging simulated real-world 3D benchmark for household rearrangements, and show that LLM-Personalize achieves more than a 30 percent increase in success rate over existing LLM planners, showcasing significantly improved alignment with human preferences. Project page: https://gdg94.github.io/projectllmpersonalize/.

Weipu Zhang, Adam Jelley, Trevor McInroe et al.

Deep reinforcement learning has achieved remarkable success in learning control policies from pixels across a wide range of tasks, yet its application remains hindered by low sample efficiency, requiring significantly more environment interactions than humans to reach comparable performance. Model-based reinforcement learning (MBRL) offers a solution by leveraging learnt world models to generate simulated experience, thereby improving sample efficiency. However, in visually complex environments, small or dynamic elements can be critical for decision-making. Yet, traditional MBRL methods in pixel-based environments typically rely on auto-encoding with an $L_2$ loss, which is dominated by large areas and often fails to capture decision-relevant details. To address these limitations, we propose an object-centric MBRL pipeline, which integrates recent advances in computer vision to allow agents to focus on key decision-related elements. Our approach consists of four main steps: (1) annotating key objects related to rewards and goals with segmentation masks, (2) extracting object features using a pre-trained, frozen foundation vision model, (3) incorporating these object features with the raw observations to predict environmental dynamics, and (4) training the policy using imagined trajectories generated by this object-centric world model. Building on the efficient MBRL algorithm STORM, we call this pipeline OC-STORM. We demonstrate OC-STORM's practical value in overcoming the limitations of conventional MBRL approaches on both Atari games and the visually complex game Hollow Knight.

Leyang Xue, Meghana Madhyastha, Myungjin Lee et al.

The ecosystem behind foundation model development today is highly centralized and limited to large-scale cloud data center operators: training foundation models is costly, needing immense compute resources. Decentralized foundation model training across edge devices, leveraging their spare compute, promises a democratized alternative. However, existing edge-training approaches fall short: they struggle to match cloud-based training performance, exhibit limited scalability with model size, exceed device memory capacity, and have prohibitive communication overhead. They also fail to satisfactorily handle device heterogeneity and dynamism. We introduce a new paradigm, Cleave, which finely partitions training operations through a novel selective hybrid tensor parallelism method. Together with a parameter server centric training framework, Cleave copes with device memory limits and avoids communication bottlenecks, thereby enabling efficient training of large models on par with the cloud. Further, with a cost optimization model to guide device selection and training workload distribution, Cleave effectively accounts for device heterogeneity and churn. Our evaluations show that Cleave matches cloud-based GPU training by scaling efficiently to larger models and thousands of devices, supporting up to 8x more devices than baseline edge-training approaches. It outperforms state-of-the-art edge training methods by up to a factor of 10 in per-batch training time and efficiently handles device failures, achieving at least 100x faster recovery than prior methods.

Benjamin R. Redhead, Thomas L. Lee, Peng Gu et al.

Probabilistic forecasting is increasingly critical across high-stakes domains, from finance and epidemiology to climate science. However, current evaluation frameworks lack a consensus metric and suffer from two critical flaws: they often assume independence across time steps or variables, and they demonstrably lack sensitivity to tail events, the very occurrences that are most pivotal in real-world decision-making. To address these limitations, we propose two kernel-based metrics: the signature maximum mean discrepancy (Sig-MMD) and our novel censored Sig-MMD (CSig-MMD). By leveraging the signature kernel, these metrics capture complex inter-variate and inter-temporal dependencies and remain robust to missing data. Furthermore, CSig-MMD introduces a censoring scheme that prioritizes a forecaster's capability to predict tail events while strictly maintaining properness, a vital property for a good scoring rule. These metrics enable a more reliable evaluation of direct multi-step forecasting, facilitating the development of more robust probabilistic algorithms.

Kale-ab Abebe Tessera, Leonard Hinckeldey, Riccardo Zamboni et al.

Cooperative multi-agent reinforcement learning (MARL) is typically formalised as a Decentralised Partially Observable Markov Decision Process (Dec-POMDP), where agents must reason about the environment and other agents' behaviour. In practice, current model-free MARL algorithms use simple recurrent function approximators to address the challenge of reasoning about others using partial information. In this position paper, we argue that the empirical success of these methods is not due to effective Markov signal recovery, but rather to learning simple conventions that bypass environment observations and memory. Through a targeted case study, we show that co-adapting agents can learn brittle conventions, which then fail when partnered with non-adaptive agents. Crucially, the same models can learn grounded policies when the task design necessitates it, revealing that the issue is not a fundamental limitation of the learning models but a failure of the benchmark design. Our analysis also suggests that modern MARL environments may not adequately test the core assumptions of Dec-POMDPs. We therefore advocate for new cooperative environments built upon two core principles: (1) behaviours grounded in observations and (2) memory-based reasoning about other agents, ensuring success requires genuine skill rather than fragile, co-adapted agreements.

