Multi-label classification is an approach which allows a datapoint to be labelled with more than one class at the same time. A common but trivial approach is to train individual binary classifiers per label, but the performance can be improved by considering associations within the labels. Like with any machine learning algorithm, hyperparameter tuning is important to train a good multi-label classifier model. The task of selecting the best hyperparameter settings for an algorithm is an optimisation problem. Very limited work has been done on automatic hyperparameter tuning and AutoML in the multi-label domain. This paper attempts to fill this gap by proposing a neural network algorithm, CascadeML, to train multi-label neural network based on cascade neural networks. This method requires minimal or no hyperparameter tuning and also considers pairwise label associations. The cascade algorithm grows the network architecture incrementally in a two phase process as it learns the weights using adaptive first order gradient algorithm, therefore omitting the requirement of preselecting the number of hidden layers, nodes and the learning rate. The method was tested on 10 multi-label datasets and compared with other multi-label classification algorithms. Results show that CascadeML performs very well without hyperparameter tuning.

Memory is increasingly often the bottleneck when training neural network models. Despite this, techniques to lower the overall memory requirements of training have been less widely studied compared to the extensive literature on reducing the memory requirements of inference. In this paper we study a fundamental question: How much memory is actually needed to train a neural network? To answer this question, we profile the overall memory usage of training on two representative deep learning benchmarks -- the WideResNet model for image classification and the DynamicConv Transformer model for machine translation -- and comprehensively evaluate four standard techniques for reducing the training memory requirements: (1) imposing sparsity on the model, (2) using low precision, (3) microbatching, and (4) gradient checkpointing. We explore how each of these techniques in isolation affects both the peak memory usage of training and the quality of the end model, and explore the memory, accuracy, and computation tradeoffs incurred when combining these techniques. Using appropriate combinations of these techniques, we show that it is possible to the reduce the memory required to train a WideResNet-28-2 on CIFAR-10 by up to 60.7x with a 0.4% loss in accuracy, and reduce the memory required to train a DynamicConv model on IWSLT'14 German to English translation by up to 8.7x with a BLEU score drop of 0.15.

Efficient deep learning computing requires algorithm and hardware co-design to enable specialization: we usually need to change the algorithm to reduce memory footprint and improve energy efficiency. However, the extra degree of freedom from the algorithm makes the design space much larger: it's not only about designing the hardware but also about how to tweak the algorithm to best fit the hardware. Human engineers can hardly exhaust the design space by heuristics. It's labor consuming and sub-optimal. We propose design automation techniques for efficient neural networks. We investigate automatically designing specialized fast models, auto channel pruning, and auto mixed-precision quantization. We demonstrate such learning-based, automated design achieves superior performance and efficiency than rule-based human design. Moreover, we shorten the design cycle by 200x than previous work, so that we can afford to design specialized neural network models for different hardware platforms.

Large scale machine learning (ML) systems such as the Alexa automatic speech recognition (ASR) system continue to improve with increasing amounts of manually transcribed training data. Instead of scaling manual transcription to impractical levels, we utilize semi-supervised learning (SSL) to learn acoustic models (AM) from the vast firehose of untranscribed audio data. Learning an AM from 1 Million hours of audio presents unique ML and system design challenges. We present the design and evaluation of a highly scalable and resource efficient SSL system for AM. Employing the student/teacher learning paradigm, we focus on the student learning subsystem: a scalable and robust data pipeline that generates features and targets from raw audio, and an efficient model pipeline, including the distributed trainer, that builds a student model. Our evaluations show that, even without extensive hyper-parameter tuning, we obtain relative accuracy improvements in the 10 to 20$\%$ range, with higher gains in noisier conditions. The end-to-end processing time of this SSL system was 12 days, and several components in this system can trivially scale linearly with more compute resources.

We present an algorithm for learning a latent variable generative model via generative adversarial learning where the canonical uniform noise input is replaced by samples from a graphical model. This graphical model is learned by a Boltzmann machine which learns low-dimensional feature representation of data extracted by the discriminator. A quantum annealer, the D-Wave 2000Q, is used to sample from this model. This algorithm joins a growing family of algorithms that use a quantum annealing subroutine in deep learning, and provides a framework to test the advantages of quantum-assisted learning in GANs. Fully connected, symmetric bipartite and Chimera graph topologies are compared on a reduced stochastically binarized MNIST dataset, for both classical and quantum annealing sampling methods. The quantum-assisted associative adversarial network successfully learns a generative model of the MNIST dataset for all topologies, and is also applied to the LSUN dataset bedrooms class for the Chimera topology. Evaluated using the Fr\'{e}chet inception distance and inception score, the quantum and classical versions of the algorithm are found to have equivalent performance for learning an implicit generative model of the MNIST dataset.

