Bruce A. Bassett

Nicholas Pather, Joshua Fouché, Sitwala Mundia et al.

Manual digitisation of structured handwritten documents is slow and costly. We benchmark 17 leading frontier multi-modal large language models and open-source models against a very challenging real-world medical form that mixes dates; structured, printed text; hand-written responses and significant variability challenges. None of the smaller or older models perform well but the latest Google and OpenAI models reach accuracies around $85\%$ with weighted F1 scores $\simeq 90\%$ across the discrete or predefined fields despite the very challenging nature of the responses. Clear task specific strengths emerge: GPT 5.4 excels in noisy date extraction as well as reliability with the lowest hallucination rate ($6\%$). Claude Sonnet 4.6 had the best average performance across formatted fields (dates and numerical values), while Gemini 3.1 delivered the best overall performance, with the lowest free text error rates (WER = $0.50$ and CER = $0.31$) and the strongest results across discrete classification metrics. We further show that prompt optimisation dramatically improves macro precision, recall and F1 by over $60\%$, but has little impact on weighted metrics (only $\sim2-5\%$ improvement). These results provide evidence that the rapid improvements of multimodal large language models offer a compelling pathway toward fully automated digitisation of complex handwritten workflows that is particularly relevant in low- and middle-income countries.

Christiaan Jacobs, Nathanaël Carraz Rakotonirina, Everlyn Asiko Chimoto et al.

We consider hate speech detection through keyword spotting on radio broadcasts. One approach is to build an automatic speech recognition (ASR) system for the target low-resource language. We compare this to using acoustic word embedding (AWE) models that map speech segments to a space where matching words have similar vectors. We specifically use a multilingual AWE model trained on labelled data from well-resourced languages to spot keywords in data in the unseen target language. In contrast to ASR, the AWE approach only requires a few keyword exemplars. In controlled experiments on Wolof and Swahili where training and test data are from the same domain, an ASR model trained on just five minutes of data outperforms the AWE approach. But in an in-the-wild test on Swahili radio broadcasts with actual hate speech keywords, the AWE model (using one minute of template data) is more robust, giving similar performance to an ASR system trained on 30 hours of labelled data.

Amy Rouillard, Sitwala Mundia, Linda Camara et al.

Evaluating medical AI systems using expert clinician panels is costly and slow, motivating the use of large language models (LLMs) as alternative adjudicators. Here, we evaluate an LLM jury composed of three frontier AI models scoring 3333 diagnoses on 300 real-world middle-income country (MIC) hospital cases. Model performance was benchmarked against expert clinician panel and independent human re-scoring panel evaluations. Both LLM and clinician-generated diagnoses are scored across four dimensions: diagnosis, differential diagnosis, clinical reasoning and negative treatment risk. For each of these, we assess scoring difference, inter-rater agreement, scoring stability, severe safety errors and the effect of post-hoc calibration. We find that: (i) the uncalibrated LLM jury scores are systematically lower than clinician panels scores; (ii) the LLM Jury preserves ordinal agreement and exhibits better concordance with the primary expert panels than the human expert re-score panels do; (iii) the probability of severe errors is lower in \lj models compared to the human expert re-score panels; (iv) the LLM Jury shows excellent agreement with primary expert panels' rankings. We find that the LLM jury combined with AI model diagnoses can be used to identify ward diagnoses at high risk of error, enabling targeted expert review and improved panel efficiency; (v) LLM jury models show no self-preference bias. They did not score diagnoses generated by their own underlying model or models from the same vendor more (or less) favourably than those generated by other models. Finally, we demonstrate that LLM jury calibration using isotonic regression improves alignment with human expert panel evaluations. Together, these results provide compelling evidence that a calibrated, multi-model LLM jury can serve as a trustworthy and reliable proxy for expert clinician evaluation in medical AI benchmarking.

Bruce A. Bassett, Amy Rouillard, Sitwala Mundia et al.

