Sign upSign inSign upSign inAnurag SinghFollow–ListenShareIn the real world, new conditions and changing scenarios often differ from training data, causing current ML models to fail. Let’s explore this with a hypothetical example.Consider a self-driving AI trained on data from various countries and driving conditions, now deployed in different vehicles with distinct regulations. Should this AI make the same decision in identical situations across all these vehicles?The AI must adapt based on the vehicle type. Private cars carry passengers, so the AI should prioritize their safety. In contrast, commercial vehicles carry payloads, and the AI might need to protect humans outside the vehicle in an accident.Such context-dependent scenarios are common in the real world. Current machine learning models struggle to adapt because developers often “hard-code” notions of generalisation, leading to AI misalignment.Imagine this: It’s your best friend’s birthday, and you’ve decided to surprise them with a gift. But there’s a catch — you have no idea what they want. You could buy something random and hope for the best, but chances are, your gift might not hit the mark. This dilemma mirrors a common challenge in machine learning: training models with fixed generalization notions often leads to alignment problems.So, what’s the solution? One might suggest giving users the raw training data to train a model themselves. This is akin to handing over cash instead of a gift, allowing them to get exactly what they want. While this approach ensures alignment, it also places a significant burden on the user and can raise privacy concerns — much like the awkwardness of giving money instead of a thoughtful present.But what if there was a middle ground? Here’s where our approach comes in. We propose that model developers train a collection of models, each tailored to different potential needs, and then let the user choose the one that best fits their requirements at test time. Think of it as presenting a selection of carefully chosen gifts, allowing your friend to pick the one they love the most. This is what our approach, Imprecise Learning executes!Sure, this method is more computationally expensive — similar to buying multiple gifts instead of just one. However, it offers the best of both worlds: solving the alignment problem while avoiding the need to release the training data. By providing a range of options, we respect user privacy and make the process more efficient and user-friendly.Broadly speaking, there are two main types of generalization frameworks: in-distribution (IID) and out-of-distribution (OOD).IID: Train and test environments have the same distribution.OOD: Train and test environments can come from different distributions.In prior frameworks IID and OOD are reduced to precise generalisation: When training under IID, developers assume a fixed distribution and optimize for it. Making the learning problem precise, i.e. precise objective to optimise is known. In OOD, developers often optimize for the worst-case scenario, without knowing the test-time distribution. This also makes the problem precise.However, this precise generalization can lead to misalignment due to the rigid commitment to a specific distribution. However, imprecise learning takes a different approach.Imprecise Learning:By using imprecise learning, we can better handle the uncertainty and variability of real-world applications, creating models that are more flexible and aligned with diverse user needs.To learn further about our research you can read our ICML 2024 [paper] or have a chat with me or my colleagues at the Rational Intelligence Lab who made it possible, Siu Lun Chau, Shahine Bouabid] and Krikamol Muandet—-HelpStatusAboutCareersPressBlogPrivacyRulesTermsText to speech