representation internal-manipulation (rim): a neuro ... · putational concepts, have been...

Representation Internal-Manipulation (RIM):A Neuro-Inspired Computational Theory of Consciousness

Gianluca Baldassarre1

Laboratory of Computational Embodied Neuroscience,Institute of Cognitive Sciences and Technologies, National

Research Council of Italy, Rome, Italy

Giovanni Granato1

Laboratory of Computational Embodied Neuroscience,Institute of Cognitive Sciences and Technologies, National

Research Council of Italy, Rome, Italy

Many theories, based on neuroscientific and psychological empirical evidence and on com-putational concepts, have been elaborated to explain the emergence of consciousness in thecentral nervous system. These theories propose key fundamental mechanisms to explain con-sciousness, but they only partially connect such mechanisms to the possible functional andadaptive role of consciousness. Recently, some cognitive and neuroscientific models try tosolve this gap by linking consciousness to various aspects of goal-directed behaviour, the piv-otal cognitive process that allows mammals to flexibly act in challenging environments. Herewe propose the Representation Internal-Manipulation (RIM) theory of consciousness, a theorythat links the main elements of consciousness theories to components and functions of goal-directed behaviour, ascribing a central role for consciousness to the goal-directed manipulationof internal representations. This manipulation relies on four specific computational operationsto perform the flexible internal adaptation of all key elements of goal-directed computation,from the representations of objects to those of goals, actions, and plans. Finally, we proposethe concept of ‘manipulation agency’ relating the sense of agency to the internal manipulationof representations. This allows us to propose that the subjective experience of consciousnessis associated to the human capacity to generate and control a simulated internal reality thatis vividly perceived and felt through the same perceptual and emotional mechanisms used totackle the external world.

1The authors have equally contributed to the paper.

Introduction

Scientific theories of consciousness often focus on spe-cific key mechanisms possibly relevant for it such as the neu-ronal networks that support conscious processing (Dehaene& Changeux, 2011; Dehaene, Kerszberg, & Changeux, 1998;Dehaene & Naccache, 2001), their informational integra-tion (Koch, Massimini, Boly, & Tononi, 2016; Tononi, 2008;Tononi, Boly, Massimini, & Koch, 2016), the brain hierar-chies (Damasio, 1989; Meyer & Damasio, 2009), and thesensorimotor interactions involving brain, body, and the en-vironment (O’Regan & Noe, 2001; O’Regan, Myin, & Noë,2005).

Similarly to Baars’ approach stressing the possible func-tional elements and adaptive role of conscious cognition(Baars, 1997, 2005; Baars, Franklin, & Ramsoy, 2013;Baars, Ramsøy, & Laureys, 2003), our approach to con-sciousness strongly focuses on the importance of consider-ing the functional (adaptive) aspects of consciousness along-side the possible mechanisms that might support them, inte-grated to form a coherent system. In particular, here we pro-pose a strong link between the main features of conscious-ness highlighted by the aforementioned theories and the pro-

cesses and mechanisms involved in goal-directed behaviour(Balleine & Dickinson, 1998; Hills, 2006; Mannella, Gur-ney, & Baldassarre, 2013). Goal-directed processes allowcomplex organisms to adapt to novel challenges of the envi-ronment through complex forms of mental simulation (Ger-lach, Spreng, Gilmore, & Schacter, 2011; Pezzulo, Rigoli,& Chersi, 2013) allowing them to plan actions ahead, tosupport complex decision making, and to solve challengingproblems. Despite various authors mention possible links be-tween consciousness and goal-directed processes (e.g., Ben-gio, 2017; Chiappe & MacDonald, 2005; Keller, 2014; Pen-nartz, 2017; Ressler, 2004; Rossano, 2003), how the compu-tational processes underlying goal-directed behaviour mightbe specifically related to conscious processing remains elu-sive.

Here we propose the Representation Internal-Manipulation (RIM) theory of consciousness. The theoryposits that consciousness emerges within the goal-directedsystem of brain pivoting on this key main componentsand processes: a set of working memories mediatingperception and imagination, an abstract working memoryrepresenting the main elements of goal-directed processes,a selection process (‘internal manipulator’) manipulatinginternal representations at all levels, and a motivational

arX

iv:1

912.

1349

0v1

[cs

.AI]

31

Dec

201

9

2 GIANLUCA BALDASSARRE1

system driving the manipulator. In this proposal, a centralrole for consciousness is played by the goal-directedmanipulation of internal representations. This manipulationprocess performs four classes of computational operationson internal representations: decomposition, composition,abstraction, and specification. The key adaptive functionplayed by consciousness through these operations is toflexibly modify as desired all key elements of goal-directedprocesses, from the representations of objects and contextsto those of goals, actions, and plans. The theory allows theintegration of the key elements proposed by the theories onconsciousness while at the same time ascribing to them afunctional/adaptive valence.