Background: Large language models (LLMs) are increasingly proposed for diagnostic support, but few evaluations use real-world multimodal inpatient data, particularly in low and middle-income country (LMIC) public hospitals. Methods: We conducted VALID, a retrospective evaluation of 539 multimodal inpatient cases from a tertiary public hospital in South Africa. Inputs included radiology imaging (CT, MRI, CXR) and reports, laboratory results, clinical notes, and vital signs. Expert panels adjudicated 300 cases (balanced and discordant subsets) to establish ground truth diagnoses, differentials, and reasoning. Ten multimodal LLMs generated zero-shot outputs. A calibrated three-model LLM Jury scored all outputs and routine ward diagnoses across diagnostic accuracy, differential quality, reasoning, and patient safety (>10,000 evaluations). Primary outcomes were composite scores ($S_3$, $S_4$) and win rates. Results: (i) LLM performance was tightly clustered (<15% variation) despite large cost differences; low-cost models performed comparably to top models. (ii) All LLMs significantly outperformed routine ward diagnoses on average diagnostic and safety scores. (iii) Top performance was achieved by GPT-5.1, followed by Gemini models. (vi) Adding radiology reports improved performance by 6%. (v) Diagnostic and reasoning scores were highly correlated ($ρ= 0.85$). (vi) Output rates varied (65-100%) due to input constraints. Results were robust across subsets and evaluation design. Conclusions: Across a real-world LMIC dataset, multimodal LLMs showed similar diagnostic performance despite large cost differences and outperformed routine care on average safety metrics. Affordability, robustness, and deployment constraints may outweigh marginal performance differences in LMIC settings.

Everlyn Asiko Chimoto, Bruce A. Bassett

Active learning aims to deliver maximum benefit when resources are scarce. We use COMET-QE, a reference-free evaluation metric, to select sentences for low-resource neural machine translation. Using Swahili, Kinyarwanda and Spanish for our experiments, we show that COMET-QE significantly outperforms two variants of Round Trip Translation Likelihood (RTTL) and random sentence selection by up to 5 BLEU points for 20k sentences selected by Active Learning on a 30k baseline. This suggests that COMET-QE is a powerful tool for sentence selection in the very low-resource limit.

Everlyn Asiko Chimoto, Bruce A. Bassett

Language-agnostic sentence embeddings generated by pre-trained models such as LASER and LaBSE are attractive options for mining large datasets to produce parallel corpora for low-resource machine translation. We test LASER and LaBSE in extracting bitext for two related low-resource African languages: Luhya and Swahili. For this work, we created a new parallel set of nearly 8000 Luhya-English sentences which allows a new zero-shot test of LASER and LaBSE. We find that LaBSE significantly outperforms LASER on both languages. Both LASER and LaBSE however perform poorly at zero-shot alignment on Luhya, achieving just 1.5% and 22.0% successful alignments respectively (P@1 score). We fine-tune the embeddings on a small set of parallel Luhya sentences and show significant gains, improving the LaBSE alignment accuracy to 53.3%. Further, restricting the dataset to sentence embedding pairs with cosine similarity above 0.7 yielded alignments with over 85% accuracy.

Alireza Vafaei Sadr, Bruce A. Bassett, Emmanuel Sekyi

Anomaly detection algorithms are typically applied to static, unchanging, data features hand-crafted by the user. But how does a user systematically craft good features for anomalies that have never been seen? Here we couple deep learning with active learning -- in which an Oracle iteratively labels small amounts of data selected algorithmically over a series of rounds -- to automatically and dynamically improve the data features for efficient outlier detection. This approach, AHUNT, shows excellent performance on MNIST, CIFAR10, and Galaxy-DESI data, significantly outperforming both standard anomaly detection and active learning algorithms with static feature spaces. Beyond improved performance, AHUNT also allows the number of anomaly classes to grow organically in response to Oracle's evaluations. Extensive ablation studies explore the impact of Oracle question selection strategy and loss function on performance. We illustrate how the dynamic anomaly class taxonomy represents another step towards fully personalized rankings of different anomaly classes that reflect a user's interests, allowing the algorithm to learn to ignore statistically significant but uninteresting outliers (e.g., noise). This should prove useful in the era of massive astronomical datasets serving diverse sets of users who can only review a tiny subset of the incoming data.

Everlyn Asiko Chimoto, Mostafa Elhoushi, Bruce A. Bassett

Quantization is an effective technique for reducing the storage footprint and computational costs of Large Language Models (LLMs), but it often results in performance degradation. Existing post-training quantization methods typically use small, English-only calibration sets; however, their impact on multilingual models remains underexplored. We systematically evaluate eight calibration settings (five single-language and three multilingual mixes) on two quantizers (GPTQ, AWQ) on data from 10 languages. Our findings reveal a consistent trend: non-English and multilingual calibration sets significantly improve perplexity compared to English-only baselines. Specifically, we observe notable average perplexity gains across both quantizers on Llama3.1 8B and Qwen2.5 7B, with multilingual mixes achieving the largest overall reductions of up to 3.52 points in perplexity. Furthermore, our analysis indicates that tailoring calibration sets to the evaluation language yields the largest improvements for individual languages, underscoring the importance of linguistic alignment. We also identify specific failure cases where certain language-quantizer combinations degrade performance, which we trace to differences in activation range distributions across languages. These results highlight that static one-size-fits-all calibration is suboptimal and that tailoring calibration data, both in language and diversity, plays a crucial role in robustly quantizing multilingual LLMs.

