Transforming task representations to perform novel tasks

Abstract

An important aspect of intelligence is the ability to adapt to a novel task without any direct experience (zero shot), based on its relationship to previous tasks. Humans can exhibit this cognitive flexibility. By contrast, models that achieve superhuman performance in specific tasks often fail to adapt to even slight task alterations. To address this, we propose a general computational framework for adapting to novel tasks based on their relationship to prior tasks. We begin by learning vector representations of tasks. To adapt to new tasks, we propose metamappings, higher-order tasks that transform basic task representations. We demonstrate the effectiveness of this framework across a wide variety of tasks and computational paradigms, ranging from regression to image classification and reinforcement learning. We compare to both human adaptability and language-based approaches to zero-shot learning. Across these domains, metamapping is successful, often achieving 80 to 90% performance, without any data, on a novel task, even when the new task directly contradicts prior experience. We further show that metamapping can not only generalize to new tasks via learned relationships, but can also generalize using novel relationships unseen during training. Finally, using metamapping as a starting point can dramatically accelerate later learning on a new task and reduce learning time and cumulative error substantially. Our results provide insight into a possible computational basis of intelligent adaptability and offer a possible framework for modeling cognitive flexibility and building more flexible artificial intelligence systems.

On the link between conscious function and general intelligence in humans and machines

“…This approach is consistent with a recent proposal to interpret the global workspace as a representational space of possible skills or tasks (a schema space) (VanRullen & Kanai, 2021). Related recent work has also provided an empirical instantiation of this concept by utilizing hypernetworks to generate task-specific networks (Lampinen & McClelland, 2020). We expect that such an system would be able to display rapid learning in the face of novel scenarios, although without the ability to internally generate experience as described by IGT, we expect it would still be orders of magnitude less sample efficient than an agent capable of mental time travel.…”

Two Computational Approaches to Visual Analogy: Task‐Specific Models Versus Domain‐General Mapping

“…Instead of learning perceptual representations that might be generally useful in multiple tasks, these models acquire representations tailored to the idiosyncrasies of the specific task used in training. A possible way to augment this task-based approach is to incorporate meta-learning, in which a model is trained to solve (and explicitly represent) multiple distinct tasks, and then transfers its acquired knowledge to solve novel tasks, based on their similarity to prior learned tasks (Lampinen & McClelland, 2020). However, it remains to be seen whether meta-learning across a range of analogy tasks can allow a task-based approach to account for the breadth of human analogical reasoning.…”

Abstraction and analogy‐making in artificial intelligence

“…However, at the time of this writing, metalearning approaches have not yet been explored for other abstraction and analogy domains. Another proposed approach is that of “metamapping” 91 that directly maps a representation of one task to a representation of a related task.…”