NUTS, NARS, and Speech

TL;DR

This work investigates whether Non Axiomatic Reasoning (NARS) can support speech recognition through NUTS, a pipeline that applies random dimensionality reduction to Mel-frequency features before reasoning with Narsese. By using a minimal few-shot setup (two labeled examples per class) on Speech Commands v2, NUTS achieves about $64\%$ accuracy on a 35-class task, with very fast inference after reduction, but still trails state-of-the-art ANAM (~$93$–$94\%$) and Whisper Tiny (~$58\%$). The study demonstrates the feasibility of perception with non-axiomatic reasoning and highlights trade-offs between data efficiency, computational cost, and accuracy, arguing for the potential of integrating reasoning with perception under resource constraints. It also discusses interpretability and the impact of resource limits on intelligent systems, suggesting directions for making such approaches more scalable and robust.

Abstract

To investigate whether "Intelligence is the capacity of an information-processing system to adapt to its environment while operating with insufficient knowledge and resources", we look at utilising the non axiomatic reasoning system (NARS) for speech recognition. This article presents NUTS: raNdom dimensionality redUction non axiomaTic reasoning few Shot learner for perception. NUTS consists of naive dimensionality reduction, some pre-processing, and then non axiomatic reasoning (NARS). With only 2 training examples NUTS performs similarly to the Whisper Tiny model for discrete word identification.

Figures (3)

• Figure 1: Accuracy as a function of the reduced dimension embedding. Number of examples per class = 3.
• Figure 2: Accuracy as a function of the number of examples. Reduced dimensions = 4.
• Figure 3: Performance as a function of the AIKR limit. Examples= 3, Dimensions=4