This classic paper in computational psychology addresses a long-standing debate: how does the brain combine different sources of information—such as the sound of a word and the visual context of a speaker's lips—to make a decision? Specifically, the authors demonstrate that the "Morton-Massaro law," a mathematical rule stating that stimulus and context influence perception independently, is entirely compatible with neural networks that contain feedback loops. They reveal that this factorization depends on the network's architectural wiring ("channel separability") rather than an absence of recurrent processing.
The Historical Conflict
In the 1980s, a conflict emerged between two schools of thought in perception. Experimental data frequently followed the Morton-Massaro law, which dictates that the probability of a specific response factorizes into independent stimulus and context components. Conversely, connectionist (neural network) models heavily relied on feedback and lateral connections, suggesting a highly interactive system where everything influences everything else.
Many researchers, notably Dominic Massaro, argued that because feedback models are inherently interactive, they could not produce the clean, independent factorization observed in human data. This paper sets out to prove that this perceived incompatibility is a myth.
Methods: Diffusion Networks and Channel Separability
To bridge this gap, the authors employ Diffusion Networks—recurrent neural networks defined by stochastic differential equations (SDEs).
- Stochastic Dynamics: By adding "Brownian motion" (noise) to standard neural network equations, the model can reach a "stochastic equilibrium." This is a steady state where the probability of the network being in a certain configuration can be calculated mathematically.
- Channel Separability: The authors define a specific architectural constraint. A network possesses "separable channels" if the units processing the stimulus and the units processing the context are disjoint and lack direct lateral connections to each other. Crucially, both channels can still maintain internal feedback loops, and both feed into a shared "response" layer.
- Mathematical Proof: Using "goodness functions" (akin to energy functions in physics), the authors prove that if a network is structurally separable, the equilibrium probability density of the response units will factorize, even in the presence of heavy internal feedback.
Empirical Validation
The paper evaluates the model against experimental data from Massaro and Cohen (1983), where subjects identified synthetic consonant sounds (ranging from /l/ to /r/) presented in different phonetic contexts. The experiment featured two response alternatives, seven stimulus levels, and four context conditions.
The authors compared the predictions of the Morton-Massaro law directly against 100 random samples from their diffusion network simulation.
As shown in the results, there is a near-perfect match between the Morton-Massaro law's predictions and the diffusion network's behavior. For instance, in one condition, the law predicted a probability of 0.9000, while the network produced 0.91. This effectively neutralized the argument that the Morton-Massaro law could be used as evidence against interactive, feedback-based models of the brain.
Key Insights
- Feedback Does Not Preclude Independence: The most surprising finding is that feedback doesn't automatically destroy independence. As long as the inputs to the system (stimulus and context) do not directly contaminate each other's sensors or relay units, the final decision can still appear as though it was made by two independent experts.
- Architecture Over Dynamics: The "factorization" property is a result of the network's layout (topology), not whether the signals flow unidirectionally or bidirectionally.
- The Role of "Sharpness": As the "gain" or sharpness of the response units increases (making them act more like all-or-nothing switches), the network's behavior converges perfectly to the Morton-Massaro law.
Limitations and Future Directions
- Symmetric Weights: The primary mathematical proof assumes symmetric connections between neurons ($w_{ij} = w_{ji}$). While approximations exist for more biologically realistic non-symmetric weights, the rigorous proof is tied to the symmetric case.
- Equilibrium Assumption: The model assumes the network has sufficient time to reach stochastic equilibrium before a response is made. In rapid real-world perception, the brain might make decisions during the transient phase before equilibrium is reached.
- Neural Testing: The authors suggest future work should investigate whether this factorization holds at the level of individual neural responses, potentially helping map the functional architecture of the brain's sensory channels.
Summary updated