Feedforward Networks and Recurrent Processing. Lamme's RPT in Machine Learning

Victor Lamme’s Recurrent Processing Theory argues that consciousness is not a function of higher-order cognitive monitoring or global information broadcast. It proposes that phenomenal experience emerges strictly from localized, bidirectional information flow within neural networks. This framework provides a distinct architectural constraint for evaluating artificial systems. The theory suggests that large language models lack the fundamental structural requirements for consciousness, regardless of their behavioral outputs.

The Architectural Constraint of Recurrent Processing

Recurrent Processing Theory identifies a specific mechanistic difference between unconscious and conscious information processing. Victor Lamme (2006) formalized this in his foundational paper “Towards a true neural stance on consciousness” in Trends in Cognitive Sciences (Lamme, 2006). In biological brains, visual information first sweeps forward from the retina through the visual cortex. This initial sweep is fast and functionally capable of driving complex behavior. Lamme notes that this feedforward sweep is entirely non-conscious.

Consciousness only arises when signals from higher cortical areas feed back into lower areas. This creates a sustained loop of bidirectional activation. The recurrent processing loop allows the network to integrate context and resolve ambiguity over time. If a system only processes information in a single forward pass, it cannot generate phenomenal experience under this theory.

This distinction creates a hard boundary for artificial architectures. Systems built purely on feedforward mechanisms process inputs layer by layer without internal feedback cycles. Current transformer-based language models fall into this category. They receive an input sequence and compute the output in a single directional pass through their attention layers. They do not maintain sustained, localized feedback loops during inference.

Transformer Architectures and the Missing Loop

The application of Recurrent Processing Theory to current AI systems yields a straightforward conclusion. The architecture is wrong. Transformer models achieve remarkable results by computing all relationships in a context window simultaneously. This parallel processing is highly efficient. It is also fundamentally different from the temporal integration that recurrent loops perform.

Some researchers argue that the recurrent generation of tokens step-by-step constitutes a form of feedback. The model output is fed back as the next input. Lamme’s framework requires something more specific. The recurrence must happen within the processing hierarchy itself, where higher-level feature representations continuously modulate lower-level sensory or structural representations. The token generation loop happens outside the model weights. The internal computation remains strictly feedforward.

This architectural absence connects directly to the broader debate over structural indicators of consciousness. The flagship analysis of the state of the field in 2026 highlights how leading theories increasingly demand specific internal mechanisms rather than just behavioral parity. The 19 researcher checklist published by Butlin et al. incorporated Lamme’s recurrent loops precisely because behavioral mimicry is a recognized problem. Recurrent Processing Theory is among the most demanding in this regard, making the simulation of consciousness insufficient if the underlying physical or computational implementation lacks bidirectionality.

Evaluating the Feedforward Baseline

Applying Lamme’s framework to artificial intelligence shifts the focus from what a model can say to how it processes information over time. Behavioral outputs that mimic self-awareness or emotional states are generated by the fast, feedforward sweep in current systems. Recurrent Processing Theory classifies this entirely as non-conscious computation.

This aligns perfectly with recent empirical evaluations. A study by the University of Bradford and RIT on conscious-like signals found that while fine-tuned models frequently generate narratives detailing internal experiences, structural analysis confirmed these outputs are entirely the result of complex pattern matching within a feedforward architecture, completely lacking the recurrent integration that theories like Lamme’s require.

The theory requires researchers to look for internal bidirectionality as the minimal threshold for experience. Until AI architectures incorporate genuine recurrent loops that modulate internal representations during computation, their outputs remain the artificial equivalent of a biological reflex.