When effective vaccines are available, vaccination programs are typically one of the best defences against the spread of an infectious disease. Such vaccination programs become particularly important during severe epidemics or pandemics to ensure sufficient vaccination coverage is achieved to increase protection or reduce transmission to a level that enables relaxation of non-pharmaceutical interventions, such as lockdowns, travel restrictions, or social distancing. Unfortunately, vaccination uptake in the community may be slow if there are substantial levels of vaccine hesitancy in the population. As a result, it is important to identify when these hesitancy behaviours are present in the community. Furthermore, understanding the main drivers of such behaviour can inform adjustments to public health strategies to improve community uptake. In this study, we consider the problem of identifying vaccination hesitancy behaviour during a vaccination roll-out that occurs in response to a severe epidemic. Specifically, our aim is to explore the extent to which mathematical modelling of reported case, death, and vaccination counts can be used to detect vaccine hesitancy and possible drivers. To do this, we develop a novel susceptible-exposed-infectious-recovered (SEIR) epidemiological model of disease transmission that incorporates changes in population behaviour relating to non-pharmaceutical interventions and vaccine uptake that are influenced by information reported through media or data dashboards about cases, deaths, and vaccination rates. We then use a Bayesian approach to analyse simulated data representing various hesitancy scenarios. Through this simulation study, our key findings are that individual parameters values related to drivers of vaccine hesitancy often cannot be identified. However, posterior correlation structures between these parameters enable the presence of vaccine hesitancy in the community to be detected and provide some insight into the relative influence of key factors, such as vaccine safety concerns or complacency. While our simulation study is inspired by the public health response to the COVID-19 pandemic, our tools and techniques are general and could enable vaccination programs of various infectious diseases to be adapted rapidly in response to community behaviours in the future.

Atomic structure prediction and associated property calculations are the bedrock of chemical physics. Since high-fidelity ab initio modeling techniques for computing the structure and properties can be prohibitively expensive, this motivates the development of machine-learning (ML) models that make these predictions more efficiently. Training graph neural networks over large atomistic databases introduces unique computational challenges, such as the need to process millions of small graphs with variable size and support communication patterns that are distinct from learning over large graphs, such as social networks. We demonstrate a novel hardware-software codesign approach to scale up the training of atomistic graph neural networks (GNN) for structure and property prediction. First, to eliminate redundant computation and memory associated with alternative padding techniques and to improve throughput via minimizing communication, we formulate the effective coalescing of the batches of variable-size atomistic graphs as the bin packing problem and introduce a hardware-agnostic algorithm to pack these batches. In addition, we propose hardware-specific optimizations, including a planner and vectorization for the gather-scatter operations targeted for Graphcore's Intelligence Processing Unit (IPU), as well as model-specific optimizations such as merged communication collectives and optimized softplus. Putting these all together, we demonstrate the effectiveness of the proposed codesign approach by providing an implementation of a well-established atomistic GNN on the Graphcore IPUs. We evaluate the training performance on multiple atomistic graph databases with varying degrees of graph counts, sizes, and sparsity. We demonstrate that such a codesign approach can reduce the training time of atomistic GNNs and can improve their performance by up to 1.5× compared to the baseline implementation of the model on the IPUs. Additionally, we compare our IPU implementation with a Nvidia GPU-based implementation and show that our atomistic GNN implementation on the IPUs can run 1.8× faster on average compared to the execution time on the GPUs.

Epidemiology models are central to understanding and controlling large-scale pandemics. Several epidemiology models require simulation-based inference such as Approximate Bayesian Computation (ABC) to fit their parameters to observations. ABC inference is highly amenable to efficient hardware acceleration. In this work, we develop parallel ABC inference of a stochastic epidemiology model for COVID-19. The statistical inference framework is implemented and compared on Intel’s Xeon CPU, NVIDIA’s Tesla V100 GPU, Google’s V2 Tensor Processing Unit (TPU), and the Graphcore’s Mk1 Intelligence Processing Unit (IPU), and the results are discussed in the context of their computational architectures. Results show that TPUs are 3×, GPUs are 4×, and IPUs are 30× faster than Xeon CPUs. Extensive performance analysis indicates that the difference between IPU and GPU can be attributed to higher communication bandwidth, closeness of memory to compute, and higher compute power in the IPU. The proposed framework scales across 16 IPUs, with scaling overhead not exceeding 8% for the experiments performed. We present an example of our framework in practice, performing inference on the epidemiology model across three countries and giving a brief overview of the results.

