In recent years, artificial intelligence technology has become increasingly mature. Nowadays, many products and services rely on artificial intelligence technology to achieve automation and intelligence, so it is closely related to our daily lives. Whether it’s the home devices that bring us various conveniences or the way the products we use all the time are made, the impact of artificial intelligence is everywhere, driving innovation in almost every aspect of our lives. But there are still shortcomings that can lead to frustration for end users and significant challenges for researchers trying to improve the performance of artificial intelligence technology.
common sense approach
Before his death in 2018, Microsoft co-founder Paul Allen devoted a lot of time and resources to solving a seemingly endless and huge challenge: the lack of basic common sense about artificial intelligence technology. Mr. Allen's Allen Institute for Artificial Intelligence (AI2) launched the Mosaic project to continue solving this problem. His conceptionis this: “In the early days of artificial intelligence research, people paid great attention to common sense, but this work has stagnated. Artificial intelligence still There is a lack of common sense that most 10-year-olds can grasp. We hope to make major breakthroughs in this area starting from this research." Allen's metaphor highlights a major problem with current deep learning technology. As smart as our AI products are often, they still can't answer the extremely simple questions we might ask our colleagues or partners. For example, “If I paint this wall red, will it still be red tomorrow?” To illustrate the extent to which we are going to solve this problem, AI2 CEO Oren Etzioni cites Here’s an example: “While Google’s artificial intelligence program AlphaGo defeated the world’s number one Go player in 2016, it had no idea that Go was a board game.” I think we can all agree that this is A very important detail, and if we cannot solve this problem, the potential for AI success will be limited to narrow application areas.
Complex solutions to common sense problems
Clearly, commonsense AI requires a multi-pronged strategy to overcome its limitations. To this end, Allen's Mosaic project "integrates machine reading and reasoning, natural language understanding, computer vision and crowdsourcing technologies to establish a new source of broad basic common sense knowledge for future artificial intelligence systems." For AI2 like this Organizationally speaking, what does this look like at a research level?
- Visual Common-Sense Reasoning (VCR) is a new task and large-scale dataset for cognitive-level visual understanding. This research focuses on creating high-order cognition and common sense reasoning for artificial intelligence-based vision systems. VCR is the result of a joint effort between researchers at the University of Washington and AI2. VCR utilized a team of crowdsourced workers to annotate data for the project.
- Common sense knowledge graph provides a semi-structured way to represent common sense concepts. This structure provides a different perspective than other sources of knowledge, but what type of knowledge is represented and how to ideally incorporate it into modern neural methods remains an important question facing research in this area. To address this, the teamis currently building and publishing resourcesto explore aspects of common sense, such as those about social conditions, mental states, and causality information.
- It is a large-scale data set that enables common sense reasoning, unified natural language reasoning, and physics-based reasoning. The dataset includes 113,000 multiple-choice questions about scenarios. Each question is a video caption from the Large-scale Movie Description and Understanding Challenge (LSMDC) or the large-scale dense event description database ActivityNet Captions, with four answers to choose from, aiming to determine what will happen next in the scene in question. . The correct answer is the (real) video subtitle of the next event in the video; the three incorrect answers are adversarially generated and verified by humans in order to confuse non-human machines. The team’s goal is to make SWAG a benchmark for the evaluation of common sense-based NLI and learning representations.