The Present and Future AI Technologies : Development and Applications in Image/Video Recognition and Natural Language Processing

Hirokatsu KataokaJournal of the Japan Society for Precision Engineering 2021 · DOI

Abstract

As of 2020, the field of Artificial Intelligence (AI) and Computer Vision (CV) has entered a phase of rapid transformation. This paper surveys the shift from a singular focus on raw accuracy toward a multi-faceted approach prioritizing Fairness, Accountability, Transparency, and Ethics (FATE), computational efficiency, and self-supervised learning. We examine key developments such as AutoML, EfficientNet, and SimCLR, which collectively aim to reduce the reliance on massive human-labeled datasets and manual architecture design.

Introduction: The Evolution of AI

The trajectory of AI has been marked by distinct waves of innovation, moving from early perceptrons to the deep learning revolution that defines the current era. As the volume of research explodes—with thousands of papers accepted annually at major conferences—the field has transitioned into what is often termed the "3rd AI" wave, characterized by deep learning and its subsequent refinements.

A timeline illustrating the evolution of AI from the 1960s to 2020, highlighting the transition from early perceptrons to modern deep learning.

However, this rapid progress has surfaced three critical challenges: the Ethical Gap (bias in datasets), the Efficiency Gap (high computational costs), and the Labeling Gap (the bottleneck of manual data annotation).

Addressing the Three Gaps

The Ethical Gap (FATE)

Fairness and transparency are no longer peripheral concerns but technical requirements. Research such as the Gender Shades study has exposed significant racial and gender biases in commercial facial recognition APIs. To combat this, the community is moving toward standardized documentation like "DataSheets for Datasets" to ensure accountability in how data is sourced and used. This is particularly urgent following the withdrawal of datasets like "80M Tiny Images" due to offensive content.

The Efficiency Gap

The manual design of deep neural networks is both time-consuming and computationally expensive. Innovations like AutoML and Neural Architecture Search (NAS) automate this process, using AI to optimize its own structure. EfficientNet represents a landmark in this area, utilizing a compound scaling method to balance depth, width, and resolution. EfficientNet-B7 achieved state-of-the-art accuracy while being 8.4x smaller and 6.1x faster during inference than existing top-tier models.

The Labeling Gap: Self-Supervised Learning

The need for millions of manually labeled images has long been a bottleneck. Self-Supervised Learning (SSL) addresses this by training models on "pretext tasks"—synthetic challenges like predicting image rotation or solving jigsaw puzzles—where the "labels" are derived from the data itself.

The workflow of self-supervised learning, showing the transition from a pretext task (learning features from unlabeled data) to a target task (fine-tuning for specific classification).

A major breakthrough in this domain is SimCLR, a contrastive learning framework. By teaching a model that different augmentations of the same image should have similar representations, SimCLR allows models to learn robust features from unlabeled data. When scaled, these models can rival fully supervised systems using only a small fraction (e.g., 1% to 10%) of labeled data.

Key Insights and Future Directions

  • Ethics as a Technical Requirement: Fairness is a technical challenge requiring new data collection and auditing methods.
  • The End of Manual Feature Engineering: Just as deep learning replaced manual feature design, AutoML is beginning to replace manual architecture design.
  • Video as the Next Frontier: Research is moving from simple classification to "Temporal Action Segmentation" (identifying exact start/end times) and "Video Captioning." The Kinetics dataset, with over 650,000 clips, provides the necessary foundation for this progress.

While AI has excelled in general object recognition, applying it to specific fields like autonomous driving or medical diagnosis requires navigating complex social and technical hurdles. The convergence of CV and Natural Language Processing (NLP) remains a work in progress, aiming for a seamless integration that understands both visual context and linguistic grammar.

Summary updated

BibTeX
@article{kataoka2021present,
  author = {Hirokatsu Kataoka},
  title = {{The Present and Future AI Technologies : Development and Applications in Image/Video Recognition and Natural Language Processing}},
  journal = {Journal of the Japan Society for Precision Engineering},
  year = {2021},
  doi = {10.2493/jjspe.87.16},
}
Figure