In predictive processing, attention is a system that manipulates the confidence intervals on your predictions. Low attention -> wide intervals -> lots of mismatch between prediction and data doesn't register as an error. High attention -> tighter intervals -> slight mismatch leads to error signal. 

Hmm... that sounds a bit different from how I've understood attention in predictive processing to work. AFAIK, it's not that attending to something tightens its confidence interval; it's that things with tight confidence intervals (relevant for the task in question) get more attention.

So bottom-up attention would be computed by the subsystems that were broadcasting content into GNW, and their attention assignments would be implicit in their output. E.g. if you are looking at someone's face and a subsystem judges that the important thing to pay attention to is the fact that they are looking angry, then it would send the message "this person looks angry" to the GNW. And subsystems would have a combination of learned and innate weights for when their messages could grab control of the GNW and dominate it with what they are paying attention to, similar to the salience cues in the basal ganglia that allow some bids to dominate in specific situations.

Top-down attention would be computed by attentional subsystems interfacing with the GNW, to pick out specific signals in the GNW that seemed most useful for the current task, and strengthening those signals.

The GNW seems like it can only broadcast "simple" or "small" things. A single image, a percept, a signal. Something like a hypothesis in the PP paradigm seems like too big and complex a thing to be "sent" on the GNW

Is it? Like, suppose that a subsystem's hypothesis is that you are seeing a person's face; as a result, the image of a face is sent to the GNW. In that case, the single image that is transmitted into the GNW is the hypothesis. That hypothesis being in consciousness may then trigger an error signal due to not matching another subsystem's data, causing an alternative hypothesis to be broadcast into consciousness.

That said, seeing a face usually involves other things than just the sight of the face: thoughts about the person in question, their intentions, etc. My interpretation has been that once one subsystem has established that "this is a face" (and sends into consciousness a signal that highlights the facial features that it has computed to get the most attention), other subsystems then grab onto those features and send additional details and related information into consciousness. The overall hypothesis is formed by many distinct pieces of data submitted by different subsystems - e.g. "(1) I'm seeing a face, (2) which belongs to my friend Mary, (3) who seems to be happy; (4) I recall an earlier time when Mary was happy".

Here's an excerpt from Consciousness and the Brain that seems relevant:

In 1959, the artificial intelligence pioneer John Selfridge introduced another useful metaphor: the “pandemonium.” He envisioned the brain as a hierarchy of specialized “daemons,” each of which proposes a tentative interpretation of the incoming image. Thirty years of neurophysiological research, including the spectacular discovery of visual cells tuned to lines, colors, eyes, faces, and even U.S. presidents and Hollywood stars, have brought strong support to this idea. In Selfridge’s model, the daemons yelled their preferred interpretation at one another, in direct proportion to how well the incoming image favored their own interpretation. Waves of shouting were propagated through a hierarchy of increasingly abstract units, allowing neurons to respond to increasingly abstract features of the image—for instance, three daemons shouting for the presence of eyes, nose, and hair would together conspire to excite a fourth daemon coding for the presence of a face. By listening to the most vocal daemons, a decision system could form an opinion of the incoming image—a conscious percept.

Selfridge’s pandemonium model received one important improvement. Originally, it was organized according to a strict feed-forward hierarchy: the daemons bellowed only at their hierarchical superiors, but a high-ranking daemon never yelled back at a low-ranking one or even at another daemon of the same rank. In reality, however, neural systems do not merely report to their superiors; they also chat among themselves. The cortex is full of loops and bidirectional projections. Even individual neurons dialogue with each other: if neuron α projects to neuron β, then β probably projects back to α. At any level, interconnected neurons support each other, and those at the top of the hierarchy can talk back to their subordinates, so that messages propagate downward at least as much as upward.

Simulation and mathematical modeling of realistic “connectionist” models with many such loops show that they possess a very useful property. When a subset of neurons is excited, the entire group self-organizes into “attractor states”: groups of neurons form reproducible patterns of activity that remain stable for a long duration. As anticipated by Hebb, interconnected neurons tend to form stable cell assemblies.