Evander Nyoni, Bruce A. Bassett

Low-resource African languages have not fully benefited from the progress in neural machine translation because of a lack of data. Motivated by this challenge we compare zero-shot learning, transfer learning and multilingual learning on three Bantu languages (Shona, isiXhosa and isiZulu) and English. Our main target is English-to-isiZulu translation for which we have just 30,000 sentence pairs, 28% of the average size of our other corpora. We show the importance of language similarity on the performance of English-to-isiZulu transfer learning based on English-to-isiXhosa and English-to-Shona parent models whose BLEU scores differ by 5.2. We then demonstrate that multilingual learning surpasses both transfer learning and zero-shot learning on our dataset, with BLEU score improvements relative to the baseline English-to-isiZulu model of 9.9, 6.1 and 2.0 respectively. Our best model also improves the previous SOTA BLEU score by more than 10.

Emmanuel Dufourq, Bruce A. Bassett

Deep facial expression recognition faces two challenges that both stem from the large number of trainable parameters: long training times and a lack of interpretability. We propose a novel method based on evolutionary algorithms, that deals with both challenges by massively reducing the number of trainable parameters, whilst simultaneously retaining classification performance, and in some cases achieving superior performance. We are robustly able to reduce the number of parameters on average by 95% (e.g. from 2M to 100k parameters) with no loss in classification accuracy. The algorithm learns to choose small patches from the image, relative to the nose, which carry the most important information about emotion, and which coincide with typical human choices of important features. Our work implements a novel form attention and shows that evolutionary algorithms are a valuable addition to machine learning in the deep learning era, both for reducing the number of parameters for facial expression recognition and for providing interpretable features that can help reduce bias.

Alireza Vafaei Sadr, Bruce A. Bassett, Martin Kunz

Anomaly detection is challenging, especially for large datasets in high dimensions. Here we explore a general anomaly detection framework based on dimensionality reduction and unsupervised clustering. We release DRAMA, a general python package that implements the general framework with a wide range of built-in options. We test DRAMA on a wide variety of simulated and real datasets, in up to 3000 dimensions, and find it robust and highly competitive with commonly-used anomaly detection algorithms, especially in high dimensions. The flexibility of the DRAMA framework allows for significant optimization once some examples of anomalies are available, making it ideal for online anomaly detection, active learning and highly unbalanced datasets.

Ethan Roberts, Bruce A. Bassett, Michelle Lochner

Statistical uncertainties are rarely incorporated in machine learning algorithms, especially for anomaly detection. Here we present the Bayesian Anomaly Detection And Classification (BADAC) formalism, which provides a unified statistical approach to classification and anomaly detection within a hierarchical Bayesian framework. BADAC deals with uncertainties by marginalising over the unknown, true, value of the data. Using simulated data with Gaussian noise, BADAC is shown to be superior to standard algorithms in both classification and anomaly detection performance in the presence of uncertainties, though with significantly increased computational cost. Additionally, BADAC provides well-calibrated classification probabilities, valuable for use in scientific pipelines. We show that BADAC can work in online mode and is fairly robust to model errors, which can be diagnosed through model-selection methods. In addition it can perform unsupervised new class detection and can naturally be extended to search for anomalous subsets of data. BADAC is therefore ideal where computational cost is not a limiting factor and statistical rigour is important. We discuss approximations to speed up BADAC, such as the use of Gaussian processes, and finally introduce a new metric, the Rank-Weighted Score (RWS), that is particularly suited to evaluating the ability of algorithms to detect anomalies.

A. Vafaei Sadr, Etienne. E. Vos, Bruce A. Bassett et al.

Point source detection at low signal-to-noise is challenging for astronomical surveys, particularly in radio interferometry images where the noise is correlated. Machine learning is a promising solution, allowing the development of algorithms tailored to specific telescope arrays and science cases. We present DeepSource - a deep learning solution - that uses convolutional neural networks to achieve these goals. DeepSource enhances the Signal-to-Noise Ratio (SNR) of the original map and then uses dynamic blob detection to detect sources. Trained and tested on two sets of 500 simulated 1 deg x 1 deg MeerKAT images with a total of 300,000 sources, DeepSource is essentially perfect in both purity and completeness down to SNR = 4 and outperforms PyBDSF in all metrics. For uniformly-weighted images it achieves a Purity x Completeness (PC) score at SNR = 3 of 0.73, compared to 0.31 for the best PyBDSF model. For natural-weighting we find a smaller improvement of ~40% in the PC score at SNR = 3. If instead we ask where either of the purity or completeness first drop to 90%, we find that DeepSource reaches this value at SNR = 3.6 compared to the 4.3 of PyBDSF (natural-weighting). A key advantage of DeepSource is that it can learn to optimally trade off purity and completeness for any science case under consideration. Our results show that deep learning is a promising approach to point source detection in astronomical images.