Adam Jelley, Yuhan Cao, Dave Bignell et al.

Training agents to act competently in complex 3D environments from high-dimensional visual information is challenging. Reinforcement learning is conventionally used to train such agents, but requires a carefully designed reward function, and is difficult to scale to obtain robust agents that generalize to new tasks. In contrast, Large Language Models (LLMs) demonstrate impressively general capabilities resulting from large-scale pre-training and post-training alignment, but struggle to act in complex environments. This position paper draws explicit analogies between decision-making agents and LLMs, and argues that agents should be trained like LLMs to achieve more general, robust, and aligned behaviors. We provide a proof-of-concept to demonstrate how the procedure for training LLMs can be used to train an agent in a 3D video game environment from pixels. We investigate the importance of each stage of the LLM training pipeline, while providing guidance and insights for successfully applying this approach to agents. Our paper provides an alternative perspective to contemporary LLM Agents on how recent progress in LLMs can be leveraged for decision-making agents, and we hope will illuminate a path towards developing more generally capable agents for video games and beyond. Project summary and videos: https://adamjelley.github.io/aligning-agents-like-llms .

Thomas L. Lee, Sigrid Passano Hellan, Linus Ericsson et al.

In continual learning (CL) -- where a learner trains on a stream of data -- standard hyperparameter optimisation (HPO) cannot be applied, as a learner does not have access to all of the data at the same time. This has prompted the development of CL-specific HPO frameworks. The most popular way to tune hyperparameters in CL is to repeatedly train over the whole data stream with different hyperparameter settings. However, this end-of-training HPO is unusable in practice since a learner can only see the stream once. Hence, there is an open question: what HPO framework should a practitioner use for a CL problem in reality? This paper looks at this question by comparing several realistic HPO frameworks. We find that none of the HPO frameworks considered, including end-of-training HPO, perform consistently better than the rest on popular CL benchmarks. We therefore arrive at a twofold conclusion: a) to be able to discriminate between HPO frameworks there is a need to move beyond the current most commonly used CL benchmarks, and b) on the popular CL benchmarks examined, a CL practitioner should use a realistic HPO framework and can select it based on factors separate from performance, for example compute efficiency.

Thomas L. Lee, Amos Storkey

For models consisting of a classifier in some representation space, learning online from a non-stationary data stream often necessitates changes in the representation. So, the question arises of what is the best way to adapt the classifier to shifts in representation. Current methods only slowly change the classifier to representation shift, introducing noise into learning as the classifier is misaligned to the representation. We propose DeepCCG, an empirical Bayesian approach to solve this problem. DeepCCG works by updating the posterior of a class conditional Gaussian classifier such that the classifier adapts in one step to representation shift. The use of a class conditional Gaussian classifier also enables DeepCCG to use a log conditional marginal likelihood loss to update the representation. To perform the update to the classifier and representation, DeepCCG maintains a fixed number of examples in memory and so a key part of DeepCCG is selecting what examples to store, choosing the subset that minimises the KL divergence between the true posterior and the posterior induced by the subset. We explore the behaviour of DeepCCG in online continual learning (CL), demonstrating that it performs well against a spectrum of online CL methods and that it reduces the change in performance due to representation shift.

Asif Khan, Alexander I. Cowen-Rivers, Antoine Grosnit et al.

Antibodies are canonically Y-shaped multimeric proteins capable of highly specific molecular recognition. The CDRH3 region located at the tip of variable chains of an antibody dominates antigen-binding specificity. Therefore, it is a priority to design optimal antigen-specific CDRH3 regions to develop therapeutic antibodies. However, the combinatorial nature of CDRH3 sequence space makes it impossible to search for an optimal binding sequence exhaustively and efficiently using computational approaches. Here, we present \texttt{AntBO}: a combinatorial Bayesian optimisation framework enabling efficient \textit{in silico} design of the CDRH3 region. Ideally, antibodies are expected to have high target specificity and developability. We introduce a CDRH3 trust region that restricts the search to sequences with favourable developability scores to achieve this goal. For benchmarking, \texttt{AntBO} uses the \texttt{Absolut!} software suite as a black-box oracle to score the target specificity and affinity of designed antibodies \textit{in silico} in an unconstrained fashion~\citep{robert2021one}. The experiments performed for $159$ discretised antigens used in \texttt{Absolut!} demonstrate the benefit of \texttt{AntBO} in designing CDRH3 regions with diverse biophysical properties. In under $200$ calls to black-box oracle, \texttt{AntBO} can suggest antibody sequences that outperform the best binding sequence drawn from 6.9 million experimentally obtained CDRH3s and a commonly used genetic algorithm baseline. Additionally, \texttt{AntBO} finds very-high affinity CDRH3 sequences in only 38 protein designs whilst requiring no domain knowledge. We conclude \texttt{AntBO} brings automated antibody design methods closer to what is practically viable for in vitro experimentation.