When processing a batch of graphs in machine learning models such as Graph Neural Networks (GNN), it is common to combine several small graphs into one overall graph to accelerate processing and remove or reduce the overhead of padding. This is for example supported in the PyG library. However, the sizes of small graphs can vary substantially with respect to the number of nodes and edges, and hence the size of the combined graph can still vary considerably, especially for small batch sizes. Therefore, the costs of excessive padding and wasted compute are still incurred when working with static shapes, which are preferred for maximum acceleration. This paper proposes a new hardware agnostic approach -- tuple packing -- for generating batches that cause minimal overhead. The algorithm extends recently introduced sequence packing approaches to work on the 2D tuples of (|nodes|, |edges|). A monotone heuristic is applied to the 2D histogram of tuple values to define a priority for packing histogram bins together with the objective to reach a limit on the number of nodes as well as the number of edges. Experiments verify the effectiveness of the algorithm on multiple datasets.

Molecular property calculations are the bedrock of chemical physics. High-fidelity \textit{ab initio} modeling techniques for computing the molecular properties can be prohibitively expensive, and motivate the development of machine-learning models that make the same predictions more efficiently. Training graph neural networks over large molecular databases introduces unique computational challenges such as the need to process millions of small graphs with variable size and support communication patterns that are distinct from learning over large graphs such as social networks. This paper demonstrates a novel hardware-software co-design approach to scale up the training of graph neural networks for molecular property prediction. We introduce an algorithm to coalesce the batches of molecular graphs into fixed size packs to eliminate redundant computation and memory associated with alternative padding techniques and improve throughput via minimizing communication. We demonstrate the effectiveness of our co-design approach by providing an implementation of a well-established molecular property prediction model on the Graphcore Intelligence Processing Units (IPU). We evaluate the training performance on multiple molecular graph databases with varying degrees of graph counts, sizes and sparsity. We demonstrate that such a co-design approach can reduce the training time of such molecular property prediction models from days to less than two hours, opening new possibilities for AI-driven scientific discovery.

Upcoming regional and National Science Foundation (NSF)-funded Cyberinfrastructure (CI) resources will give researchers opportunities to run their artificial intelligence / machine learning (AI/ML) workflows on accelerators. To effectively leverage this burgeoning CI-rich landscape, researchers need extensive benchmark data to maximize performance gains and map their workflows to appropriate architectures. This data will further assist CI administrators, NSF program officers, and CI allocation-reviewers make informed determinations on CI-resource allocations. Here, we compare the performance of two very different architectures: the commonly used Graphical Processing Units (GPUs) and the new generation of Intelligence Processing Units (IPUs), by running training benchmarks of common AI/ML models. We leverage the maturity of software stacks, and the ease of migration among these platforms to learn that performance and scaling are similar for both architectures. Exploring training parameters, such as batch size, however finds that owing to memory processing structures, IPUs run efficiently with smaller batch sizes, while GPUs benefit from large batch sizes to extract sufficient parallelism in neural network training and inference. This comes with different advantages and disadvantages as discussed in this paper.As such considerations of inference latency, inherent parallelism and model accuracy will play a role in researcher selection of these architectures. The impact of these choices on a representative image compression model system is discussed.

