Consciousness in Artificial Intelligence: Insights from the Science of Consciousness — part 10/14

Cleaned-text code points [139057, 154001). Bibliography omitted here; source.md includes it.

Implementing GWT in artificial systems has been the subject of several studies, including recent research by VanRullen and Kanai (2021) and Goyal et al. (2022). Here we give brief overviews of these two studies before discussing indicators GWT-1 through GWT-4 in turn.

VanRullen and Kanai’s (2021) proposal involves a set of specialised neural modules, at least some of which are generative networks that can produce behavioural outputs or drive sensory processing top-down. Each of these modules has its own low-dimensional latent space in which key information is represented. The modules are integrated by a workspace, which is a shared latent space trained to perform unsupervised translation of representations in the latent spaces of the modules so that the information they carry is available to the others. The workspace has a lower capacity than the sum of the module latent spaces, and a task-dependent key-query attention mechanism is used to determine which information it broadcasts. So this architecture includes the important features of a bottleneck, global broadcast and state-dependent selection. However, this work is a “roadmap” to a possible implementation, rather than a working system. It faces a substantial open question about how the attention mechanism could be trained to select among the potential inputs to the workspace, and especially how this could achieve the sequences of operations of attention needed to control extended, functional sequences of operations by relevant modules.

Meanwhile, Goyal et al. (2022) experimented with a method for implementing global workspace which similarly involved using key-query attention to select which of multiple modules would write to a shared space, with the contents of the shared space then being broadcast to all modules. In this case, the modules were trained together to generate mutually usable representations, avoiding the need for translation in the workspace. A limitation of the specific implementations developed in this work was that the “modules” were elements processing tokens in a sequence or parts of an image, so it is questionable whether they were specialised subsystems capable of operating in parallel; instead, they each contributed in similar ways to the performance of a single task.

Although neither of these studies produced a working system that clearly satisfies all four GWT indicators, this does seem to be a realistic objective, as we will argue by considering the indicators in turn. Indicator GWT-1 states that the system must have specialised systems, or modules, capable of working in parallel. To make global broadcast possible, these modules must be implemented by recurrent neural networks, unless they are “output” modules for the system as a whole, which do not provide information to the workspace. The modules might take as input:

1. Sensory input in one or more modalities.

2. Input from a small number of other modules that typically work in tandem. For instance, a “saccade” module might take input from a “visual saliency” module in order to enable quick bottom-up saccades towards potentially important objects.

3. Top-down signals coming from an executive Global Workspace module.

These modules might be trained independently on narrow tasks, as suggested by VanRullen and Kanai (2021). Or they might be jointly trained end-to-end with the workspace in order to achieve some system-wide objective, from which module specialisation to subtasks would naturally emerge. The end-to-end training approach was employed by Goyal et al. (2022), although for relatively simple tasks (see also Goyal et al. 2020).

The second element of an implementation of GWT is a limited-capacity workspace, which is a further neural module with different properties. The simplest way to limit the capacity of the workspace is to limit the number of dimensions of its activity space. Another interesting option is to train a recurrent neural network that exhibits attractor dynamics. An attractor is a state in a dynamical system such that when that state is reached, it will remain stable in the absence of inputs or noise to the system. The reason that attractor dynamics limits capacity is that it induces a many-to-one mapping from initial conditions in a neural trajectory to attractors (any neural trajectory that enters an attractor’s basin of attraction will converge to that attractor). Thus, these attractor dynamics induce an information bottleneck by contracting the size of the space of stable states. Ji, Elmoznino et al. (2023) argue that attractor dynamics in the workspace can help to explain the apparent richness and ineffability of conscious experience.

For indicator GWT-3, global broadcast, the basic requirement is that all modules take workspace representations as input. As we have seen, this means that some mechanism must be in place to ensure that these inputs can be used by all modules, such as the translation mechanism proposed by VanRullen and Kanai (2021). In the global neuronal workspace theory developed by Dehaene and colleagues (Dehaene & Naccache 2001, Dehaene & Changeaux 2011, Mashour et al. 2020), the workspace exhibits particular dynamic properties: for a representation to be in the workspace and globally broadcast, it must be sustained by recurrent loops. While it is not clear that this is essential, this behaviour could be replicated in AI if a network exhibiting attractor dynamics was used to implement the workspace. In this case, the broadcast mechanism might consist of a leaky neural integrator whose dynamics have slow timescales such that sustained inputs are required to place it in a particular state, and in the absence of these sustained inputs it relaxes back to some baseline state (as in models of decision-making through evidence accumulation). This broadcast mechanism would generate the top-down signals feeding into each specialised module.

Indicator GWT-4 includes the conditions that the system must use a state-dependent attention mechanism and that the workspace must be able to compose modules to perform complex tasks by querying them in succession. For the state-dependent attention mechanism, both VanRullen and Kanai (2021) and Goyal et al. (2022) propose the use of key-query attention, which is common in current AI models. A query can be computed from the workspace’s current state, and keys can be computed for all other modules. The similarity between the workspace’s query and a given module’s key would be normalised by the similarities across all other modules in order to introduce competition between modules, and these normalised similarities would determine the degree to which each module’s value contributes to the net input to the workspace. That is, a standard key-query attention mechanism would be applied at each timepoint to compute the input to the workspace in a way that depends on its current state.

