1. Global Workspace Theory (GWT) and AI

#1.1 Van Rullen & Kanai — the actual architectural proposal

"Deep Learning and the Global Workspace Theory", Rufin VanRullen & Ryota Kanai, arXiv 2012.10390 (v2 Feb 2021; published in Trends in Neurosciences 2021). Retrieved: full abstract.

Actual claim, verbatim:

"We argue that the time is ripe to consider explicit implementations of this theory using deep learning techniques. We propose a roadmap based on unsupervised neural translation between multiple latent spaces (neural networks trained for distinct tasks, on distinct sensory inputs and/or modalities) to create a unique, amodal global latent workspace (GLW). Potential functional advantages of GLW are reviewed, along with neuroscientific implications."

So the architecture is: several specialist networks, each already trained and frozen on its own task/modality; a shared amodal latent space; bidirectional encoders/decoders between each specialist's latent space and the shared space; trained by cycle-consistency (unsupervised translation), not by labels.

What they claim would be testable. The abstract does not state testable predictions about consciousness. The testable content is functional/behavioural, and the group's follow-ups supply the actual tests:

• Devillers, Maytié, VanRullen, *Semi-supervised Multimodal Representation Learning through a Global Workspace*, arXiv 2306.15711 (IEEE TNNLS 2024). Claim: with a shared workspace + cycle-consistency, alignment/translation between modalities needs "4 to 7 times less" matched data than fully supervised approaches, and *"Ablation studies reveal that both the shared workspace and the self-supervised cycle-consistency training are critical to the system's performance."* This is the measurable, falsifiable core of the GW proposal: data-efficiency + ablation necessity.

• Chateau-Laurent & VanRullen, *Learning to Chain Operations by Routing Information Through a Global Workspace*, arXiv 2503.01906: a controller gates information between modules through a shared workspace. Claim: *"the Global Workspace model, while having fewer parameters, outperforms LSTMs and Transformers when tested on unseen addition operations (both interpolations and extrapolations of addition operations seen during training)."* → testable: systematic generalisation to unseen compositions, at lower parameter count.

• Maytié, Johannet, VanRullen, Multimodal Dreaming: A Global Workspace Approach to World Model-Based RL, arXiv 2502.21142: GW latent space + world model ("GW-Dreamer"). Claim: *"performing the dreaming process (i.e., mental simulation) inside the GW latent space allows for training with fewer environment steps" and, as an emergent property, "the resulting model (but not its comparison baselines) displays strong robustness to the absence of one of its observation modalities (images or simulation attributes)."* → testable: sample efficiency + modality-dropout robustness.

• Goyal et al., Coordination Among Neural Modules Through a Shared Global Workspace, arXiv 2103.01197 (ICLR'22). This is the cleanest statement of why the bottleneck is functional: *"due to limits on the communication bandwidth, specialist modules must compete for access. We show that capacity limitations have a rational basis in that (1) they encourage specialization and compositionality and (2) they facilitate the synchronization of otherwise independent specialists."* → testable: vary workspace bandwidth and measure specialisation + compositionality.

• Juliani, Arulkumaran, Sasai, Kanai, *On the link between conscious function and general intelligence in humans and machines*, arXiv 2204.05133. (The user asked about a "Juliani et al. Deep Learning and the GWT" paper as possibly separate from Van Rullen & Kanai — it is separate, but its title/subject is the consciousness↔intelligence link, not the deep-learning GW roadmap. It examines GWT, "Information Generation Theory" and AST and argues all three link conscious function to domain-general intelligence, proposing "mental time travel" as an implementable near-term target.)

#1.2 Attention Schema Theory (AST) and its tests

• Graziano & Webb, The attention schema theory: a mechanistic account of subjective awareness, Frontiers in Psychology 6:500 (2015), doi 10.3389/fpsyg.2015.00500. (Citation verified via OpenAlex; I did not retrieve the full text.)

• Graziano, The Attention Schema Theory: A Foundation for Engineering Artificial Consciousness, Frontiers in Robotics and AI 4:60 (2017), doi 10.3389/frobt.2017.00060.

• Wilterson & Graziano, *The attention schema theory in a neural network agent: Controlling visuospatial attention using a descriptive model of attention*, PNAS (2021), doi 10.1073/pnas.2102421118. Title verified; abstract not retrieved. This is the key machine test from the AST group: an RL/neural agent that controls spatial attention using a descriptive model of its own attention.

• Wilterson et al., Attention control and the attention schema theory of consciousness, Progress in Neurobiology (2020), doi 10.1016/j.pneurobio.2020.101844. Six human experiments. Verbatim from abstract: *"Because AST is a control-engineering style theory, it can make specific predictions in complex situations."* Key dissociation reported: exogenous attentional cuing persisted when participants were unaware of the cue, but was impaired when they were aware of it; and implicit learning of the cue–target association generalised to adjacent untrained locations. This is a testable double dissociation between awareness of a cue and its attentional effect.

• Piefke, Doerig, Kietzmann, Thorat, *Computational characterization of the role of an attention schema in controlling visuospatial attention*, arXiv 2402.01056 (CogSci 2024). Verbatim: *"the more uncertain the agent was about the location of its attentional window, the more it benefited from these additional resources, which developed an attention schema. Together, these results indicate that an attention schema emerges in simple learning systems where attention is important and difficult to track."* → testable: the emergence condition is measurable — attention-tracking uncertainty must be high. If your agent's attention is trivially observable, an attention schema buys nothing.

• Farrell, Ziman, Graziano, *Testing Components of the Attention Schema Theory in Artificial Neural Networks*, arXiv 2411.00983. This is the strongest AST evidence in machines. Verbatim: *"we found that an agent with an attention schema is better at categorizing the attention states of other agents (higher accuracy). Second, an agent with an attention schema develops a pattern of attention that is easier for other agents to categorize. Third, in a joint task where two agents must predict each other to paint a scene together, adding an attention schema improves performance. Finally, the performance improvements are not caused by a general increase in network complexity. Instead, improvement is specific to tasks involving judging, categorizing, or predicting the attention of other agents."* → the decisive control is the parameter-matched complexity control. This is what makes AST's central claim — that the schema is used to model others' attention — a real prediction rather than a free lunch.

• Liu, Bolotta, Zhu, Bengio, Dumas, Attention Schema in Neural Agents, arXiv 2305.17375: *"agents that implement the AS as a recurrent internal control achieve the best performance... these exploratory experiments suggest that equipping artificial agents with a model of attention can enhance their social intelligence."*

• A theoretical warning worth noting: Steel, *Modelling aspects of consciousness: a topological perspective, arXiv 2011.05294, proves that "a complete representation of attention is not possible, since it cannot faithfully represent streams of attention"* — supporting AST's claim that the attention schema is necessarily incomplete. This is a genuine mathematical result with a testable consequence: an attention schema can never be a perfect predictor of the agent's own attention.

Strength of evidence (GWT/AST): moderate but only functional. GWT has real implemented systems with ablations and generalisation results. AST has an implemented agent with a parameter-matched control and a specific claimed benefit (modelling others' attention). Neither has any validated link to phenomenal consciousness. Both are best read as architectures with measurable functional payoffs.