Effective training of today's large language models (LLMs) depends on large batches and long sequences for throughput and accuracy. To handle variable-length sequences on hardware accelerators, it is common practice to introduce padding tokens, so that all sequences in a batch have the same length. We show in this paper that the variation in sequence lengths in common NLP datasets is such that up to 50% of all tokens can be padding. In less common, but not extreme, cases (e.g. GLUE-COLA with sequence length 128), the ratio is up to 89%. Existing methods to address the resulting inefficiency are complicated by the need to avoid "cross-contamination" in self-attention, by a reduction in accuracy when sequence ordering information is lost, or by customized kernel implementations only valid for specific accelerators. This paper introduces a new formalization of sequence packing in the context of the well-studied bin packing problem, and presents new algorithms based on this formulation which, for example, confer a 2x speedup for phase 2 pretraining in BERT while preserving downstream performance. We show how existing models can be adapted to ensure mathematical equivalence between the original and packed models, meaning that packed models can be trained with existing pre-training and fine-tuning practices.

We find that at sequence length 512 padding tokens represent in excess of 50% of the Wikipedia dataset used for pretraining BERT (Bidirectional Encoder Representations from Transformers). Therefore by removing all padding we achieve a 2x speed-up in terms of sequences/sec. To exploit this characteristic of the dataset, we develop and contrast two deterministic packing algorithms. Both algorithms rely on the assumption that sequences are interchangeable and therefore packing can be performed on the histogram of sequence lengths, rather than per sample. This transformation of the problem leads to algorithms which are fast and have linear complexity in dataset size. The shortest-pack-first histogram-packing (SPFHP) algorithm determines the packing order for the Wikipedia dataset of over 16M sequences in 0.02 seconds. The non-negative least-squares histogram-packing (NNLSHP) algorithm converges in 28.4 seconds but produces solutions which are more depth efficient, managing to get near optimal packing by combining a maximum of 3 sequences in one sample. Using the dataset with multiple sequences per sample requires adjusting the model and the hyperparameters. We demonstrate that these changes are straightforward to implement and have relatively little impact on the achievable performance gain on modern hardware. Finally, we pretrain BERT-Large using the packed dataset, demonstrating no loss of convergence and the desired 2x speed-up.

Origami is becoming more and more relevant to research. However, there is no public dataset yet available and there hasn't been any research on this topic in machine learning. We constructed an origami dataset using images from the multimedia commons and other databases. It consists of two subsets: one for classification of origami images and the other for difficulty estimation. We obtained 16000 images for classification (half origami, half other objects) and 1509 for difficulty estimation with $3$ different categories (easy: 764, intermediate: 427, complex: 318). The data can be downloaded at: https://github.com/multimedia-berkeley/OriSet. Finally, we provide machine learning baselines.

Differentially private SGD (DPSGD) has recently shown promise in deep learning. However, compared to non-private SGD, the DPSGD algorithm places computational overheads that can undo the benefit of batching in GPUs. Micro-batching is a common method to alleviate this and is fully supported in the TensorFlow Privacy library (TFDP). However, it degrades accuracy. We propose NanoBatch Privacy, a lightweight add-on to TFDP to be used on Graphcore IPUs by leveraging batch size of 1 (without microbatching) and gradient accumulation. This allows us to achieve large total batch sizes with minimal impacts to throughput. Second, we illustrate using Cifar-10 how larger batch sizes are not necessarily optimal from a privacy versus utility perspective. On ImageNet, we achieve more than 15x speedup over TFDP versus 8x A100s and significant speedups even across libraries such as Opacus. We also provide two extensions: 1) DPSGD for pipelined models and 2) per-layer clipping that is 15x faster than the Opacus implementation on 8x A100s. Finally as an application case study, we apply NanoBatch training for use on private Covid-19 chest CT prediction.

