Synthetic consciousness: the distributed adaptive control perspective

Abstract

Understanding the nature of consciousness is one of the grand outstanding scientific challenges. The fundamental methodological problem is how phenomenal first person experience can be accounted for in a third person verifiable form, while the conceptual challenge is to both define its function and physical realization. The distributed adaptive control theory of consciousness (DACtoc) proposes answers to these three challenges. The methodological challenge is answered relative to the hard problem and DACtoc proposes that it can be addressed using a convergent synthetic methodology using the analysis of synthetic biologically grounded agents, or quale parsing. DACtoc hypothesizes that consciousness in both its primary and secondary forms serves the ability to deal with the hidden states of the world and emerged during the Cambrian period, affording stable multi-agent environments to emerge. The process of consciousness is an autonomous virtualization memory, which serializes and unifies the parallel and subconscious simulations of the hidden states of the world that are largely due to other agents and the self with the objective to extract norms. These norms are in turn projected as value onto the parallel simulation and control systems that are driving action. This functional hypothesis is mapped onto the brainstem, midbrain and the thalamo-cortical and cortico-cortical systems and analysed with respect to our understanding of deficits of consciousness. Subsequently, some of the implications and predictions of DACtoc are outlined, in particular, the prediction that normative bootstrapping of conscious agents is predicated on an intentionality prior. In the view advanced here, human consciousness constitutes the ultimate evolutionary transition by allowing agents to become autonomous with respect to their evolutionary priors leading to a post-biological Anthropocene.This article is part of the themed issue 'The major synthetic evolutionary transitions'.

Keywords: consciousness; distributed adaptive control; memory; neuronal substrate; normativity; social brain.

Figure 1.

A highly abstracted representation of the distributed adaptive control (DAC) theory of mind and brain showing its main processes (boxes) and dominant information flows (arrows). DAC is organized along four layers (soma, reactive, adaptive and contextual) and three columns (world, self, action). The ‘soma’ designates the body with its sensors, organs and actuators. It defines the needs, or self-essential functions (SEF) the organism must satisfy in order to survive. The reactive layer (RL) comprises dedicated behaviour systems (BS) each implementing predefined sensorimotor mappings serving the SEFs. In order to allow for action selection, task switching and conflict resolution, all BSs are regulated via a, so-called, allostatic controller that sets the internal homeostatic dynamics of BSs relative to overall demands and opportunities [28]. The AL acquires a state space of the agent–environment interaction combining perceptual and behavioural learning constrained by value functions defined by the allostatic control of the RL, minimizing perceptual and behavioural prediction error [29,30]. The contextual layer (CL) further expands the time horizon in which the agent can operate through the use of episodic and sequential short- and long-term memory systems (STM and LTM, respectively). STM acquires conjunctive sensorimotor representations assisted by episodic memory as the agent acts in the world. STM sequences are retained as goal-oriented sequences in LTM when positive value is encountered, as defined by the RL and/or AL. The contribution of stored LTM policies to decision-making depends on four factors: goals, perceptual evidence, memory chaining and valence while action selection is further biased by the expected cost of the actions that pertain to reaching a goal state. The content of working memory (WM) is defined by the memory dynamics that represent this four-factor decision-making model. See text for further explanation.

Figure 2.

The DACX model maps the DAC architecture (figure 1) to main brain system and validates the resultant model in a foraging task using a mobile robot. (a) The reactive layer (blue) comprises midbrain/brainstem core behaviour systems (CBS) including trigeminal nucleus (TN), the superior colliculus (SC), central grey (actuation) and the hypothalamus (HP). The adaptive layer (red) includes models of the cerebellum (CRB), amygdala (AMY), ventral tegmental area (VTA) and basal ganglia (BG). The contextual layer (green) includes prefrontal cortex (PFC) and the hippocampus (HPC). All components are realised with synthetic neurons. (b) Top: the foraging robot (10×10 cm at its base) comprises a camera (1), wheels (2), a gripper (3) and five proximity sensors (4). The environment contains a home base (yellow), obstacles (green) and two kinds of rewards (blue and red). The robot can only consume the reward items at the home base and thus has to perform hoarding. Collisions sensed by the proximity sensors are mapped onto TN. Visual states are mapped to AMY/VTA to extract the reward quality from colour, while features extracted from the full image are processed by CRB (for the acquisition of adaptive proximal object related actions such as collision avoidance), HPC (map creation and utilization) and PFC (decision-making). The internal states of the robot are mapped onto HP. Bottom: example trajectories of the agent foraging in the environment at trial 1 (blue) and at trial 12 after reaching a stable hoarding trajectory from between the home base and a red target. (c) The dynamics of the neuronal control system at three distinct stages of behaviour: E1, cue-based navigation; E2, collision avoidance; E3, decision-making and reward delivery. C, colour detection; PR, proximity sensor; M, motor (left: grey, right: purple); TN, ‘pain’ response; SC, orienting response; HPT, internal drive state displaying two competing SEFs; CRBa, cerebellar circuits associated with avoidance responses; CRBo, cerebellar circuits for acquired approach responses. E1: Adaptive cue-based navigation, during exploration of a visual cue (C) triggers an adaptive approach response in the cerebellum (CRBa) conveyed to the motors (M) which partially overwrites the reactive orienting response triggered by SC, allowing the agent to turn towards a resource location. E2: Acquired obstacle avoidance: a proximity sensor signal (PR) allows the robot to efficiently prevent collisions with obstacles by being associated with an avoidance response (CRBo). The unconditioned stimulus, i.e. collision, associated response in the TN is partially removed due to the peripheral disruption of the US. E3: Decision-making and reward delivery, when the agent reaches the home location, visually identified by a landmark (yellow patch) as well as represented by the acquired internal spatial representation (HPC), a reward value associated with the hoarded item is delivered (STR, green) and the allostatic value for the related internal state encoded in HPT (blue), increasing its values and, consequently, decreasing the motivation to pursue that type of resource encoded in VTA (green). Activity in the VTA module directly affects the decision-making process (PFC) performed at the beginning of the trial, biasing the decision towards the most urgent need (PFC, blue).

