Gordon Pask and the System That Grew Its Own Sensor

In the late 1950s Gordon Pask demonstrated a device that the cybernetics community would spend decades interpreting. He immersed an array of platinum electrodes in an acidic ferrous sulphate solution and passed a current between selected pairs. Under the current, iron deposited out of solution and grew as conductive dendritic threads between the electrodes. The structure was not stable in the ordinary sense. Each thread dissolved continuously in the acid and persisted only while it carried enough current to rebuild itself faster than it decayed. The assemblage therefore existed as a dynamic equilibrium, a continuous competition among threads for a limited supply of current.

Reinforcement as physical structure

Pask used that current as the reward signal. When more current was made available, the structures carrying it could extend and reinforce, while structures that were not rewarded dissolved back into the solution. Reinforcement here was not a numerical parameter adjusted inside a fixed model. It was the physical material of the system itself. The device was constructed, quite literally, from its own reward.

It did not optimise a mapping within a fixed feature space. It added a dimension to that space by constructing the hardware to measure a new variable.

The consequential result concerns sensitivity. A growing thread structure is weakly responsive to many physical disturbances, among them mechanical vibration, local changes in acidity, and magnetic fields. In an ordinary device such incidental sensitivity is noise. In Pask's assemblage, when a disturbance happened to modulate the current in a way that was subsequently rewarded, the region of the structure responsible was reinforced and enlarged. Over a training period of roughly half a day this process produced a discrete physical organ. In his 1960 account Pask reported that the device grew a structure sensitive to sound, able to discriminate two frequencies near fifty and one hundred cycles per second, the responsible region consisting of fibrils that resonated with the stimulus. He described it without embellishment as an ear.

Extending the input space

The distinction this draws is precise, and it remains unresolved in contemporary systems. A trained model adjusts parameters behind a sensor that its designers fixed in advance. Its microphone, its pixel grid, and its token vocabulary are given. Within that fixed input space it can become arbitrarily capable, but it cannot, by training alone, become sensitive to a variable that was never represented to it. Pask's assemblage did precisely that. It did not optimise a mapping within a predetermined feature space. It added a dimension to the feature space by constructing the hardware required to measure a new variable. Peter Cariani, whose analysis remains the standard reference, treats the device as the clearest physical demonstration of a system that creates its own observables.

The relevance to current work is specific. Parameter learning, whether by backpropagation or reinforcement, operates strictly within a given input space and cannot manufacture a sensor that was not provided. The architectural search methods that appear to relax this constraint, including neural architecture search and automated machine learning, in fact explore a space of possibilities the designer has already enumerated. The one mechanism in a contemporary system that genuinely widens its own input space at runtime is tool acquisition: an agent that writes code to query an instrument it was not given, that calls an interface returning a class of data it has not previously encountered, that incorporates a new measurement channel into its own loop. This is the digital analogue of a structure seeking current in order to build a receptor. The capacity to grow a new organ of perception was demonstrated in an electrochemical assemblage before it was achieved in any digital system, and it has still not been achieved by learning alone.