Researchers at MIT and IBM Research have created a method that enables a user to rank the results of a machine-learning model’s behavior. Their technique, called Shared Interest, incorporates metrics that compare how well a model’s thinking matches people’s. “Today, AI is capable of reaching (and, in some cases, exceeding) human performance in specific tasks, including image recognition and language understanding,” Pieter Buteneers, director of engineering in machine learning and AI at the communications company Sinch, told Lifewire in an email interview. “With natural language processing (NLP), AI systems can interpret, write and speak languages as well as humans, and the AI can even adjust its dialect and tone to align with its human peers.”
Artificial Smarts
AI often produces results without explaining why those decisions are correct. And tools that help experts make sense of a model’s reasoning often only provide insights, only one example at a time. AI is usually trained using millions of data inputs, making it hard for a human to evaluate enough decisions to identify patterns. In a recent paper, the researchers said that Shared Interest could help a user uncover trends in a model’s decision-making. And these insights could allow the user to decide whether a model is ready to be deployed. “In developing Shared Interest, our goal is to be able to scale up this analysis process so that you could understand on a more global level what your model’s behavior is,” Angie Boggust, a co-author of the paper, said in the news release. Shared Interest uses a technique that shows how a machine-learning model made a particular decision, known as saliency methods. If the model is classifying images, saliency methods highlight areas of an image that are important to the model when it makes its decision. Shared Interest works by comparing saliency methods to human-generated annotations. Researchers used Shared Interest to help a dermatologist determine if he should trust a machine-learning model designed to help diagnose cancer from photos of skin lesions. Shared Interest enabled the dermatologist to quickly see examples of the model’s correct and incorrect predictions. The dermatologist decided he could not trust the model because it made too many predictions based on image artifacts rather than actual lesions. “The value here is that using Shared Interest, we are able to see these patterns emerge in our model’s behavior. In about half an hour, the dermatologist was able to decide whether or not to trust the model and whether or not to deploy it,” Boggust said.
Measuring Progress
The work by MIT researchers could be a significant step forward for AI’s progress toward human-level intelligence, Ben Hagag, head of research at Darrow, a company that uses machine learning algorithms, said that told Lifewire in an email interview. “The reasoning behind a model’s decision is important to both machine learning researcher and the decision-maker,” Hagag said. “The former wants to understand how good the model is and how it can be improved, whereas the latter wants to develop a sense of confidence in the model, so they need to understand why that output was predicted.” But Hagag cautioned that the MIT research is based on the assumption that we understand or can annotate human understanding or human reasoning. “However, there is a possibility that this might not be accurate, so more work on understanding human decision-making is necessary,” Hagag added. Advances in AI could speed up the development of computers’ ability to understand language and revolutionize the way AI and humans interact, Buteneers said. Chatbots can understand hundreds of languages at a time, and AI assistants can scan bodies of text for answers to questions or irregularities. “Some algorithms can even identify when messages are fraudulent, which can help businesses and consumers alike to weed out spam messages,” Buteneers added. But, said Buteneers, AI still makes some mistakes that humans never would. “While some worry that AI will replace human jobs, the reality is we’ll always need people working alongside AI bots to help keep them in check and keep these mistakes at bay while maintaining a human touch in business,” he added.
