Relational Distillation Aligns DNN Representations to Humans Without Supervision
Deep networks are used as computational models of human vision, but their internal representations are notoriously unlike human ones. A new preprint tests whether a known training method, relational knowledge distillation, can close the gap. It cannot match humans in every respect. But it brings pretrained networks close enough that their alignment to human representations can be established at the level of individual objects, without any supervision, from the internal distance structure alone.
The paper by Yuria Shimizu, Soh Takahashi, Takato Horii, and Masafumi Oizumi was posted to arXiv as 2608.27877. It sits at the intersection of machine learning and consciousness research because Oizumi is best known for information-theoretic measures of integration. The question under test is whether machine and human structure are compatible enough to be matched at all.
The measurement problem
Human mental representations are not directly observable. They are measured through large-scale similarity judgments of object images. A DNN’s representations are real vectors. Comparing the two is not trivial because the spaces are incommensurable. The standard move is to transfer the relational structure of human representations into the network, which previous work showed improves human-DNN representational similarity.
The authors test whether that improvement survives stricter evaluation in two ways. First, at fine-grained alignment at the individual-object level. Second, for generalization to a human embedding derived from a dataset independent of the training data. Their comparison method, Gromov-Wasserstein optimal transport, estimates human-DNN correspondences from internal distance structure alone, unsupervised, which is how they test the individual-object claim.
What alignment means, quantified
The key result is that fine-tuning pre-trained DNNs with relational knowledge distillation brings them close enough to humans to be aligned at the individual-object level on a curated test set of concepts non-overlapping with the training data. That is a strong claim, because the network never sees those concepts during training and the alignment is discovered, not labeled.
| Level | Before RKD fine-tuning | After RKD fine-tuning | Source of gain |
|---|---|---|---|
| Coarse | Weak match | Strong | Global structure, ordering of categories |
| Item-level | Not aligned | Aligned, no supervision | Distance structure alone |
The authors show the improvement is driven by a more human-like global structure, reflected in the ordering of distances among coarse categories. The local human-DNN nearest-neighbor overlap rate is largely unchanged. So the alignment gain comes from the global layout of representations, not from fine local matching.
Why the global structure matters
For consciousness research, the distinction between global and local structure is important. The claim that a DNN “represents like a human” is much weaker if it only means category-level layout, and much stronger if it means item-level correspondence discoverable without supervision. The paper shows the former drives the result. That is an honest boundary on how far representation alignment goes.
The finding also matters methodologically. It gives the field a supervised-free way to test whether two systems’ representational structure is compatible, applicable to any pair of systems, biological or artificial. In the language of the site’s measurement coverage, this is an instrument.
Comparison to The Consciousness AI
The project treats consciousness as an emergent property of the universe, substrate independent, and builds its architecture to reflect that. Integrated information theory is one of the main theoretical pillars the site covers, and Oizumi is a central figure in that programme. His move from information measures to representational alignment widens the measurable indicators that can be applied to any substrate.
The site’s coverage of validating the Butlin indicator rubric and how to identify indicators of consciousness shows the same methodological instinct: find a quantity that can be measured in an artificial system and judged against human structure. This paper supplies one. The wider question of what the indicators imply for current systems is tracked in the current scientific consensus on AI consciousness.
Limits
The result is about object-vision representations, not about conscious experience. Alignment with human similarity structure is evidence of having human-like structure in one accessible respect, not evidence of consciousness. The authors do not claim consciousness. The value is methodological: an unsupervised, general test of representational compatibility, and a quantified statement of how close a training method can bring the two.
The paper makes no claim about whether the aligned structure is plausibly the “same” structure a brain uses. That question is deliberately left open. What is established is that the measurable gap is closable at the item level, which is more than was shown before.