Explanatory Learning: Beyond Empiricism in Neural Networks (2201.10222v1)

Abstract: We introduce Explanatory Learning (EL), a framework to let machines use existing knowledge buried in symbolic sequences -- e.g. explanations written in hieroglyphic -- by autonomously learning to interpret them. In EL, the burden of interpreting symbols is not left to humans or rigid human-coded compilers, as done in Program Synthesis. Rather, EL calls for a learned interpreter, built upon a limited collection of symbolic sequences paired with observations of several phenomena. This interpreter can be used to make predictions on a novel phenomenon given its explanation, and even to find that explanation using only a handful of observations, like human scientists do. We formulate the EL problem as a simple binary classification task, so that common end-to-end approaches aligned with the dominant empiricist view of machine learning could, in principle, solve it. To these models, we oppose Critical Rationalist Networks (CRNs), which instead embrace a rationalist view on the acquisition of knowledge. CRNs express several desired properties by construction, they are truly explainable, can adjust their processing at test-time for harder inferences, and can offer strong confidence guarantees on their predictions. As a final contribution, we introduce Odeen, a basic EL environment that simulates a small flatland-style universe full of phenomena to explain. Using Odeen as a testbed, we show how CRNs outperform empiricist end-to-end approaches of similar size and architecture (Transformers) in discovering explanations for novel phenomena.

• The paper introduces a novel Explanatory Learning framework enabling machines to interpret symbolic sequences in a human-like scientific process.
• It demonstrates that Critical Rationalist Networks in the Odeen environment outperform conventional approaches with improved Nearest Rule Score and Tagging Accuracy.
• The approach enhances intrinsic explainability and adaptive processing, paving the way for AI systems that mirror rigorous scientific inquiry.

An Overview of Explanatory Learning: Beyond Empiricism in Neural Networks

The paper introduces Explanatory Learning (EL), a novel framework conceived to enable machines to autonomously interpret symbolic sequences and utilize them as explanations, in a manner that emulates human scientific practice. This approach diverges from traditional program synthesis, where human-coded compilers are used to process symbols, by engaging a learned interpreter developed from a limited array of symbolic sequences paired with observations of various phenomena.

Key Contributions and Concepts

The authors present three main contributions:

1. Explanatory Learning Framework: The paper posits a problem framework where machines must learn to interpret languages comprising symbolic sequences to predict new phenomena. Unlike previous methodologies focusing on meta-learning or program synthesis, EL emphasizes data-driven learning of interpreters instead of instantiating them through hand-crafted logic expressions.
2. Odeen Environment: As a simulation of knowledge discovery, Odeen represents an environment with a constrained universe (reminiscent of flatland) populated by geometric figures. It serves as a testbed for EL approaches, simulating various phenomena to be explained and predicted.
3. Critical Rationalist Networks (CRNs): In alignment with the critical rationalist epistemological stance, CRNs are proposed, emphasizing conjectures that are either accepted or rejected through testing, rather than being continually modified. CRNs comprise a Conjecture Generator and a learned Interpreter, which work together to formulate and verify potential explanations for the observed phenomena.

Experimental Insights

The experiments conducted using the Odeen environment demonstrate that CRNs, while limited to similar size and architecture as conventional Transformers, outperform end-to-end empiricist approaches (referred to as EMP-C and EMP-R) in discovering explanations and producing accurate predictions for novel phenomena. The Odeen dataset's various training configurations reveal CRNs' superior capability in terms of Nearest Rule Score (NRS) and Tagging Accuracy (T-Acc), thus asserting their generalizable nature.

Theoretical and Practical Implications

The paper posits several theoretical implications of EL and CRNs:

• Improved Generalization: By detaching the conjectures from the network’s adjustable parameters, CRNs promote more robust explanations with greater reach, akin to how scientific theories must withstand critical evaluation.
• Adaptive Processing: CRNs provide a mechanism to dynamically adjust processing time for complex inferences, showcasing resilience against ambiguous and contradictory propositions, which is critical in domains requiring interpretative flexibility.
• Intrinsic Explainability: Unlike post hoc interpretability methods, CRNs inherently provide explanations for their predictions, aligning with calls for transparent and accountable AI systems.

Future Directions

The framework suggests intriguing avenues for future AI development. One such direction extends EL's application to interactive learning environments where the machine actively seeks observations to improve learning, a more realistic simulation of human scientific processes. Additionally, further exploration toward enhancing CRNs' resistance to adversarial manipulations could magnify their applicability in security-sensitive domains.

In conclusion, this paper illuminates a path forward in AI research by integrating symbolic interpretation into the learning paradigm, inviting further inquiry into the epistemological underpinnings of machine intelligence and its applications across various domains.