Finally, we relate the internal manipulation of representa-tions to the sense of agency. This allows us to contribute toface the ‘hard problem of consciousness’ (Chalmers, 1995),in particular to propose a possible explanation of the subjec-tive experience that accompanies conscious processes (Den-nett, 1988; Haggard, 2017). The contribution rests on a novelconcept that we formulate here, of ‘manipulation agency’.The idea is that the internal manipulation system under-lying consciousness allows humans to actively and flexi-bly generate, control, and change as desired an internalsimulated/imagined reality. The sense of agency given bythis flexible and powerful control, together with the vividlysensed and felt experience of the imagined reality that is pro-cessed through the same perceptual and emotional mecha-nisms used to tackle the external world, gives rise to the sub-jective experience accompanying consciousness.

Overview of some theories of consciousness

Many theories have been proposed to explain conscious-ness as a propriety emergent from specific mechanisms ofbrain. Here we consider some major theories that proposearticulated systems of concepts on consciousness that give areasonable coverage of the relevant ideas on consciousnessalso developed in other theories. We start from the theoryof Baars that has the merit to more strongly aim to proposea functional interpretation of conscious processes, i.e. whichcognitive functions useful for adaptation these might play, anaim that we also follow here.

The global workspace theory of Baars (GWT, Baars 1997)proposes that consciousness relies on a set of fundamentalelements that might metaphorically correspond to the ele-ments of a theatre, the ‘theatre of consciousness’. The fun-damental elements are the conscious contents such as per-cepts, thoughts, and judgements (‘actors in the stage’), theglobal workspace of working memory where conscious con-tents are represented (‘theatre stage’), the selective attentionthat allocates limited processing resources to specific con-tents (‘theatre spotlight’), the executive functions (‘director’)that, on the basis of the agent’s current goals, decide whichcontents to activate, the unconscious background processes

(‘audience’), that contribute to furnish row information tothe conscious contents (e.g. contextual information). Theauthor recently grounded the model on neuroscientific evi-dence (Baars, 2005; Baars et al., 2013, 2003), in particularby linking the ‘theatre elements’ with neural structures andprocesses. Baars’ theory follows an approach that best fitsour functional approach to consciousness, and in this respectit can be considered a precursor of our proposal.

A related biologically-grounded theory, the neuronalworkspace model (NWM; Dehaene and Changeux 2011),proposes the existence of two computational sub-spaces ofbrain networks to explain the emergence of conscious expe-rience. A first sub-space is formed by many specialised func-tional modules (e.g., sensory areas or motor systems) char-acterised by high-density short/medium range connections.A second sub-space is formed by a distributed set of neu-rons, communicating trough excitatory long-range projec-tions, involving the associative cortices (prefrontal, parieto-temporal, and cingulate cortex) and including the fibers ofcorpus callosum and of the cortico-thalamic system (De-haene & Changeux, 2011). The NWM has been extendedto envisage the existence of buffers (working memories) be-tween the sensorial cortices and the neuronal workspace(Raffone, Srinivasan, & van Leeuwen, 2014, 2015). Basedon these anatomical organisation, the model proposes thatthe reverberating networks formed by the top-down projec-tions (second interconnected computational space) and thebottom-up projections (first computational space) implementspecific selective processes that lead to the emergence of arestricted set of relevant contents within consciousness.

Another influential theory of consciousness, the inte-grated information theory (IIT; Tononi 2008; Tononi et al.2016), proposes that a high anatomical and functional in-tegration of a key brain sub-network, namely the thalamo-cortical system, represents a critical element of conscious-ness. In particular, this theory proposes that the represen-tations of the thalamo-cortical system are characterised by‘specificity’, i.e. an substantial level of differentiation andsegregation with respect to other competitive representations,and ‘integration’, i.e. a relevant level of information ex-changed by the units that compose the same representation.A major strength of the theory is the proposal of a quanti-tative measure of the level of such specificity and integra-tion, namely the Φ value. The theory postulates that a highΦ value represents the neural correlate of a subjective con-sciousness state. A recent update of the theory (Koch et al.,2016) distinguishes a brain hot zone that supports the forma-tion of conscious contents localised in the posterior corticesand a frontal-parietal system that supports task-monitoringprocesses, making conscious contents coherent.

