Paper ID: 2304.00002

Beyond Interpretable Benchmarks: Contextual Learning through Cognitive and Multimodal Perception

Nick DiSanto

With state-of-the-art models achieving high performance on standard benchmarks, contemporary research paradigms continue to emphasize general intelligence as an enduring objective. However, this pursuit overlooks the fundamental disparities between the high-level data perception abilities of artificial and natural intelligence systems. This study questions the Turing Test as a criterion of generally intelligent thought and contends that it is misinterpreted as an attempt to anthropomorphize computer systems. Instead, it emphasizes tacit learning as a cornerstone of general-purpose intelligence, despite its lack of overt interpretability. This abstract form of intelligence necessitates contextual cognitive attributes that are crucial for human-level perception: generalizable experience, moral responsibility, and implicit prioritization. The absence of these features yields undeniable perceptual disparities and constrains the cognitive capacity of artificial systems to effectively contextualize their environments. Additionally, this study establishes that, despite extensive exploration of potential architecture for future systems, little consideration has been given to how such models will continuously absorb and adapt to contextual data. While conventional models may continue to improve in benchmark performance, disregarding these contextual considerations will lead to stagnation in human-like comprehension. Until general intelligence can be abstracted from task-specific domains and systems can learn implicitly from their environments, research standards should instead prioritize the disciplines in which AI thrives.

Submitted: Dec 4, 2022