The Association for the Advancement of Artificial Intelligence’s annual Conference on Artificial Intelligence (AAAI) received about 9,000 paper submissions this year, requiring a relatively large program committee with two program chairs and four associate program chairs.
One of the associate program directors is Kai-Wei Chang, an associate professor of computer science at the University of California, Los Angeles, and an Amazon Visiting Academic in the Alexa AI organization. This year, Chang was also named a senior member of the Association for the Advancement of Artificial Intelligence—and he chairs the AAAI conference’s Best Papers Committee. So he has an unusually good vantage point on trends in this year’s AAAI entries.
With more than 1,600 accepted papers, the AAAI program naturally spans a wide range of topics. “There are papers from all different fields—computer vision, natural language processing, neural networks,” Chang says. “Robotics have a big part of the conference, and there are papers in traditional areas, like search and planning.”
Still, Chang says, two topics stand out to him: rapid construction and reasoning.
Fast engineering
“Prompt engineering” refers to efforts to extract accurate, consistent, and fair outputs from large generative models, such as text-to-image synthesizers or large language models (LLMs). LLMs are trained in large texts, so they encode a lot of factual information about the world. But they are trained to produce sequences of words that are likely in the general case—not accurate in the particular case.
“For example, I asked a model to generate a biosketch of me, and it actually generated something that’s pretty good,” Chang says. “Maybe the model was trained on my website, but it said that I’m a professor at UCLA and that I do NLP research and submit papers to conferences like ACL, which is all true. But it also gave some random facts—for example, that I won a certain award, which I didn’t.”
“It is important for these models to have some form of fact-checking to filter out content that is inappropriate. There are several AAAI articles on how to ensure that the texts generated are personal, reliable and consistent.”
At Amazon, one of the topics Chang is researching is LLMs’ justice. Again, because LLMs’ output is based on statistical averages, it can reinforce stereotypes prevalent in the models’ training data. For example, if an LLM receives an input (a prompt) that mentions a doctor, it may default to using male pronouns to refer to that doctor in its generated output.
“Similar observations occur in text-to-image generation,” adds Chang. “If you ask the model to generate a doctor, it will probably generate a male doctor. We find that you can fix this by providing a description along with the prompt – like ‘all individuals can be lawyers regardless of their gender and skin color.’ Alternatively, you can improve the diversity of the generation by adding more diverse training data.”
Reasoning
Reasoning involves making inferences about the logical relationships between entities or concepts to perform tasks more complicated than the type of classification that machine learning models currently excel at. Many researchers believe that this will necessarily involve symbolic reasoning—an approach to AI that machine learning for years seemed to replace.
“You can define a loss function or a layer of a neural network called a semantic probabilistic layer to enable the model to learn to use symbolic knowledge for reasoning,” explains Chang. “For example, you can define some rules and define a loss based on how likely the model’s prediction is to violate those rules. Then you can train the model by minimizing the loss to avoid rule violations.”
“For example, for language generation, you can say, ‘I want to generate a sentence, and it must contain certain concepts or certain words’ – or conversely, that it can’t contain any bad words. The constraints can also be ‘soft’. For example, if you’re doing robot planning, you can have a constraint that says the robot shouldn’t go into a certain area unless it’s necessary. So it’s not necessary. cannot enter the region, but the model is trained to avoid it.”
In fact, Chang says, he has also worked on just such an approach, where a second, supplementary network helps guide the primary model toward output that satisfies a set of constraints.
“You can train a supplementary neural network to help you break down complicated constraints into smaller pieces so it’s easier to incorporate into the model,” Chang explains. “So in the language generation example, say you want to generate a story that must contain certain user-defined words, but also the sentiment of the story must be positive. These constraints are difficult to incorporate into text generation, as the generated output must be coherent, and the model may not know where to insert those words and keep the sentiment positive. The neural network can learn to break down these problems and the compatibility rules that address these issues. to guide the primary model.”
However, Chang emphasizes that while rapid engineering and reasoning are popular topics at this year’s AAAI, they still make up only a small fraction of the conference’s program. “AI is very popular today,” he says. “There are several subfields, such as machine learning, computer vision, NLP and robotics. And there are quite diverse submissions from all these different fields.”