Artificial intelligence has experienced significant growth and development in recent years. As AI capabilities continue to improve at an exponential pace, researchers are exploring new applications of the technology and investigating how it will impact various aspects of society. This research paper provides an in-depth look at advancements in artificial intelligence from 2018 as presented in academic papers and published studies.
One of the major breakthroughs in 2018 was the creation of more powerful AI systems that are able to learn from vast amounts of unlabeled data in an unsupervised manner. Unsupervised learning allows AI to discover hidden patterns and insights without being explicitly programmed what to learn. An influential paper from OpenAI described an autoregressive language model called GPT-2 that was trained on a huge corpus of unlabeled web text. GPT-2 showed an ability to generate coherent multi-paragraph passages of text while avoiding biases present in its training data. This marked progress toward building AI systems that can learn in a more human-like fashion from raw experiences rather than relying solely on narrow, pre-defined tasks.
Other significant papers focused on self-supervised learning, a related area that involves learning representations of data without reliance on manual labelling. A team from Google Brain published research demonstrating how self-supervised methods allowed them to train state-of-the-art visual models using only unlabeled images from YouTube thumbnails. Separately, researchers from DeepMind presented a method called Skip-Thought Vectors which enabled learning useful and generalizable embeddings of sequences (such as sentences) without requiring explicit supervision during training. These embeddings could then be applied to downstream tasks like question answering, summarization and translation. Advancements in self-supervised and unsupervised learning are broadening AI capabilities while reducing the need for costly data annotation efforts.
2018 also saw further development of AI assistants like Amazon’s Alexa, Microsoft’s Cortana and Anthropic’s Claude which can carry out conversational interactions based on natural language. These systems still face limitations in terms of their ability to reason, explain their knowledge and adapt their responses based on context or prior conversations. To address this, numerous papers explored new techniques for endowing AI with deeper language understanding. A team from Salesforce Research applied self-supervised methods to learn multi-modal alignments between language, knowledge graphs and images which improved their conversational agent’s grounding capabilities. Researchers from Harvard proposed a method called ConveRT which pre-trained transformer-based language models on large corpora for use in goal-oriented dialog. Their model demonstrated more comprehensive responses compared to prior sequence-to-sequence models.
In the realm of computer vision, noteworthy progress was seen in few-shot learning – the ability of models to learn new concepts from just a handful of training examples. Researchers from DeepMind introduced a method called Meta-Learning for Fast Adaptation of Deep Networks which enabled training convolutional networks that could rapidly learn new classification tasks from very small datasets. Separately, researchers from UC Berkeley proposed a model called Propagation Networks which leveraged information across examples to generalize learning of new visual categories from limited data. These types of approaches hold promise for enabling computer vision systems that require less data to learn new image recognition capabilities.
Progress was also made in reinforcement learning, a subfield of machine learning concerned with agents that learn optimal behaviors through trial-and-error interactions with an environment. DeepMind researchers published a landmark paper describing AlphaZero, a revolutionary AI system that achieved superhuman performance at Go, chess and shogi with no prior domain knowledge. AlphaZero defeated the world’s strongest programs by learning via self-play reinforcement learning without human data or expertise, culminating months of uninterrupted self-play training. In another breakthrough, OpenAI Five defeated the best professional esports team in Dota 2 after training exclusively through self-play with no human demonstrations. This demonstrated the capability of reinforcement learning to master challenging multi-agent interactions and continuous high-dimensional control problems.
Applying deep reinforcement learning to the real-world remains difficult due to imperfect simulations and high sample complexity, but progress continued in 2018. Researchers from Google introduced a method called Sim-to-Real which enabled policies trained in simulation to generalize to the real world, allowing a robotic hand to complete challenging in-hand object manipulation tasks after learning solely through simulation. Scientists from DeepMind applied model-based reinforcement learning techniques called PlaNet to enable an agent to solve visual navigation tasks and control a Sawyer robotic arm in the real world. Separately, researchers from OpenAI introduced the methodology of constitutional AI which uses self-supervised auxiliary tasks to help agents learn control policies that transfer between simulations and reality with far fewer real-world interactions. These works moved reinforcement learning closer to solving practical robotics applications.
While AI made remarkable progress in various technical fronts in 2018, there remains open challenges to address. Limited understanding, bias and lack of transparency plague some self-supervised models that learn representations from vast datasets in an uninterpretable manner. Researchers are actively exploring techniques like model summarization and explanation generation to demystify complex AI systems. Additionally, safe and beneficial reinforcement learning poses unique difficulties due to the open-ended nature of interactions enabled by such systems. Addressing these design challenges will be crucial to ensuring AI safety and aligning the development of advanced autonomous agents with human values and priorities. Overall, 2018 saw abundant creativity and innovation in AI research, bringing closer the realization of intelligent systems that assist and empower humanity. Continued progress responsibly advancing the field promises to substantially benefit society.
In summary, 2018 was an extremely productive year for artificial intelligence research as evidenced by the numerous influential papers published across various subfields. Advances in unsupervised, self-supervised and reinforcement learning widened the applications of AI while reducing dependency on large human-annotated datasets. Advances in computer vision, natural language processing and robotics brought AI capabilities closer to solving real-world problems. While challenges remain regarding interpretability, robustness and safe development of advanced autonomous systems, steady progress is being made to ensure AI is developed and applied in a beneficial manner. Continued responsible research promises to fulfill the promise of artificial intelligence to improve lives at scale.
Andrew Stroud