TL;DR
This paper introduces a novel connectionist model that integrates explicit instruction with experience-based learning. It proposes that direct instruction influences learning by dynamically shifting a network's activation state, rather than by modifying connection weights, allowing for rapid learning and complex interactions between instruction and induction.
What Problem Does This Paper Address?
Humans seamlessly blend knowledge gained from direct instruction with what they learn through experience. This paper tackles the challenge of computationally modeling this integration, specifically addressing how explicit instructions can rapidly influence learning and interact with inductive generalization from examples. Existing models often struggle to account for the speed and complexity of human learning in this regard, particularly the immediate impact of "learning by being told."
Methods
The authors propose a connectionist model that represents instructed knowledge as "articulated attractors" within the network's dynamic activation space. Direct instruction is modeled as input activity that nudges the network into a specific basin of attraction, representing the instructed knowledge. This approach decouples the rapid effects of instruction from the slower process of connection weight modification, which is handled by standard inductive learning techniques. This allows induction and instruction to operate concurrently and interact in complex ways.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Instruction as Dynamic State Shift: The core idea is that instruction doesn't immediately rewire the network but rather guides its current state, enabling rapid learning without slow weight updates.
- Attractors for Knowledge: Instructed knowledge is encoded as stable states (attractors) in the network's dynamics, providing a structured representation of what has been learned explicitly.
- Bridging Explicit and Implicit Learning: The model offers a unified framework for understanding how explicit rules and implicit experience can work together, potentially explaining phenomena like the primacy of instruction over examples.
- Explaining Human Learning Puzzles: The model is designed to account for observed human learning behaviors, such as the rapid improvement after instruction, the advantage of instruction before examples, and the sometimes-stubborn adherence to rules even when contradictory evidence exists.
Key Results
The paper reports that the model has demonstrated the ability to learn a simple instructional language and exhibit rapid instruction-following behavior. It is also being applied to experimental data from category learning and implicit/explicit learning literature, with comparisons planned against symbolic models like Instructo-SOAR.
Limitations and Future Work
The authors acknowledge that the model is still under development and is being applied to existing literature. Future work includes detailed comparisons with symbolic models and further experimental validation against human learning data. The paper suggests that this approach opens doors to understanding complex interactions between induction and instruction and could potentially lead to neurological explanations for learning.
Summary updated