Self-supervised learning (SSL) foundation models have emerged as powerful, domain-agnostic, general-purpose feature extractors applicable to a wide range of tasks. Such models pre-trained on human speech have demonstrated high transferability for bioacoustic processing. This paper investigates (i) whether SSL models pre-trained directly on animal vocalizations offer a significant advantage over those pre-trained on speech, and (ii) whether fine-tuning speech-pretrained models on automatic speech recognition (ASR) tasks can enhance bioacoustic classification. We conduct a comparative analysis using three diverse bioacoustic datasets and two different bioacoustic tasks. Results indicate that pre-training on bioacoustic data provides only marginal improvements over speech-pretrained models, with comparable performance in most scenarios. Fine-tuning on ASR tasks yields mixed outcomes, suggesting that the general-purpose representations learned during SSL pre-training are already well-suited for bioacoustic tasks. These findings highlight the robustness of speech-pretrained SSL models for bioacoustics and imply that extensive fine-tuning may not be necessary for optimal performance.

Computer-based modelling and simulation have become useful tools to facilitate humans to understand systems in different domains, such as physics, astrophysics, chemistry, biology, economics, engineering and social science. A complex system is featured with a large number of interacting components (agents, processes, etc.), whose aggregate activities are nonlinear and self-organized. Complex systems are hard to be simulated or modelled by using traditional computational approaches due to complex relationships among system components, distributed features of resources, and dynamics of environments. Meanwhile, smart systems such as multi-agent systems have demonstrated advantages and great potentials in modelling and simulating complex systems.

The Deep Learning on Supercomputers workshop is with ISC'20 on June 25th, 2020 in Frankfurt, Germany. It is the fourth workshop in the Deep Learning on Supercomputers series. The workshop provides a forum for practitioners working on any and all aspects of DL for scientific research in the High Performance Computing (HPC) context to present their latest research results and development, deployment, and application experiences. The general theme of this workshop series is the intersection of DL and HPC, while the theme of this particular workshop is centered around the applications of deep learning methods in scientific research: novel uses of deep learning methods, e.g., convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial network (GAN), and reinforcement learning (RL), for both natural and social science research, and innovative applications of deep learning in traditional numerical simulation. Its scope encompasses application development in scientific scenarios using HPC platforms; DL methods applied to numerical simulation; fundamental algorithms, enhanced procedures, and software development methods to enable scalable training and inference; hardware changes with impact on future supercomputer design; and machine deployment, performance evaluation, and reproducibility practices for DL applications with an emphasis on scientific usage. Topics include but are not limited to: - DL as a novel approach of scientific computing - Emerging scientific applications driven by DL methods - Novel interactions between DL and traditional numerical simulation - Effectiveness and limitations of DL methods in scientific research - Algorithms and procedures to enhance reproducibility of scientific DL applications - DL for science workflows - Data management through the life cycle of scientific DL applications - General algorithms and procedures for efficient and scalable DL training - Scalable DL methods to address the challenges of demanding scientific applications - General algorithms and systems for large scale model serving for scientific use cases - New software, and enhancements to existing software, for scalable DL - DL communication optimization at scale - I/O optimization for DL at scale - DL performance evaluation and analysis on deployed systems - DL performance modeling and tuning of DL on supercomputers - DL benchmarks on supercomputers - Novel hardware designs for more efficient DL - Processors, accelerators, memory hierarchy, interconnect changes with impact on deep learning in the HPC context As part of the reproducibility initiative, the workshop requires authors to provide information such as the algorithms, software releases, datasets, and hardware configurations used.

Dialogue response selection is an important part of Task-oriented Dialogue Systems (TDSs); it aims to predict an appropriate response given a dialogue context. Obtaining key information from a complex, long dialogue context is challenging, especially when different sources of information are available, e.g., the user's utterances, the system's responses, and results retrieved from a knowledge base (KB). Previous work ignores the type of information source and merges sources for response selection. However, accounting for the source type may lead to remarkable differences in the quality of response selection. We propose the Source-aware Recurrent Entity Network (SEntNet), which is aware of different information sources for the response selection process. SEntNet achieves this by employing source-specific memories to exploit differences in the usage of words and syntactic structure from different information sources (user, system, and KB). Experimental results show that SEntNet obtains 91.0% accuracy on the Dialog bAbI dataset, outperforming prior work by 4.7%. On the DSTC2 dataset, SEntNet obtains an accuracy of 41.2%, beating source unaware recurrent entity networks by 2.4%.

Background Based Conversations (BBCs) have been developed to make dialogue systems generate more informative and natural responses by leveraging background knowledge. Existing methods for BBCs can be grouped into two categories: extraction-based methods and generation-based methods. The former extract spans frombackground material as responses that are not necessarily natural. The latter generate responses thatare natural but not necessarily effective in leveraging background knowledge. In this paper, we focus on generation-based methods and propose a model, namely Context-aware Knowledge Pre-selection (CaKe), which introduces a pre-selection process that uses dynamic bi-directional attention to improve knowledge selection by using the utterance history context as prior information to select the most relevant background material. Experimental results show that our model is superior to current state-of-the-art baselines, indicating that it benefits from the pre-selection process, thus improving in-formativeness and fluency.