As a coding scheme, these recurrent networks possess an additional advantage—they often converge to a consensus. In neuronal networks that are endowed with recurrent connections, unlike Selfridge’s daemons, the neurons do not simply yell stubbornly at one another: they progressively come to an intelligent agreement, a unified interpretation of the perceived scene. The neurons that receive the greatest amount of activation mutually support one another and progressively suppress any alternative interpretation. As a result, missing parts of the image can be restored and noisy bits can be removed. After several iterations, the neuronal representation encodes a cleaned-up, interpreted version of the perceived image. It also becomes more stable, resistant to noise, internally coherent, and distinct from other attractor states. Francis Crick and Christof Koch describe this representation as a winning “neural coalition” and suggest that it is the perfect vehicle for a conscious representation.

The term “coalition” points to another essential aspect of the conscious neuronal code: it must be tightly integrated. Each of our conscious moments coheres as one single piece. When contemplating Leonardo da Vinci’s Mona Lisa, we do not perceive a disemboweled Picasso with detached hands, Cheshire cat smile, and floating eyes. We retrieve all these sensory elements and many others (a name, a meaning, a connection to our memories of Leonardo’s genius)—and they are somehow bound together into a coherent whole. Yet each of them is initially processed by a distinct group of neurons, spread centimeters apart on the surface of the ventral visual cortex. How do they get attached to one another?

One solution is the formation of a global assembly, thanks to the hubs provided by the higher sectors of cortex. These hubs, which the neurologist Antonio Damasio calls “convergence zones,” are particularly predominant in the prefrontal cortex but also in other sectors of the anterior temporal lobe, inferior parietal lobe, and a midline region called the precuneus. All send and receive numerous projections to and from a broad variety of distant brain regions, allowing the neurons there to integrate information over space and time. Multiple sensory modules can therefore converge onto a single coherent interpretation (“a seductive Italian woman”). This global interpretation may, in turn, be broadcast back to the areas from which the sensory signals originally arose. The outcome is an integrated whole. Because of neurons with long-distance top-down axons, projecting back from the prefrontal cortex and its associated high-level network of areas onto the lower-level sensory areas, global broadcasting creates the conditions for the emergence of a single state of consciousness, at once differentiated and integrated.

This permanent back-and-forth communication is called “reentry” by the Nobel Prize winner Gerald Edelman. Model neuronal networks suggest that reentry allows for a sophisticated computation of the best possible statistical interpretation of the visual scene. Each group of neurons acts as an expert statistician, and multiple groups collaborate to explain the features of the input. For instance, a “shadow” expert decides that it can account for the dark zone of the image—but only if the light comes from the top left. A “lighting” expert agrees and, using this hypothesis, explains why the top parts of the objects are illuminated. A third expert then decides that, once these two effects are accounted for, the remaining image looks like a face. These exchanges continue until every bit of the image has received a tentative interpretation.

If you put three targets all in a row, people are able to detect them just fine. Additionally, if you ask people to remember the entire sequence they can do better than when you ask them to just remember only some of the characters (up till the point where you max out working memory). This makes no sense if the earlier experiments where interacting with a fundamental processing period that anything being attended to requires. 

Huh! Nice find. That's weird, I'm confused now.

Easy tasks can route around the global workspace, hard ones or ones that produce error have to go through it. That's the previous idea. Now, this paper has begun to shift my thinking. For a specific set of tasks, it claims to show that training doesn't shift activity away from a bottleneck location, but instead makes the processing at the point of the bottleneck more efficient.

Interesting! This would make a lot of intuitive sense - after the subsystems responsible for some task have been sufficiently trained, they can mostly just carry out the task using their trained-up predictive models, and need a lot less direct sensory data.

This might also explain some aspects of meditation: for example, The Mind Illuminated talks about "increasing the power of consciousness", which it describes as literally increasing the amount of experience-moments per unit of time. I was never quite sure of how exactly to explain that in terms of a global workspace model... but maybe if the system for generating moments of introspective awareness also got more efficient somehow? Hmm...

If subsystems had to route through the GNW to trigger motor actions, then this system or some variation could totally account for the serial conflict resolution function. But if subsystems can directly send motor commands without going through the GNW, how would would subsystems in conflict be "told to stop" while the conflict resolution happens? The GNW is not a commander, it can't order subsystems around. Though it may be central to consciousness, it's not the "you" that commands and thinks.

