Last week, Apple sponsored the largest speech technology conference Interspeech 2019. Being also a participant, the company demonstrated its new machine learning system called Overton.
As part of their work at the conference, the Apple team presented several research papers with a detailed analysis of speech processing technology. There were documents describing methods for determining a user’s intentions by the intonation of their voice, improving voice recognition, developing accurate tools for understanding the nuances of speech, and building relationship between users and voice assistants.
In the nearest future, the video by Apple should appear on the Interspeech YouTube channel and tell more about their achievements in the field of machine learning. The fact that Apple engineers and analysts interact with the scientific community is no longer surprising now. Since 2017 they occasionally publish articles about machine learning on their website.
Apple claims that Overton is the first of its kind to ensure that most of the personalization of ML models are controlled by the machine itself, and not by humans. When you ask Siri a question, only voice interaction occurs on the client side. Next, ML models try to understand the question, contextualise it, and find the most accurate answer.
Providing a high-quality answer is actually harder than it sounds. In response to many questions, Siri can only provide encyclopaedic data, for example, from Wikipedia. Although in the future, the assistant should become an effective source of answers for intricate problems. Perhaps, it will even learn to predict questions. But it will be very difficult to implement this feature.
How can data scientists be sure that Siri’s answer is the most accurate possible? This is the problem solved by Apple with the help of Overton, which is designed to ‘automate the life cycle of model construction, deployment, and monitoring’. The machine itself corrects the ML models in response to external stimuli, making them more accurate and correcting logical errors that could lead to incorrect conclusions.
The idea is that people will be able to focus on quality monitoring of ML models. When they need to make small but necessary adjustments, people will be able to request a set of changes that Overton will apply itself instead of delving into more complex code.
The company’s ambitions for Siri extend further than using it as a digital know-it-all for users to ask questions without a-hundred-percent certainty to get a useful answer. It is designed to become a real assistant providing high-level information, conducting contextual analysis, and expanding current tasks.
Siri Suggestions also point to this direction, although their implementation is still limited. Apple says:
A major direction of ongoing work is the systems that build on Overton to aid in managing data augmentation, programmatic supervision, and collaboration.
Most likely, the new product will affect user privacy. Apple engineers build models that run on iOS devices and are highly accurate in their opinion. Overton provides these models with some independence, while the ML system adjusts the models to achieve accuracy and relevance without the intervention of a researcher in each specific action. This means that data analysts will play a strategic role and will not possess information about each user.
Apple creates ML machines so that they could personalize the models they use. The company claims that Overton is the first ML management system that can monitor an application and improve its quality. Along with Siri’s improvement, this technology also helps iOS 13 devices detect pets when a user points the camera at pets.