Transformers are powerful sequence models, but require time and memory that grows quadratically with the sequence length. In this paper we introduce sparse factorizations of the attention matrix which reduce this to $O(n \sqrt{n})$. We also introduce a) a variation on architecture and initialization to train deeper networks, b) the recomputation of attention matrices to save memory, and c) fast attention kernels for training. We call networks with these changes Sparse Transformers, and show they can model sequences tens of thousands of timesteps long using hundreds of layers. We use the same architecture to model images, audio, and text from raw bytes, setting a new state of the art for density modeling of Enwik8, CIFAR-10, and ImageNet-64. We generate unconditional samples that demonstrate global coherence and great diversity, and show it is possible in principle to use self-attention to model sequences of length one million or more.

Multimodal learning has been lacking principled ways of combining information from different modalities and learning a low-dimensional manifold of meaningful representations. We study multimodal learning and sensor fusion from a latent variable perspective. We first present a regularized recurrent attention filter for sensor fusion. This algorithm can dynamically combine information from different types of sensors in a sequential decision making task. Each sensor is bonded with a modular neural network to maximize utility of its own information. A gating modular neural network dynamically generates a set of mixing weights for outputs from sensor networks by balancing utility of all sensors' information. We design a co-learning mechanism to encourage co-adaption and independent learning of each sensor at the same time, and propose a regularization based co-learning method. In the second part, we focus on recovering the manifold of latent representation. We propose a co-learning approach using probabilistic graphical model which imposes a structural prior on the generative model: multimodal variational RNN (MVRNN) model, and derive a variational lower bound for its objective functions. In the third part, we extend the siamese structure to sensor fusion for robust acoustic event detection. We perform experiments to investigate the latent representations that are extracted; works will be done in the following months. Our experiments show that the recurrent attention filter can dynamically combine different sensor inputs according to the information carried in the inputs. We consider MVRNN can identify latent representations that are useful for many downstream tasks such as speech synthesis, activity recognition, and control and planning. Both algorithms are general frameworks which can be applied to other tasks where different types of sensors are jointly used for decision making.

Power meters are becoming a widely used tool for measuring training and racing effort in cycling, and are now spreading also to other sports. This means that increasing volumes of data can be collected from athletes, with the aim of helping coaches and athletes analyse and understanding training load, racing efforts, technique etc. In this project, we have collaborated with Skisens AB, a company producing handles for cross country ski poles equipped with power meters. We have conducted a pilot study in the use of machine learning techniques on data from Skisens poles to identify which "gear" a skier is using (double poling or gears 2-4 in skating), based only on the sensor data from the ski poles. The dataset for this pilot study contained labelled time-series data from three individual skiers using four different gears recorded in varied locations and varied terrain. We systematically evaluated a number of machine learning techniques based on neural networks with best results obtained by a LSTM network (accuracy of 95% correctly classified strokes), when a subset of data from all three skiers was used for training. As expected, accuracy dropped to 78% when the model was trained on data from only two skiers and tested on the third. To achieve better generalisation to individuals not appearing in the training set more data is required, which is ongoing work.

In this paper, we propose a semi-supervised deep learning method for detecting the specific types of reads that impede the de novo genome assembly process. Instead of dealing directly with sequenced reads, we analyze their coverage graphs converted to 1D-signals. We noticed that specific signal patterns occur in each relevant class of reads. Semi-supervised approach is chosen because manually labelling the data is a very slow and tedious process, so our goal was to facilitate the assembly process with as little labeled data as possible. We tested two models to learn patterns in the coverage graphs: M1 M2 and semi-GAN. We evaluated the performance of each model based on a manually labeled dataset that comprises various reads from multiple reference genomes with respect to the number of labeled examples that were used during the training process. In addition, we embedded our detection in the assembly process which improved the quality of assemblies.

Complex-valued representations have demonstrated promising results on modeling relational data, i.e., knowledge graphs. This paper proposes a new knowledge graph embedding method. More concretely, we move beyond standard complex representations, adopting expressive hypercomplex representations for learning representations of entities and relations. Hypercomplex embeddings, or Quaternion embeddings (QuatE), are complex valued embeddings with three imaginary components. Different from standard complex (Hermitian) inner product, latent interdependencies (between all components) are aptly captured via the Hamilton product in Quaternion space, encouraging a more efficient and expressive representation learning process. Moreover, Quaternions are intuitively desirable for smooth and pure rotation in vector space, preventing noise from sheer/scaling operators. Finally, Quaternion inductive biases enjoy and satisfy the key desiderata of relational representation learning (i.e., modeling symmetry, anti-symmetry and inversion). Experimental results demonstrate that QuatE achieves state-of-the-art performance on four well-established knowledge graph completion benchmarks.