Justin Engelmann, Amos Storkey, Miguel O. Bernabeu

Deep learning (DL) models are very effective on many computer vision problems and increasingly used in critical applications. They are also inherently black box. A number of methods exist to generate image-wise explanations that allow practitioners to understand and verify model predictions for a given image. Beyond that, it would be desirable to validate that a DL model \textit{generally} works in a sensible way, i.e. consistent with domain knowledge and not relying on undesirable data artefacts. For this purpose, the model needs to be explained globally. In this work, we focus on image modalities that are naturally aligned such that each pixel position represents a similar relative position on the imaged object, as is common in medical imaging. We propose the pixel-wise aggregation of image-wise explanations as a simple method to obtain label-wise and overall global explanations. These can then be used for model validation, knowledge discovery, and as an efficient way to communicate qualitative conclusions drawn from inspecting image-wise explanations. We further propose Progressive Erasing Plus Progressive Restoration (PEPPR) as a method to quantitatively validate that these global explanations are faithful to how the model makes its predictions. We then apply these methods to ultra-widefield retinal images, a naturally aligned modality. We find that the global explanations are consistent with domain knowledge and faithfully reflect the model's workings.

Asif Khan, Amos Storkey

We introduce \textit{HALO} -- a deep generative model utilising HAmiltonian Latent Operators to reliably disentangle content and motion information in image sequences. The \textit{content} represents summary statistics of a sequence, and \textit{motion} is a dynamic process that determines how information is expressed in any part of the sequence. By modelling the dynamics as a Hamiltonian motion, important desiderata are ensured: (1) the motion is reversible, (2) the symplectic, volume-preserving structure in phase space means paths are continuous and are not divergent in the latent space. Consequently, the nearness of sequence frames is realised by the nearness of their coordinates in the phase space, which proves valuable for disentanglement and long-term sequence generation. The sequence space is generally comprised of different types of dynamical motions. To ensure long-term separability and allow controlled generation, we associate every motion with a unique Hamiltonian that acts in its respective subspace. We demonstrate the utility of \textit{HALO} by swapping the motion of a pair of sequences, controlled generation, and image rotations.

Benedict Leimkuhler, Tiffany Vlaar, Timothée Pouchon et al.

We employ constraints to control the parameter space of deep neural networks throughout training. The use of customized, appropriately designed constraints can reduce the vanishing/exploding gradients problem, improve smoothness of classification boundaries, control weight magnitudes and stabilize deep neural networks, and thus enhance the robustness of training algorithms and the generalization capabilities of neural networks. We provide a general approach to efficiently incorporate constraints into a stochastic gradient Langevin framework, allowing enhanced exploration of the loss landscape. We also present specific examples of constrained training methods motivated by orthogonality preservation for weight matrices and explicit weight normalizations. Discretization schemes are provided both for the overdamped formulation of Langevin dynamics and the underdamped form, in which momenta further improve sampling efficiency. These optimization schemes can be used directly, without needing to adapt neural network architecture design choices or to modify the objective with regularization terms, and see performance improvements in classification tasks.

Mateusz Ochal, Massimiliano Patacchiola, Amos Storkey et al.

Few-Shot Learning (FSL) algorithms are commonly trained through Meta-Learning (ML), which exposes models to batches of tasks sampled from a meta-dataset to mimic tasks seen during evaluation. However, the standard training procedures overlook the real-world dynamics where classes commonly occur at different frequencies. While it is generally understood that class imbalance harms the performance of supervised methods, limited research examines the impact of imbalance on the FSL evaluation task. Our analysis compares 10 state-of-the-art meta-learning and FSL methods on different imbalance distributions and rebalancing techniques. Our results reveal that 1) some FSL methods display a natural disposition against imbalance while most other approaches produce a performance drop by up to 17\% compared to the balanced task without the appropriate mitigation; 2) contrary to popular belief, many meta-learning algorithms will not automatically learn to balance from exposure to imbalanced training tasks; 3) classical rebalancing strategies, such as random oversampling, can still be very effective, leading to state-of-the-art performances and should not be overlooked; 4) FSL methods are more robust against meta-dataset imbalance than imbalance at the task-level with a similar imbalance ratio ($ρ<20$), with the effect holding even in long-tail datasets under a larger imbalance ($ρ=65$).