Damasio (Damasio, 1989; Meyer & Damasio, 2009)has proposed another relevant theory, the convergence-divergence zones theory (CDZt), that focuses on the hierar-

CONSCIOUS MANIPULATION OF INTERNAL REPRESENTATIONS 3

chical organisation of brain systems to explain the differencebetween conscious and subconscious representations. In par-ticular, this theory postulates the existence of minor conver-gence zones, corresponding to the sensory cortices, and amajor central convergence zone, corresponding to associa-tive cortical areas (prefrontal cortex, parietal cortex, and tem-poral cortex), that together form a hierarchy of convergencezones. The theory proposes that this architecture supportsa progressive integration of information from perceptual tocentral areas that supports perceptual processes, and an in-verse flow of information that supports imaginary processes.The theory also focuses on the emotional dimension of con-scious experience by referring to the somatic marker theoryto explain the emotional nuance of conscious representations(Bechara & Damasio, 2005). In particular, the term ‘somaticmarker’ refers to the influence, taking place within the majorassociative convergence zones, of somatic signals onto brainrepresentations that thus receive a prioritisation bias for be-ing consciously processed (Verdejo-García, Pérez-García, &Bechara, 2006).

O’Regan (O’Regan & Noe, 2001; O’Regan et al., 2005)proposed a sensorimotor theory of consciousness that, inline with the approaches of embodied cognition (Ander-son, 2003; Borghi & Cimatti, 2010; Garbarini & Adenzato,2004) and enactivism (Hutto, 2005), goes beyond strictlycognitive processes to involve the close interaction betweenbrain, body, and environment. In particular, the theory ex-plains the features of conscious visual experience, proposingthat ‘sensorimotor contingencies’ (the perception of worldchanges caused by actions) are fundamental features for theemergence of conscious experience. Furthermore, this the-ory explains the difference between a sensory experienceand other types of neural processes (e.g. imagination orreasoning) suggesting the existence of two properties ofaction-contingent stimuli defined ‘allerting’ and ‘corporal-ity’. While allerting corresponds to the capacity of chang-ing stimuli to exogenously attract our attention, corporalitydescribes a phenomenal change of perception that is conse-quent to a motor act of the agent.

The goal-directed system

Goal-directed behaviour relies on sophisticated internalrepresentations, such as the representations of objects, goals,and world models, and is involved in many higher order cog-nitive processes such as planning, reasoning, and problemsolving. In this section we present contributions from differ-ent disciplines, in particular neurobiology, neuropsychology,psychology, and computational literature, that clarify rele-vant aspects of the processes that underlie goal directed be-haviour in brain, behaviour, and artificial intelligence sys-tems. The contributions are in part overlapping; they are alsoso wide that here we can only consider few main aspects ofthem.

Biological contributions

Given a specific condition of the environment, goal-directed behaviour allows an agent to exhibit different be-haviours depending on the goal that it intends to pur-sue. In particular, the behaviour is not automatically trig-gered by specific stimuli, as in habitual behaviour relyingon stimulus-response (S-R) associations (Yin & Knowlton,2006). Rather, it is guided by the future desired environmentstates on the basis of previously acquired action-outcome (A-O) associations (Balleine & Dickinson, 1998). Goal-directedbehaviour, initially studied in animal species such as rats,is phylogenetically more recent than habitual behaviour andrelies on the evolutionary emergence of complex brain pro-cesses (Passingham & Wise, 2012). Among these, planningprocesses allow animals to flexibly assemble sequences ofactions in novel ways (Delatour & Gisquet-Verrier, 2000;Pfeiffer & Foster, 2013), reasoning processes allow the infer-ence of new knowledge from previously acquired informa-tion (Aust, Range, Steurer, & Huber, 2008; Blaisdell, Sawa,Leising, & Waldmann, 2006), and problem solving processesallow the accomplishment of new goals by adjusting previ-ously acquired knowledge (Newman, Carpenter, Varma, &Just, 2003).

We now consider some main aspects of the brain goal-directed system (for more details, see Baldassarre, Caligiore,& Mannella, 2013; Caligiore, Arbib, Miall, & Baldassarre,2019; Cisek & Kalaska, 2010).

States and the cortical pathways. Brain processes per-ceptual information through multiple hierarchies of corticalareas partially segregated by sensory modality, e.g. for hear-ing or vision (Felleman & Van Essen, 1991). For example,visual information is processed from early occipital cortex,performing basic information processing such as the detec-tion of image edges, towards higher associative cortices, pro-cessing information for object identification in the infero-temporal cortex (‘ventral visual pathway’) or for supportingthe interaction with object based on and supporting objectinteraction, e.g. related to space location and size encodedin the parietal cortex (‘dorsal visual pathway’; Goodale &Milner, 1992). These areas support action performance bysending information to motor cortices (Rizzolatti, Luppino,& Matelli, 1998). They also send information to the multi-modal integrative areas of prefrontal cortex (Passingham &Wise, 2012), which then exert a top-down control both overmotor areas and backward towards the perceptual cortical hi-erarchies to support attentional processes.

