Popper’s Knowledge Tensor (PKT)
Toward Hard Falsification in Neuro-Symbolic AI
Modern AI systems learn by induction — absorbing statistical patterns from massive datasets. They are powerful pattern matchers, but they cannot reject a hypothesis the way a scientist can. When a large language model hallucinates a plausible-sounding falsehood, there is no internal mechanism that says “this contradicts what I know to be true.” The model has no capacity for falsification.
PKT proposes a different kind of architecture: one where knowledge is not merely accumulated but actively tested and pruned. Inspired by Karl Popper’s philosophy of science, the framework treats every learned representation as a conjecture — provisional until it survives deductive stress tests.
The Core Idea
Most neuro-symbolic approaches (Logic Tensor Networks, DeepProbLog, NeurASP) integrate logic as a soft constraint: logical rules nudge the model toward consistency, penalizing violations through differentiable loss terms. The model is encouraged to be logical but never required to be.
PKT takes a harder line. It proposes a falsification operator that eliminates representations violating deductive rules, rather than merely penalizing them. The distinction matters:
| Approach | Mechanism | Failure Mode |
|---|---|---|
| Soft constraint (existing) | Penalty term in loss function | Model can learn to “pay the penalty” and keep inconsistencies |
| Hard falsification (PKT) | Projection operator zeros out violating entries | Inconsistent representations are structurally impossible |
This is analogous to the difference between a tax on pollution (soft) and a physical filter that removes pollutants (hard).
Roadmap
This blog develops the PKT idea across six threads, each building toward a formal paper:
- Motivation — Why pure induction isn’t enough, and why hallucination is a symptom of a deeper architectural gap.
- Philosophy — The epistemological foundations: Popper’s falsificationism, Hegel’s dialectic, and the question of what knowledge is.
- Landscape — What already exists in neuro-symbolic AI, and where PKT fits (and differs).
- Framework — The formal definition: the Knowledge Tensor, the Falsification Operator, and a proposed loss function.
- Open Questions — The deep questions this work raises about the nature of knowledge, agency, and AI.
- Research Log — A Popperian falsification log: what I’ve conjectured, tested, and discarded.
Status
This is active research in progress. The framework is speculative — definitions are proposed rather than proven, and no experiments have been run yet. The blog is the thinking tool; the paper comes later.
What you’re reading is the conjecture. The refutation is the work ahead.