We find that at sequence length 512 padding tokens represent in excess of $50\%$ of the Wikipedia dataset used for pretraining BERT (Bidirectional Encoder Representations from Transformers). Therefore by removing all padding we achieve a 2x speed-up in terms of sequences/sec. To exploit this characteristic of the dataset, we develop and contrast two deterministic packing algorithms. Both algorithms rely on the assumption that sequences are interchangeable and therefore packing can be performed on the histogram of sequence lengths, rather than per sample. This transformation of the problem leads to algorithms which are fast and have linear complexity in dataset size. The shortest-pack-first histogram-packing (SPFHP) algorithm determines the packing order for the Wikipedia dataset of over $16$M sequences in $0.02$ seconds. The non-negative least-squares histogram-packing (NNLSHP) algorithm converges in $28.4$ seconds but produces solutions which are more depth efficient, managing to get near optimal packing by combining a maximum of $3$ sequences in one sample. Using the dataset with multiple sequences per sample requires additional masking in the attention layer and a modification of the MLM loss function. We demonstrate that both of these changes are straightforward to implement and have relatively little impact on the achievable performance gain on modern hardware. Finally, we pretrain BERT-Large using the packed dataset, demonstrating no loss of convergence and the desired 2x speed-up.

Developing models of natural phenomena by capturing their underlying complex interactions is a core tenet of various scientific disciplines. These models are useful as simulators and can help in understanding the natural processes being studied. One key challenge in this pursuit has been to enable statistical inference over these models, which would allow these simulation-based models to learn from real-world observations. Recent efforts, such as Approximate Bayesian Computation (ABC), show promise in performing a new kind of inference to leverage these models. While the scope of applicability of these inference algorithms is limited by the capabilities of contemporary computational hardware, they show potential of being greatly parallelized. In this work, we explore hardware accelerated simulation-based inference over probabilistic models, by combining massively parallelized ABC inference algorithm with the cutting-edge AI chip solutions that are uniquely suited for this purpose. As a proof-of-concept, we demonstrate inference over a probabilistic epidemiology model used to predict the spread of COVID-19. Two hardware acceleration platforms are compared - the Tesla V100 GPU and the Graphcore Mark1 IPU. Our results show that while both of these platforms outperform multi-core CPUs, the Mk1 IPUs are 7.5x faster than the Tesla V100 GPUs for this workload.

Regression evaluation has been performed for decades. Some metrics have been identified to be robust against shifting and scaling of the data but considering the different distributions of data is much more difficult to address (imbalance problem) even though it largely impacts the comparability between evaluations on different datasets. In classification, it has been stated repeatedly that performance metrics like the F-Measure and Accuracy are highly dependent on the class distribution and that comparisons between different datasets with different distributions are impossible. We show that the same problem exists in regression. The distribution of odometry parameters in robotic applications can for example largely vary between different recording sessions. Here, we need regression algorithms that either perform equally well for all function values, or that focus on certain boundary regions like high speed. This has to be reflected in the evaluation metric. We propose the modification of established regression metrics by weighting with the inverse distribution of function values $Y$ or the samples $X$ using an automatically tuned Gaussian kernel density estimator. We show on synthetic and robotic data in reproducible experiments that classical metrics behave wrongly, whereas our new metrics are less sensitive to changing distributions, especially when correcting by the marginal distribution in $X$. Our new evaluation concept enables the comparison of results between different datasets with different distributions. Furthermore, it can reveal overfitting of a regression algorithm to overrepresented target values. As an outcome, non-overfitting regression algorithms will be more likely chosen due to our corrected metrics.

On image data, data augmentation is becoming less relevant due to the large amount of available training data and regularization techniques. Common approaches are moving windows (cropping), scaling, affine distortions, random noise, and elastic deformations. For electroencephalographic data, the lack of sufficient training data is still a major issue. We suggest and evaluate different approaches to generate augmented data using temporal and spatial/rotational distortions. Our results on the perception of rare stimuli (P300 data) and movement prediction (MRCP data) show that these approaches are feasible and can significantly increase the performance of signal processing chains for brain-computer interfaces by 1% to 6%.

Memorization is worst-case generalization. Based on MacKay's information theoretic model of supervised machine learning, this article discusses how to practically estimate the maximum size of a neural network given a training data set. First, we present four easily applicable rules to analytically determine the capacity of neural network architectures. This allows the comparison of the efficiency of different network architectures independently of a task. Second, we introduce and experimentally validate a heuristic method to estimate the neural network capacity requirement for a given dataset and labeling. This allows an estimate of the required size of a neural network for a given problem. We conclude the article with a discussion on the consequences of sizing the network wrongly, which includes both increased computation effort for training as well as reduced generalization capability.