All this leaves me thinking that I'm either missing a big obvious chunk of research, or that various motor-planning parts of the brain can't send motor commands except via the GNW. Please point me at any relevant research that you know of.

So AFAIK, the command bottleneck is in the basal ganglia, which are linked to the GNW. A lot of the brain works by lateral inhibition, where each of neurons A, B and C may fire, but any of them firing sends inhibitory signals to the others, causing only one of them to be active at a time.

My understanding from the scholarpedia article is that something similar is going on in the basal ganglia - different subsystems send various motor command "bids" to the BG, which then get different weights depending on various background factors (e.g. food-seeking behaviors get extra weight when you are hungry). Apparently there's a similar mechanism of a strong bid for one system inhibiting the bidding signals for all the others. So if multiple subsystems are issuing conflicting bids at the same time, their bids would end up inhibiting each other and none of them would reach the level necessary for carrying out actions.

Scholarpedia links to Prescott et al. 2006 (sci-hub) as offering a more detailed model, including a concrete version that was implemented in a robot. I've only skimmed it, but they note that their model has some biologically plausible behaviors. In some situations where the robot experiences conflicting high-weight bids, it seems to "react to uncertainty by going half-speed": doing a bit of one one response and then a bit of another, and failing to fully commit to any single procdure:

The control architecture of the robot includes five behaviors, or action sub-systems, which it can switch between at any time. These are: searching for cylinders (cylinder-seek or Cs), picking up a cylinder (cylinderpickup, Cp), looking for a wall (wall-seek, Ws), following alongside a wall (wall-follow, Wf), and depositing the cylinder in a corner (cylinder-deposit, Cd). [...]

three of the five action sub-systems—cylinder-seek, wall-seek, and wall-follow— map patterns of input from the peripheral sensor array into movements that orient the robot towards or away from specific types of stimuli (e.g. object contours). [...]

By comparison, the fixed action pattern for cylinder-pickup in the robot model constitutes five sequential elements: (i) slowly approach the cylinder (to ensure correct identification and good alignment for pickup), (ii) back away (to allow room to lower the arm) whilst opening the gripper, (iii) lower the arm to floor level, (iv) close the gripper (hopefully around the cylinder), (v) return the arm to vertical. [...]

A centralized action selection system requires mechanisms that can assimilate relevant perceptual, motivational, and contextual signals to determine, in some form of ‘common currency’ the relative salience or urgency of each competing behavior (McFarland, 1989; Redgrave et al., 1999a). In the embedding architecture of our model, at each time-step, a salience value for each action sub-system is calculated as a weighted sum of relevant perceptual and motivational variables, and may also be dependant on the current activity status of the action sub-system itself. [...]

... a breakdown of clean selection can occur in the disembodied model when two competitors have high salience levels. To examine the behavioral consequences of this pattern of selection, the robot model was tested over five trials of 120s (approximately 800 robot time-steps) in which the salience of every channel was increased, on every time-step, by a constant amount (C0.4) [...]

During a continuous sequence of high salience competitions the robot exhibited patterns of behavioral disintegration characterized by: (i) reduced efficiency and distortion of a selected behavior, and (ii) rapid switching and incomplete foraging behavior.

Fig. 11 shows the effect of increased salience intensity on exploration of the winner/most-salient-loser salience-space over all trials. The graph demonstrates that virtually all (w4000) salience competitions appeared in the region of salience space (compare with Fig. 7) where reduced efficiency and distorted selection can be expected

The initial avoidance sequence followed the expected pattern but the transition to foraging activity did not begin cleanly, instead showing reduced efficiency and intermittent, partial selection of (losing) avoidance behaviors. To the observer the movement of the robot behavior during the transition appeared somewhat slowed and ‘tentative’. During the foraging bout there was an extended period of rapid switching between cylinder-seek and cylinder-pickup with the robot repeatedly approaching the cylinder but failing to grasp it. The pattern initially observed (t=60–85s) was for the robot to approach the cylinder; back up as if to collect it in the gripper; then move forward without lowering the gripper-arm, pushing the cylinder forward slightly. Later (t=85–90s, 110–115s), where both behaviors showed some partial selection, the robot would lower the arm whilst moving forward but fail to grasp the cylinder due to being incorrectly aligned.

