Lässig's paper measures neural ambiguity linking networks to consciousness

Francesco Lässig's new paper, presented at the Models of Consciousness 6 conference, tackles a fundamental gap in neural network theory: how to measure whether a network's internal representations are unambiguous. The author formalizes ambiguity as conditional entropy H(I|R) — the uncertainty about what a representation means given the representation itself. This mirrors a key property of conscious experience, where a neural state cannot double as something else. Using MNIST digit classification as a testbed, Lässig shows that the relational structure of network connectivity can encode representational content with surprising fidelity.

Dropout-trained networks achieved perfect 100% accuracy in decoding which output neuron corresponded to which digit class solely from connectivity patterns, while standard backpropagation networks reached only 38% (10% chance). Strikingly, both training methods achieved identical task performance, proving that representational ambiguity can vary independently of behavioral accuracy. The paper also demonstrates that spatial position of input neurons — analogous to visual field location in biological vision — can be decoded from connectivity with R² up to 0.844. These results provide a quantitative framework for assessing representational ambiguity and suggest that neural networks can already possess the low-ambiguity representations that theorists like IIT proponents argue are necessary (though not sufficient) for consciousness.