Pointer Value Retrieval: A new benchmark for understanding the limits of neural network generalization

Abstract

Central to the success of artificial neural networks is their ability to generalize. But does neural network generalization primarily rely on seeing highly similar training examples (memorization)? Or are neural networks capable of human-intelligence styled reasoning, and if so, to what extent? These remain fundamental open questions on artificial neural networks. In this paper, as steps towards answering these questions, we introduce a new benchmark, Pointer Value Retrieval (PVR) to study the limits of neural network reasoning. The PVR suite of tasks is based on reasoning about indirection, a hallmark of human intelligence, where a first stage (task) contains instructions for solving a second stage (task). In PVR, this is done by having one part of the task input act as a pointer, giving instructions on a different input location, which forms the output. We show this simple rule can be applied to create a diverse set of tasks across different input modalities and configurations. Importantly, this use of indirection enables systematically varying task difficulty through distribution shifts and increasing functional complexity. We conduct a detailed empirical study of different PVR tasks, discovering large variations in performance across dataset sizes, neural network architectures and task complexity. Further, by incorporating distribution shift and increased functional complexity, we develop nuanced tests for reasoning, revealing subtle failures and surprising successes, suggesting many promising directions of exploration on this benchmark.

References in corpus (19)

• Learning Transferable Visual Models From Natural Language Supervision
• Prototypical Networks for Few-shot Learning
• Language Models are Few-Shot Learners
• MLP-Mixer: An all-MLP Architecture for Vision
• Understanding deep learning requires rethinking generalization
• Large Batch Optimization for Deep Learning: Training BERT in 76 minutes
• Measuring abstract reasoning in neural networks
• Zero-Shot Task Generalization with Multi-Task Deep Reinforcement Learning
• What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation
• Making Neural Programming Architectures Generalize via Recursion
• Analysing Mathematical Reasoning Abilities of Neural Models
• A Benchmark for Systematic Generalization in Grounded Language Understanding
• Measuring Compositional Generalization: A Comprehensive Method on Realistic Data
• Generalization through Memorization: Nearest Neighbor Language Models
• Learning Compositional Rules via Neural Program Synthesis
• Bongard-LOGO: A New Benchmark for Human-Level Concept Learning and Reasoning
• Compositional Processing Emerges in Neural Networks Solving Math Problems
• Can Deep Reinforcement Learning Solve Erdos-Selfridge-Spencer Games?
• Solving Raven's Progressive Matrices with Neural Networks