Preface

Geoffrey Holmes, Tie-Yan LiuACML 2015

The 7th Asian Conference on Machine Learning (ACML 2015) represents a significant milestone in the evolution of the machine learning landscape within the Asia-Pacific region. As documented in the preface to the proceedings, the 2015 conference captured a pivotal transition: the movement of deep learning from experimental frameworks to a dominant industrial force. This volume serves as a "time capsule," reflecting the challenges of the "Data Explosion" and the community's shift toward scalable, high-capacity models.

Organizational Methodology and Rigor

To maintain high scientific standards, ACML 2015 employed a rigorous organizational methodology designed to bridge the gap between theoretical analysis and real-world applications. A key feature was the two-cycle submission process, which allowed for a "revise and resubmit" workflow. This enabled authors to address reviewer comments and refine their work between cycles, a practice aimed at elevating the final quality of accepted papers.

The review process was strictly double-blind, involving at least three reviewers and one meta-reviewer per submission. The selectivity of the conference was high; out of 96 valid submissions, only 28 were accepted, resulting in a competitive 29% acceptance rate. Unlike many larger conferences that utilize parallel sessions, ACML 2015 maintained a single-track program to foster a unified community dialogue among the participants.

The "Big Model" Challenge and Distributed Learning

A primary theme of the 2015 proceedings was the challenge of "Big Models." Researchers recognized that the hurdle was no longer just the volume of data ("Big Data") but the complexity and parameter count of the models themselves. Billions of data points required billions of parameters, necessitating new "bridging models" to coordinate distributed CPU and GPU cores effectively.

The preface highlights the Stale Synchronous Parallel (SSP) framework as a critical development for managing synchronization in distributed learning environments, such as the Petuum framework. These innovations were essential for scaling algorithms to meet the demands of modern industrial applications.

Deep Learning and the Perception-Inference Gap

In 2015, Deep Learning was undergoing a massive resurgence, moving beyond simple classification tasks toward an "end-to-end learning paradigm." This shift allowed for direct connections between sensory input, such as vision, and motor output, such as robot manipulation. However, the organizers and keynote speakers also acknowledged the limitations of this "young" field.

A recurring theme was the perception-inference gap. While neural networks excelled at perception, they were seen as limited in higher-level tasks like reasoning and planning. This led to a burgeoning interest in Bayesian Deep Learning, which sought to combine the perceptual strengths of neural networks with the probabilistic reasoning of graphical models.

Beyond neural networks, the proceedings emphasized the importance of moving from correlation to causation. Research in causal discovery aimed to understand cause-and-effect relationships within non-Gaussian and time-series data, a crucial step for making machine learning models more robust and interpretable.

Additionally, the conference explored the current state of semantic search. The authors noted that existing knowledge-base techniques represented only the "tip of the iceberg," with significant work remaining in natural language interpretation and attribute ranking to improve how machines understand human queries.

The Rise of the Chinese ML Ecosystem

The 2015 proceedings also documented the rapid expansion of machine learning research in China. This growth was fueled by major industrial players—including Huawei, Tencent, Baidu, and Alibaba—and a vibrant academic community. The inclusion of a dedicated workshop on "Machine Learning in China" underscored the region's increasing influence on the global stage and its role in driving the next generation of ML innovations.

Conclusion

The ACML 2015 proceedings provide a comprehensive overview of a field in flux. By balancing theoretical analysis with practical applications in distributed computing, causal inference, and probabilistic modeling, the conference established itself as a premier venue for high-quality research. The volume remains a testament to the community's efforts to navigate the transition from small-scale models to the massive, distributed systems that define modern artificial intelligence.

Summary updated

BibTeX
@inproceedings{holmes2015preface,
  author = {Geoffrey Holmes and Tie-Yan Liu},
  title = {{Preface}},
  booktitle = {ACML},
  year = {2015},
}
Figure