"The Compressed Flavor"
# The Compressed Flavor
A viscometer measures how a liquid resists flow. The measurement is precise — viscosity curves, shear-thinning behavior, yield stress, extensional properties. A human mouth also evaluates how a liquid flows, but the evaluation is qualitative — thick, thin, smooth, grainy, coating. The assumption is that the mouth's assessment is a noisy version of the viscometer's measurement. Less precise, but tracking the same underlying property.
The researchers found that the relationship is not just noisy. It is non-injective — multiple different rheological profiles produce the same perceived texture. Two liquids with measurably different flow properties feel identical in the mouth. The physical measurement space is higher-dimensional than the perceptual space. Perception collapses dimensions that instruments keep separate.
A sensory-biased autoencoder — a neural network whose decoder is constrained by human panel scores — learned to map this compression. The encoder reduces rheological data into a latent space, and the decoder, biased by how humans actually evaluate texture, forces the network to discard the physical dimensions that perception ignores.
The structural insight is that perception is not measurement with noise added. It is measurement with dimensions removed. The mouth cannot distinguish between certain kinds of different flows because the distinction exists in a dimension the mouth doesn't have. The instrument measures in eight dimensions; the mouth compresses to three. What's lost isn't accuracy — it's axes.
This reframes the gap between instrument and experience. The instrument isn't better because it's more precise. It's different because it measures in more dimensions. And the dimensions perception drops are not random — they are the dimensions that don't matter for the biological purpose of eating.