Figure 3.

Hemispatial neglect. (a) Standard drawing test performed by a patient (male, 64 years) with a right haemorrhagic stroke showing a reduced ability to fill the left side of a workspace with crosses. (b) A visual search task where subjects have to detect a local (top), global (middle) or combined (bottom) distractor in a visual display, which occurred with 50% probability in 72 trials. The distractors are indicated with a circle. The global distractor is a Kanizsa triangle. (c) Performance of control subjects and right hemispheric (RH) stroke patients in terms of accuracy (left column) and reaction time (RT; right column). For visibility, the RT of the control group is displayed in a reduced range, which is reflected with a dashed box in RT plots for RH. Asterisks indicates a significant difference at p < 0.01. N_control: 10; N_RH: 8. Adapted from Campillo et al. [124]. See text for further explanation.

Figure 4.

The validation gate hypothesis and the dynamic coupling of consciousness to tasks and sensor states. (a) The validation gate hypothesis of perception [178] proposes that information seeking is guided by predictions in a dual form by defining both regions in input space that are expected to provide data and those that do not and thus do not need to be scrutinized. If we follow dots moving along linear trajectories, data (orange dots) is classified relative to areas in input space where it is expected to occur given the properties of stimuli, or their validation gate (light blue area). Resources are only allocated to data which falls outside of the validation gates or when validation gates overlap, i.e. resolving novelty and ambiguity respectively. (b) Using a displacement detection task together with reverse correlation allowed us to exactly define the modulation of the validation gates by cognitive load in humans. Increased cognitive load induced a distinct increase of the eccentricity of the validation gate of consciously detected displacements, indicating an expansion of the area that was ignored in sensory processing due to a secondary working memory task (right-hand side kernel). The kernels are displayed as probability distribution of displacements that were followed by a button press from low (left) to high (right) cognitive load. The kernels of fast and slow eye movements did not show a similar effect of cognitive load. Adapted from Mathews et al. [179]. (c) An fMRI analysis of the validation gate displacement task showed that the explicit detection of displacements were correlated with activation of fronto-parietal networks involving middle and right superior frontal gyrus (Brodmann area 10, 11, yellow), right anterior cingulate cortex (ACC, BA32, yellow) and left precuneus (BA7, orange) (data from Malekshahi et al. [180]). These areas are projected onto a three-dimensional reconstruction of the human connectome using brainx3, visualizing the structural and functional connectivity between the nodes (green) of the conscious detection network [181,182].

Figure 5.

Conceptual diagram of the neuronal substrate of the consciousness autonomous virtualization normative memory system (CVNM). The core behaviour systems (CBS) of the midbrain provide the substrate for primary consciousness driving both the neocortex (1) and the brainstem activating system (BAS) that in turn provide the activation and valuation of the thalamus (2) and neocortex (3). The neocortex (CRT) shares direct and indirect recurrent connections with the thalamus (4) and normative virtualization memory of secondary consciousness is implemented by the cortico-cortical and thalamo-cortical system of the fronto-parietal network (4,5). Cortical networks in turn can exert control over BAS (6) and CBS (1) closing a normative valuation feedback loop. In addition, frontal cortex can inhibit input to sensory thalamic nuclei through projections to the reticular nucleus of the thalamus imposing validation gate driven ‘blindness’ for task irrelevant states (7). See text for further explanation.

Figure 6.

Quale parsing in DACX. (a) Example trajectory of a trained DACX agent, showing initial position (left upper corner) and three salient events in the arena where coloured objects are encountered as seen through the robot's camera (locations P1: red, P2: green, P3: blue). (b) (i) Population activity in the CA3 episodic memory system of the agent during the three salient navigation events. Colour code as in (a). Insets: Rate maps of model single place cells at P1, P2 and P3 showing their spatial specificity. (ii) DACX mind travel by means of a memory sweep showing the goal-oriented engagement of three place cells due to the PFC triggered spreading of activation through the associative connections among the place cells acquired during the exploration of the arena. The actual position of the agent in the upper left corner of the arena is signalled by the first left-most place cell. The drive state was ‘hunger’ while the goal state was a ‘food’ item. The right-most panel gives the projected imagined trajectory based on a shortest path search between the activated place cells. Hence, we have parsed a ‘hunger’ induced, imagined trajectory which leads via two intermediate positions to a location where the agent expects to satisfy its ‘food’ goal state.