In all five trials, the selection behavior of the robot was similarly inefficient and distorted with the robot frequently displaying rapid alternation of foraging acts. This is illustrated in the transition matrix in Table 2, which shows that the behavior of the robot was dominated by the sequence Cs–Cp–Cs–Cp. with no trials leading to a successful foraging sequence. [...]

Whilst the performance of the model in these circumstances is clearly sub-optimal from a purely action selection viewpoint, it shows interesting similarities to the findings of a large number of studies investigating the behavior of animals in conflict situations (Fentress, 1973; Hinde, 1953, 1966; Roeder, 1975). For instance, Hinde (1966) describes a number of possible outcomes that have been observed in ethological studies of strong behavioral conflicts: (i) inhibition of all but one response, (ii) incomplete expression of a behavior (generally the preparatory stages of behavior are performed), (iii) alternation between behaviors (or ‘dithering’), (iv) ambivalent behavior (a mixture of motor responses), (v) compromise behavior (similar to ambivalence, except that the pattern of motor activity is compatible with both behavioral tendencies), (vi) autonomic responses (for instance defecation or urination), (vii) displacement activity (expression of a behavior that seems irrelevant to the current motivational context, e.g. grooming in a ‘fight or flight’ conflict situation) . Of these outcomes, several show clear similarities with the behavior of the robot in the high salience condition. Specifically, the distortion observed in the early stages of the trial could be understood as a form of ambivalent behavior (iv), whilst the later repetitive behavioral switching has elements of both incomplete expression of behavior (ii) and alternation (iii).

More generally, the behavior of the embodied basal ganglia model is consistent a wide range of findings in psychology and ethology demonstrating that behavioral processes are most effective at intermediate levels of activation (Berlyne, 1960; Bindra, 1969; Fentress, 1973; Malmo, 1959), These findings can also be viewed as expressing the Yerkes-Dodson law (Yerkes & Dodson, 1908) that predicts an ‘inverted U’-shaped relationship between arousal and performance. Our model is consistent with this law in that the robot shows little or no behavioral expression when only low salience inputs are present, demonstrates effective action selection for a range of intermediate level salience inputs (and for high salience inputs where there is no high salience competitor), and exhibits disintegrated behavior in circumstance of conflict between multiple high-salience systems. The robot model, therefore, suggests that the basal ganglia form an important element of the neural substrate mediating the effects of arousal on behavioral effectiveness.

Connecting this with the GNW, several of the salience cues used in the model are perceptual signals, e.g. whether or not a wall or a cylinder is currently perceived. We also know that signals which get to the GNW have a massively boosted signal strength over ones that do not. So while the GNW does not "command" any particular subsystem to take action, salience cues that get into the GNW can get a significant boost, helping them win the action selection process.

Compare e.g. the Stanford Marshmallow Experiment, where the children used a range of behaviors to distract themselves from the sight/thought of the marshmallow - or any situation where you yourself keep getting distracted by signals making their way to consciousness:

The three separate experiments demonstrate a number of significant findings. Effective delay of gratification depends heavily on the cognitive avoidance or suppression of the reward objects while waiting for them to be delivered. Additionally, when the children thought about the absent rewards, it was just as difficult to delay gratification as when the reward items were directly in front of them. Conversely, when the children in the experiment waited for the reward and it was not visibly present, they were able to wait longer and attain the preferred reward. The Stanford marshmallow experiment is important because it demonstrated that effective delay is not achieved by merely thinking about something other than what we want, but rather, it depends on suppressive and avoidance mechanisms that reduce frustration.

The frustration of waiting for a desired reward is demonstrated nicely by the authors when describing the behavior of the children. “They made up quiet songs…hid their head in their arms, pounded the floor with their feet, fiddled playfully and teasingly with the signal bell, verbalized the contingency…prayed to the ceiling, and so on. In one dramatically effective self-distraction technique, after obviously experiencing much agitation, a little girl rested her head, sat limply, relaxed herself, and proceeded to fall sound asleep.”