Emmanuel Dufourq, Bruce A. Bassett

The ability to identify sentiment in text, referred to as sentiment analysis, is one which is natural to adult humans. This task is, however, not one which a computer can perform by default. Identifying sentiments in an automated, algorithmic manner will be a useful capability for business and research in their search to understand what consumers think about their products or services and to understand human sociology. Here we propose two new Genetic Algorithms (GAs) for the task of automated text sentiment analysis. The GAs learn whether words occurring in a text corpus are either sentiment or amplifier words, and their corresponding magnitude. Sentiment words, such as 'horrible', add linearly to the final sentiment. Amplifier words in contrast, which are typically adjectives/adverbs like 'very', multiply the sentiment of the following word. This increases, decreases or negates the sentiment of the following word. The sentiment of the full text is then the sum of these terms. This approach grows both a sentiment and amplifier dictionary which can be reused for other purposes and fed into other machine learning algorithms. We report the results of multiple experiments conducted on large Amazon data sets. The results reveal that our proposed approach was able to outperform several public and/or commercial sentiment analysis algorithms.

Emmanuel Dufourq, Bruce A. Bassett

Deep neural networks continue to show improved performance with increasing depth, an encouraging trend that implies an explosion in the possible permutations of network architectures and hyperparameters for which there is little intuitive guidance. To address this increasing complexity, we propose Evolutionary DEep Networks (EDEN), a computationally efficient neuro-evolutionary algorithm which interfaces to any deep neural network platform, such as TensorFlow. We show that EDEN evolves simple yet successful architectures built from embedding, 1D and 2D convolutional, max pooling and fully connected layers along with their hyperparameters. Evaluation of EDEN across seven image and sentiment classification datasets shows that it reliably finds good networks -- and in three cases achieves state-of-the-art results -- even on a single GPU, in just 6-24 hours. Our study provides a first attempt at applying neuro-evolution to the creation of 1D convolutional networks for sentiment analysis including the optimisation of the embedding layer.

Emmanuel Dufourq, Bruce A. Bassett

Can textual data be compressed intelligently without losing accuracy in evaluating sentiment? In this study, we propose a novel evolutionary compression algorithm, PARSEC (PARts-of-Speech for sEntiment Compression), which makes use of Parts-of-Speech tags to compress text in a way that sacrifices minimal classification accuracy when used in conjunction with sentiment analysis algorithms. An analysis of PARSEC with eight commercial and non-commercial sentiment analysis algorithms on twelve English sentiment data sets reveals that accurate compression is possible with (0%, 1.3%, 3.3%) loss in sentiment classification accuracy for (20%, 50%, 75%) data compression with PARSEC using LingPipe, the most accurate of the sentiment algorithms. Other sentiment analysis algorithms are more severely affected by compression. We conclude that significant compression of text data is possible for sentiment analysis depending on the accuracy demands of the specific application and the specific sentiment analysis algorithm used.

Emmanuel Dufourq, Bruce A. Bassett

Regression or classification? This is perhaps the most basic question faced when tackling a new supervised learning problem. We present an Evolutionary Deep Learning (EDL) algorithm that automatically solves this by identifying the question type with high accuracy, along with a proposed deep architecture. Typically, a significant amount of human insight and preparation is required prior to executing machine learning algorithms. For example, when creating deep neural networks, the number of parameters must be selected in advance and furthermore, a lot of these choices are made based upon pre-existing knowledge of the data such as the use of a categorical cross entropy loss function. Humans are able to study a dataset and decide whether it represents a classification or a regression problem, and consequently make decisions which will be applied to the execution of the neural network. We propose the Automated Problem Identification (API) algorithm, which uses an evolutionary algorithm interface to TensorFlow to manipulate a deep neural network to decide if a dataset represents a classification or a regression problem. We test API on 16 different classification, regression and sentiment analysis datasets with up to 10,000 features and up to 17,000 unique target values. API achieves an average accuracy of $96.3\%$ in identifying the problem type without hardcoding any insights about the general characteristics of regression or classification problems. For example, API successfully identifies classification problems even with 1000 target values. Furthermore, the algorithm recommends which loss function to use and also recommends a neural network architecture. Our work is therefore a step towards fully automated machine learning.