Robert McCraith

Yushi Yang, Andrew M. Bean, Robert McCraith et al.

Fine-tuning Large Language Models (LLMs) incurs considerable training costs, driving the need for data-efficient training with optimised data ordering. Human-inspired strategies offer a solution by organising data based on human learning practices. This study evaluates the fine-tuning efficiency of five human-inspired strategies across four language models, three datasets, and both human- and LLM-labelled data in the context of medical question answering. These strategies achieve the best accuracy gain of 1.81% and an average gain of 1.02% across datasets, with interleaved strategies delivering the best average results. However, the best strategy varies across model-dataset combinations, limiting the generalisability of the effects of any single strategy. Additionally, LLM-defined question difficulty outperforms human-defined labels in curriculum-based learning, showing the potential of model-generated data as a cost-effective alternative for optimising fine-tuning.

Andrew M. Bean, Karolina Korgul, Felix Krones et al.

Large language models (LLMs) have made rapid improvement on medical benchmarks, but their unreliability remains a persistent challenge for safe real-world uses. To design for the use LLMs as a category, rather than for specific models, requires developing an understanding of shared strengths and weaknesses which appear across models. To address this challenge, we benchmark a range of top LLMs and identify consistent patterns across models. We test $16$ well-known LLMs on $874$ newly collected questions from Polish medical licensing exams. For each question, we score each model on the top-1 accuracy and the distribution of probabilities assigned. We then compare these results with factors such as question difficulty for humans, question length, and the scores of the other models. LLM accuracies were positively correlated pairwise ($0.39$ to $0.58$). Model performance was also correlated with human performance ($0.09$ to $0.13$), but negatively correlated to the difference between the question-level accuracy of top-scoring and bottom-scoring humans ($-0.09$ to $-0.14$). The top output probability and question length were positive and negative predictors of accuracy respectively (p$< 0.05$). The top scoring LLM, GPT-4o Turbo, scored $84\%$, with Claude Opus, Gemini 1.5 Pro and Llama 3/3.1 between $74\%$ and $79\%$. We found evidence of similarities between models in which questions they answer correctly, as well as similarities with human test takers. Larger models typically performed better, but differences in training, architecture, and data were also highly impactful. Model accuracy was positively correlated with confidence, but negatively correlated with question length. We find similar results with older models, and argue that these patterns are likely to persist across future models using similar training methods.

M. Mehdi Mojarradi, Lingyi Yang, Robert McCraith et al.

Large language models (LLMs) have shown impressive capabilities across various tasks, but their performance on domain-specific tasks remains limited. While methods like retrieval augmented generation and fine-tuning can help to address this, they require significant resources. In-context learning (ICL) is a cheap and efficient alternative but cannot match the accuracies of advanced methods. We present Ensemble SuperICL, a novel approach that enhances ICL by leveraging the expertise of multiple fine-tuned small language models (SLMs). Ensemble SuperICL achieves state of the art (SoTA) results on several natural language understanding benchmarks. Additionally, we test it on a medical-domain labelling task and showcase its practicality by using off-the-shelf SLMs fine-tuned on a general language task, achieving superior accuracy in large-scale data labelling compared to all baselines. Finally, we conduct an ablation study and sensitivity analyses to elucidate the underlying mechanism of Ensemble SuperICL. Our research contributes to the growing demand for efficient domain specialisation methods in LLMs, offering a cheap and effective method for practitioners.

Robert McCraith, Lukas Neumann, Andrea Vedaldi

Vision is one of the primary sensing modalities in autonomous driving. In this paper we look at the problem of estimating the velocity of road vehicles from a camera mounted on a moving car. Contrary to prior methods that train end-to-end deep networks that estimate the vehicles' velocity from the video pixels, we propose a two-step approach where first an off-the-shelf tracker is used to extract vehicle bounding boxes and then a small neural network is used to regress the vehicle velocity from the tracked bounding boxes. Surprisingly, we find that this still achieves state-of-the-art estimation performance with the significant benefit of separating perception from dynamics estimation via a clean, interpretable and verifiable interface which allows us distill the statistics which are crucial for velocity estimation. We show that the latter can be used to easily generate synthetic training data in the space of bounding boxes and use this to improve the performance of our method further.

Robert McCraith, Eldar Insafutdinov, Lukas Neumann et al.

We present a system for automatic converting of 2D mask object predictions and raw LiDAR point clouds into full 3D bounding boxes of objects. Because the LiDAR point clouds are partial, directly fitting bounding boxes to the point clouds is meaningless. Instead, we suggest that obtaining good results requires sharing information between \emph{all} objects in the dataset jointly, over multiple frames. We then make three improvements to the baseline. First, we address ambiguities in predicting the object rotations via direct optimization in this space while still backpropagating rotation prediction through the model. Second, we explicitly model outliers and task the network with learning their typical patterns, thus better discounting them. Third, we enforce temporal consistency when video data is available. With these contributions, our method significantly outperforms previous work despite the fact that those methods use significantly more complex pipelines, 3D models and additional human-annotated external sources of prior information.

Robert McCraith, Lukas Neumann, Andrea Vedaldi

In the recent years, many methods demonstrated the ability of neural networks to learn depth and pose changes in a sequence of images, using only self-supervision as the training signal. Whilst the networks achieve good performance, the often over-looked detail is that due to the inherent ambiguity of monocular vision they predict depth up to an unknown scaling factor. The scaling factor is then typically obtained from the LiDAR ground truth at test time, which severely limits practical applications of these methods. In this paper, we show that incorporating prior information about the camera configuration and the environment, we can remove the scale ambiguity and predict depth directly, still using the self-supervised formulation and not relying on any additional sensors.

Robert McCraith, Lukas Neumann, Andrew Zisserman et al.

Recent advances in self-supervised learning havedemonstrated that it is possible to learn accurate monoculardepth reconstruction from raw video data, without using any 3Dground truth for supervision. However, in robotics applications,multiple views of a scene may or may not be available, depend-ing on the actions of the robot, switching between monocularand multi-view reconstruction. To address this mixed setting,we proposed a new approach that extends any off-the-shelfself-supervised monocular depth reconstruction system to usemore than one image at test time. Our method builds on astandard prior learned to perform monocular reconstruction,but uses self-supervision at test time to further improve thereconstruction accuracy when multiple images are available.When used to update the correct components of the model, thisapproach is highly-effective. On the standard KITTI bench-mark, our self-supervised method consistently outperformsall the previous methods with an average 25% reduction inabsolute error for the three common setups (monocular, stereoand monocular+stereo), and comes very close in accuracy whencompared to the fully-supervised state-of-the-art methods.

Samuel Albanie, Jaime Thewmore, Robert McCraith et al.

Peer review forms the backbone of modern scientific manuscript evaluation. But after two hundred and eighty-nine years of egalitarian service to the scientific community, does this protocol remain fit for purpose in 2020? In this work, we answer this question in the negative (strong reject, high confidence) and propose instead State-Of-the-Art Review (SOAR), a neoteric reviewing pipeline that serves as a 'plug-and-play' replacement for peer review. At the heart of our approach is an interpretation of the review process as a multi-objective, massively distributed and extremely-high-latency optimisation, which we scalarise and solve efficiently for PAC and CMT-optimal solutions. We make the following contributions: (1) We propose a highly scalable, fully automatic methodology for review, drawing inspiration from best-practices from premier computer vision and machine learning conferences; (2) We explore several instantiations of our approach and demonstrate that SOAR can be used to both review prints and pre-review pre-prints; (3) We wander listlessly in vain search of catharsis from our latest rounds of savage CVPR rejections.