skip to main content


Title: Abstract representations emerge naturally in neural networks trained to perform multiple tasks
Abstract

Humans and other animals demonstrate a remarkable ability to generalize knowledge across distinct contexts and objects during natural behavior. We posit that this ability to generalize arises from a specific representational geometry, that we call abstract and that is referred to as disentangled in machine learning. These abstract representations have been observed in recent neurophysiological studies. However, it is unknown how they emerge. Here, using feedforward neural networks, we demonstrate that the learning of multiple tasks causes abstract representations to emerge, using both supervised and reinforcement learning. We show that these abstract representations enable few-sample learning and reliable generalization on novel tasks. We conclude that abstract representations of sensory and cognitive variables may emerge from the multiple behaviors that animals exhibit in the natural world, and, as a consequence, could be pervasive in high-level brain regions. We also make several specific predictions about which variables will be represented abstractly.

 
more » « less
Award ID(s):
1707398
NSF-PAR ID:
10398600
Author(s) / Creator(s):
;
Publisher / Repository:
Nature Publishing Group
Date Published:
Journal Name:
Nature Communications
Volume:
14
Issue:
1
ISSN:
2041-1723
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. 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.

     
    more » « less
  2. Abstract

    Advances in artificial intelligence have raised a basic question about human intelligence: Is human reasoning best emulated by applying task‐specific knowledge acquired from a wealth of prior experience, or is it based on the domain‐general manipulation and comparison of mental representations? We address this question for the case of visual analogical reasoning. Using realistic images of familiar three‐dimensional objects (cars and their parts), we systematically manipulated viewpoints, part relations, and entity properties in visual analogy problems. We compared human performance to that of two recent deep learning models (Siamese Network and Relation Network) that were directly trained to solve these problems and to apply their task‐specific knowledge to analogical reasoning. We also developed a new model using part‐based comparison (PCM) by applying a domain‐general mapping procedure to learned representations of cars and their component parts. Across four‐term analogies (Experiment 1) and open‐ended analogies (Experiment 2), the domain‐general PCM model, but not the task‐specific deep learning models, generated performance similar in key aspects to that of human reasoners. These findings provide evidence that human‐like analogical reasoning is unlikely to be achieved by applying deep learning with big data to a specific type of analogy problem. Rather, humans do (and machines might) achieve analogical reasoning by learning representations that encode structural information useful for multiple tasks, coupled with efficient computation of relational similarity.

     
    more » « less
  3. Learning from multiple representations (MRs) is not an easy task for most people, despite how easy it is for experts. Different combinations of representations (e.g., text + photograph, graph + formula, map + diagram) pose different challenges for learners, but across the literature researchers find these to be challenging learning tasks. Each representation typically includes some unique information, as well as some information shared with the other representation(s). Finding one piece of information is only somewhat challenging, but linking information across representations and especially making inferences are very challenging and important parts of using multiple representations for learning. Coordination of multiple representations skills are rarely taught in classrooms, despite the fact that learners are frequently tested on them. Learning from MRs depends on the specific learning tasks posed, learner characteristics, the specifics of which representation(s) are used, and the design of each representation. These various factors act separately and in combination (which can be compensatory, additive, or interactive). Learning tasks can be differentially effective depending on learner characteristics, especially prior knowledge, self-regulation, and age/grade. Learning tasks should be designed keeping this differential effectiveness in mind, and researchers should test for such interactions. 
    more » « less
  4. This paper presents a domain-guided approach for learning representations of scalp-electroencephalograms (EEGs) without relying on expert annotations. Expert labeling of EEGs has proven to be an unscalable process with low inter-reviewer agreement because of the complex and lengthy nature of EEG recordings. Hence, there is a need for machine learning (ML) approaches that can leverage expert domain knowledge without incurring the cost of labor-intensive annotations. Self-supervised learning (SSL) has shown promise in such settings, although existing SSL efforts on EEG data do not fully exploit EEG domain knowledge. Furthermore, it is unclear to what extent SSL models generalize to unseen tasks and datasets. Here we explore whether SSL tasks derived in a domain-guided fashion can learn generalizable EEG representations. Our contributions are three-fold: 1) we propose novel SSL tasks for EEG based on the spatial similarity of brain activity, underlying behavioral states, and age-related differences; 2) we present evidence that an encoder pretrained using the proposed SSL tasks shows strong predictive performance on multiple downstream classifications; and 3) using two large EEG datasets, we show that our encoder generalizes well to multiple EEG datasets during downstream evaluations. 
    more » « less
  5. Graph few-shot learning is of great importance among various graph learning tasks. Under the few-shot scenario, models are often required to conduct classification given limited labeled samples. Existing graph few-shot learning methods typically leverage Graph Neural Networks (GNNs) and perform classification across a series of meta-tasks. Nevertheless, these methods generally rely on the original graph (i.e., the graph that the meta-task is sampled from) to learn node representations. Consequently, the learned representations for the same nodes are identical in all meta-tasks. Since the class sets are different across meta-tasks, node representations should be task-specific to promote classification performance. Therefore, to adaptively learn node representations across meta-tasks, we propose a novel framework that learns a task-specific structure for each meta-task. To handle the variety of nodes across meta-tasks, we extract relevant nodes and learn task-specific structures based on node influence and mutual information. In this way, we can learn node representations with the task-specific structure tailored for each meta-task. We further conduct extensive experiments on five node classification datasets under both single- and multiple-graph settings to validate the superiority of our framework over the state-of-the-art baselines. 
    more » « less