The classification of complex data usually requires the composition of processing steps. Here, a major challenge is the selection of optimal algorithms for preprocessing and classification (including parameterizations). Nowadays, parts of the optimization process are automized but expert knowledge and manual work are still required. We present three steps to face this process and ease the optimization. Namely, we take a theoretical view on classical classifiers, provide an approach to interpret the classifier together with the preprocessing, and integrate both into one framework which enables a semiautomatic optimization of the processing chain and which interfaces numerous algorithms.

Multi-context model learning is crucial for marine robotics where several factors can cause disturbances to the system's dynamics. This work addresses the problem of identifying multiple contexts of an AUV model. We build a simulation model of the robot from experimental data, and use it to fill in the missing data and generate different model contexts. We implement an architecture based on long-short-term-memory (LSTM) networks to learn the different contexts directly from the data. We show that the LSTM network can achieve high classification accuracy compared to baseline methods, showing robustness against noise and scaling efficiently on large datasets.

The development of technologies for the treatment of movement disorders, like stroke, is still of particular interest in brain-computer interface (BCI) research. In this context, source localization methods (SLMs), that reconstruct the cerebral origin of brain activity measured outside the head, e.g., via electroencephalography (EEG), can add a valuable insight into the current state and progress of the treatment. However, in BCIs SLMs were often solely considered as advanced signal processing methods that are compared against other methods based on the classification performance alone. Though, this approach does not guarantee physiological meaningful results. We present an empirical comparison of three established distributed SLMs with the aim to use one for single-trial movement prediction. The SLMs wMNE, sLORETA, and dSPM were applied on data acquired from eight subjects performing voluntary arm movements. Besides the classification performance as quality measure, a distance metric was used to asses the physiological plausibility of the methods. For the distance metric, which is usually measured to the source position of maximum activity, we further propose a variant based on clusters that is better suited for the single-trial case in which several sources are likely and the actual maximum is unknown. The two metrics showed different results. The classification performance revealed no significant differences across subjects, indicating that all three methods are equally well-suited for single-trial movement prediction. On the other hand, we obtained significant differences in the distance measure, favoring wMNE even after correcting the distance with the number of reconstructed clusters. Further, distance results were inconsistent with the traditional method using the maximum, indicating that for wMNE the point of maximum source activity often did not coincide with the nearest activation cluster. In summary, the presented comparison might help users to select an appropriate SLM and to understand the implications of the selection. The proposed methodology pays attention to the particular properties of distributed SLMs and can serve as a framework for further comparisons.

MOTIVATION:For deep learning on image data, a common approach is to augment the training data by artificial new images, using techniques like moving windows, scaling, affine distortions, and elastic deformations. In contrast to image data, electroencephalographic (EEG) data suffers even more from the lack of sufficient training data.METHODS:We suggest and evaluate rotational distortions similar to affine/rotational distortions of images to generate augmented data.RESULTS:Our approach increases the performance of signal processing chains for EEG-based brain-computer interfaces when rotating only around y- and z-axis with an angle around ±18 degrees to generate new data.CONCLUSION:This shows that our processing efficient approach generates meaningful data and encourages to look for further new methods for EEG data augmentation.

The work presented in this paper describes the use and evaluation of machine learning techniques like neural networks and support vector regression to learn a model of magnetic field distortions often induced in inertial measurement units using magnetometers by changing currents, postures or configurations of a robotic system. Such a model is needed in order to compensate the local dynamic distortions, especially for complex and confined robotic systems, and to achieve more robust and accurate ambient magnetic field measurements. This is crucial for a wide variety of autonomous navigation purposes from simple heading estimation over standard SLAM approaches to sophisticated magnetic field based localization techniques. The approach was evaluated in a laboratory setup and with a complex robotic system in an outdoor environment.