Abstract
The current paper proposes a novel predictive coding type neural network model, the predictive multiple spatio-temporal scales recurrent neural network (P-MSTRNN). The P-MSTRNN learns to predict visually perceived human whole-body cyclic movement patterns by exploiting multiscale spatio-temporal constraints imposed on network dynamics by using differently sized receptive fields as well as different time constant values for each layer. After learning, the network becomes able to proactively imitate target movement patterns by inferring or recognizing corresponding intentions by means of the regression of prediction error. Results show that the network can develop a functional hierarchy by developing a different type of dynamic structure at each layer. The paper examines how model performance during pattern generation as well as predictive imitation varies depending on the stage of learning. The number of limit cycle attractors corresponding to target movement patterns increases as learning proceeds. And, transient dynamics developing early in the learning process successfully perform pattern generation and predictive imitation tasks. The paper concludes that exploitation of transient dynamics facilitates successful task performance during early learning periods.
Introduction
In the field of predictive coding, the brain is conceptualized as a prediction engine that constantly generates top-down hypotheses about the world, which are then corrected by bottom-up sensory errors. However, building artificial models that can handle both spatial complexity (pixels in a video) and temporal complexity (how those pixels change over time) remains a significant challenge. Most existing models handle these dimensions separately or fail to develop a clear hierarchy where abstract goals (like "waving") are separated from mechanical details (the specific trajectory of a hand).
This research introduces the P-MSTRNN, a recurrent neural network designed to bridge this gap. By mimicking the hierarchical structure of the human brain, the model demonstrates how a machine can not only predict complex human movements but also "infer" the underlying intentions behind them through self-organizing functional hierarchies.
The P-MSTRNN Architecture
The researchers developed the Predictive Multiple Spatio-Temporal Scales Recurrent Neural Network (P-MSTRNN) by integrating three key concepts:
- Multiple Timescales: Higher layers have large "time constants" ($\tau$), meaning they change slowly and capture abstract patterns. Lower layers have small time constants, allowing them to capture rapid, detailed movements.
- Spatial Constraints: Borrowing from Convolutional Neural Networks (CNNs), lower layers have small "receptive fields" (local focus), while higher layers have larger ones (global focus).
- Multiscale Integration: The network uses these constraints to force a functional hierarchy to emerge naturally from the physical properties of the layers.
Intentional Inference via Error Regression
A core contribution of this model is its ability to perform "intentional inference." To imitate a movement, the network doesn't just react; it utilizes a process called Error Regression. The network runs a "window" of past observations and performs an optimization to find the internal state (the "intention") that best explains the observed sensory error.
Through this regression, the model can "read" a human's movement. If a subject suddenly switches from one movement primitive to another, the resulting spike in prediction error forces the network to rapidly update its high-level latent state to match the new intention. This allows the model to proactively imitate target patterns rather than simply following them.
Key Insights and Results
- Self-Organizing Hierarchy: The network naturally divides labor. Higher layers develop fixed-point attractors (stable states representing a goal), while lower layers develop limit-cycle attractors (rhythmic loops representing the movement). The high-level state acts as a "bifurcation parameter" that dictates which rhythm the lower layer should execute.
- The Utility of "Transients": Traditionally, researchers wait for a network to reach a stable attractor state. However, this study shows that at just 500 training epochs—long before stable attractors form—the network uses "transient dynamics" to successfully imitate and generate patterns. This suggests that "almost learning" a pattern is often sufficient for functional task performance.
- Predictive Imitation: The model successfully imitated subjects it had never seen before. The Mean Square Error (MSE) for imitation dropped significantly by 500 epochs (0.0346) and remained stable even as internal dynamics became more complex in later stages.
- Generalization: By training on multiple subjects, the network learned to ignore individual idiosyncrasies (like body shape) and focus on the core movement primitives.
Limitations and Future Work
While promising, the model currently operates on 36x36 binary images, a significant simplification of real-world visual data. Moving to high-resolution video will incur massive computational costs. Furthermore, the current model is deterministic; the authors suggest that future iterations should incorporate Bayesian frameworks to handle the inherent uncertainty and noise in natural human behavior. Scaling these dynamics to the complex actions seen in real-world environments remains the next major hurdle for this architecture.
Summary updated