Soft-prototypical networks improve concept grounding without task-specific losses
Soft Symbol Grounding for Prototypical Concepts
Artificial Intelligence
Summary
Some AI models learn to recognize complex things by only seeing the final answer, which can cause them to guess the right answer but misunderstand the parts that make it up. The authors created a method called Soft-PNet to help AI link these parts to real examples in a flexible way that works across many tasks without needing special tweaking. Their approach uses a smart search among possible answers, guided by a few example concepts, making training easier and faster. They tested it on different problems and found it works as well as older methods that needed a lot of special rules.
What this means in practice
- •For ai developers: Use Soft-PNet to train models that better learn intermediate concepts from limited labeled examples, reducing errors caused by misinterpreted concepts.
- •For mobile app developers: Incorporate Soft-PNet for tasks requiring interpretable reasoning from visual inputs with scarce supervision, improving concept-level accuracy without task-specific loss design.
Authors
Marcos Galván-López, Nijesh Upreti, Hiram Calvo, Carlos Aguilar-Ibáñez, Vaishak Belle
Abstract
Neuro-symbolic models are usually trained with supervision only on final labels, leaving the intermediate concepts unobserved. Since many concept assignments are consistent with a given label, training can predict labels correctly while recovering the wrong concepts, a failure known as a reasoning shortcut. Prototypical networks reduce shortcuts by anchoring each concept to a few labeled examples, but existing methods still couple perception and reasoning through a hand-crafted, task-specific differentiable loss that must be redesigned for every task. We introduce \textbf{Soft-PNet}, which removes this loss: it reframes concept grounding as a Metropolis walk over a precomputed cache of feasible symbolic solutions, guided by a prototype distribution built from a single labeled anchor per concept, and trains against one KL objective between the prototype-weighted cache and the network's concept predictions. The objective is identical across tasks and remains applicable when the solution space cannot be enumerated. On \texttt{MNIST-EvenOdd}, Visual Sudoku, and \texttt{Kand-Logic} under scarce supervision, Soft-PNet matches loss-engineered prototypical networks at the concept and label levels and recovers concepts that soft-grounding baselines miss, with no loss engineering and lower training time.