The model described here would be able to meet the second part of GWT-4—the capacity to use the workspace to query modules in succession to perform complex tasks— when it is unrolled through time because there are computational loops between the workspace and the modules. The modules receive input from bottom-up sensory input and from a small number of other modules, but they also receive top-down input from the workspace. This means that, for instance, it is possible for one module to control others by controlling what is represented in the workspace. The sequential recruitment of modules by the workspace is within the computational repertoire of the system, so it could emerge if it is beneficial during training. However, suitable training would be required for such a system to learn to compose modules in useful ways and to perform complex tasks, and constructing a suitable training regime may be a significant challenge for implementing GWT.

#### 3.1.3 Implementing PRM

We now consider how perceptual reality monitoring theory, as a representative computational HOT, could be implemented in AI. That is, we consider how an AI system could be constructed with indicator properties HOT-1 through HOT-4. Although PRM researchers claim that there are no current AI systems that meet all of the requirements (Dehaene et al. 2017, Michel & Lau 2021, Lau 2022), and there have been no implementation attempts that we are aware of, standard machine learning methods are sufficient for plausible implementations of most elements of the theory. We begin by considering the implementation of a “quality space”—that is, a space of possible representations which satisfies the PRM account of the basis of phenomenal qualities—before turning to the core claim of PRM, that first-order representations become conscious as a result of being represented as “real” by a higher-order monitoring mechanism.

Indicator HOT-4 states that the system must have sparse and smooth coding generating a “quality space”. One of the most important features of all DNNs is that each layer endows the model with a smooth representation space, in which a continuous range of possible activations codes for the range of possible inputs. The smoothness of these representation spaces is thought to be one of the primary reasons why they can generalise to novel inputs at test time; even if a novel input drives a layer into a different activation pattern, this pattern can still be interpreted by subsequent layers so long as it is within the model’s training distribution (that is, so long as the layer activated with similar patterns at training time) (Belkin et al. 2018, Bartlett et al. 2021). In fact, there is evidence that the perceptual representation spaces learned by current DNNs already closely resemble those of the human visual system, meaning their corresponding “quality spaces” might already be substantially aligned. For instance, Kriegeskorte (2015) finds that the matrix of pairwise dissimilarities between representations of many visual stimuli in a DNN is closely aligned to what is observed in neural activity and that there is even a correspondence between successive layers in the network and the hierarchical organisation of the visual cortex—although more recent studies have complicated this picture (see e.g. Golan et al. 2020, Lindsay 2021, Bowers et al. 2022).

Standard methods in machine learning can be used to satisfy the condition that representations be sparse. For example, regularization techniques can enforce sparse representations in DNN models, by minimising the magnitude of the network’s representations or the mutual information between inputs and representations (Tishby 2000). Layers that use normalisation functions like SoftMax, in which higher values suppress lower values in the representation, also increase sparsity.

Indicators HOT-1 and HOT-2 state that the model must contain both first-order perceptual representations of sensory data and higher-order representations that assign a measure of reliability or “realness” to particular first-order representations. To meet these conditions, a vast number of known deep learning solutions are possible, of which we review a few. Importantly, all of these solutions share a simple architectural design of two components: (1) a (first-order) neural network takes sensory data and/or top-down signals as input and produces a number of perceptual representations distributed across a hierarchy of layers (2) in parallel, a series of separate (higher-order—specifically second-order) neural networks each take a first-order layer’s activations as input and then output a single scalar, representing the probability that the first-order representation of that layer is veridical. Solutions for meeting the conditions will differ primarily in terms of how the second-order networks are trained.

If supervision signals are occasionally present that provide the second-order networks with “ground-truth” about the reliability of first-order representations, then the second-order network can be trained to estimate the probability of correctness by standard supervised learning. Obtaining this ground-truth signal may be difficult, but not impossible. For instance, if a source of first-order representation errors is internal noise in the network, ground-truth can be estimated simply by averaging noisy first-order representations over time. Another possibility is to obtain ground-truth by comparing representations of the same percept in different sensory modalities (e.g., verifying the veracity of a sound using visual feedback) or through movement (e.g., checking if a visual percept behaves as it should given known motor actions).

If ground-truth is not directly available, the second-order networks can be trained on other surrogate tasks where the reliability of a signal is an implicit factor in performance. For instance, second-order networks might try to predict upcoming first-order representations using past ones. Because externally generated signals are sometimes more predictable than internally generated signals—a perception of a falling ball is more predictable than a hallucination generated by random internal noise—when the networks’ predictions have high error, the second-order network can assign a lower probability to the veracity of the first-order representation. Alternatively, when imagery is under cognitive control, engaging in imagination can lead to more predictable sensory consequences (I imagine a pink bear and a pink bear appears) than when it is not (I allow my mind to wander). Thus the level of effortful control can serve as another internal signal that could train a reality monitoring system (Dijkstra et al. 2022). In these ways, a second-order network could learn to use predictability as a cue to “realness” even in the absence of supervision signals. These methods are closely related to predictive coding, which has already seen applications in modern deep learning (e.g., Millidge et al. 2022, Oord et al. 2019, Alamia et al. 2023), except that here the prediction is done across time rather than across layers in a hierarchy.

Other methods of training the second-order network involve thinking of it as a world-model. For instance, Bayesian methods in deep learning view perception as an inference process, in which a neural network attempts to infer the values of latent variables that might have generated the data. For ideal Bayesian inference, these latent variables must be sampled according to their posterior probability, which is proportional to their prior probability multiplied by their likelihood of having generated the data under some world model that specifies how latent variables produce sensory observations. This sort of perspective fits well with PRM; the inference machinery for latent variable values can be seen as producing perceptual first-order representations, while the second-order networks evaluate the probability that the first-order representations are true.