Goals and frontal cortex. Goals represent the pivotalcomponents of goal-directed behaviour that distinguish itfrom habitual behaviour. In particular, within the behaviouralpsychology literature a goal is defined as an action out-come having biological value, such as a food pellet, that candrive the performance of a behaviour for the balancing ofhomeostasis (i.e for the satisfaction of an homeostatic drive;


Balleine and Dickinson 1998; Berridge 2004). More in gen-eral, in cognitive sciences a goal is defined as an internalrepresentation of a desirable future state of the world. Thisrepresentation can be internally re-activated, in the absenceof its referent in the environment, for triggering actions di-rected to accomplish the world state corresponding to thegoal (Thill, Caligiore, Borghi, Ziemke, & Baldassarre, 2013).A goal representation (e.g., the internal representation of afood pellet) can be supported by different sensory modalities,as the smell and visual appearance, and by different hierar-chical processes with an increasing level of abstraction andintegration of sensory modalities (Mechelli, Price, Friston,& Ishai, 2004). In the human brain, goals are formed withinthe orbitofrontal cortex and ventromedial prefrontal cortexand are given motivational valence by lower-level represen-tations within amygdala and nucleus accumbens (Mannellaet al., 2013; Mirolli, Mannella, & Baldassarre, 2010; Roy,Shohamy, & Wager, 2012). Dorsolateral prefrontal cortexalso encodes the more complex and abstract rules/situationsthat are instrumental for the attainment of biologicallyrele-vant outcomes (Tsujimoto, Genovesio, & Wise, 2011).

Selection and the basal ganglia. A key brain mecha-nism supporting the selection of goal-directed behaviour el-ements, such as states, goals, and actions, is implemented bybasal-ganglia (Redgrave, Prescott, & Gurney, 1999). Theseform basal ganglia-thalamo-cortical loops that originatefrom the associative and frontal regions of cortex, and feed-back onto the same regions via the thalamus. These loops arepartially segregated into parts specialised to support selectionof cortical contents involving in particular (Yin & Knowlton,2006): motor cortices; associative cortices (parietal, tempo-ral, and dorsal prefrontal cortex); and limbic cortices (orbitaland ventromedial preforntal cortex). An overall fundamentalpattern of brain is the one where the cortical pathways dis-cussed above, hosting bottom-up and top-down informationflow, are ‘intercepted’ by the basal-ganglia loops that per-form selection of contents within them (Baldassarre, Cali-giore, & Mannella, 2013). We will see that this pattern isvery important for the internal manipulation of informationof the RIM theory proposed here.

Motivations. Motivations represent another componentof GDB, strongly related to the formation and the internalactivation of goals. In particular motivations are internalprocesses that, based on the organism’s ‘needs’, guide theselection of certain behaviours and the modulation of otherinternal processes such as attention and learning. There arevarious types of motivations, but in general literature dividedthem in extrinsic (Panksepp, 1998) and intrinsic motivations(Baldassarre, 2011). Extrinsic motivations refer to the basicphysiological motivations linked to survival and the repro-duction such as pain, hunger, and thirst (Panksepp, 1998).These motivations can lead to form goals, for example in theform of visual/auditive internal representations of the envi-

ronment resources that can satisfy them (e.g. food or water).In brain, extrinsic motivations are supported by areas such asthe hypothalamus and amygdala and involve reward-relatedsystems such as the dopamine system (Schultz, 2002).

Intrinsic motivation, endowed with a more sophisticatedcognition (Baldassarre & Mirolli, 2013; Deci, Vallerand, Pel-letier, & Ryan, 1991; Fantz, 1964; White, 1959) drive theacquisition of knowledge that might be later employed forthe acquisition of biological relevant resources and are re-lated to processes such as curiosity, exploration, play, nov-elty, surprise, the setting of own goals and the competence toaccomplish them. Intrinsic motivations involve various brainstructures, for example hippocampus for novelty detection(Kumaran & Maguire, 2007; Lisman & Grace, 2005) or theanterior cingulate cortex for surprise (O’Reilly et al., 2013;Paus, 2001). Also intrinsic motivations can support the for-mation of goals (Baldassarre, Mannella, et al., 2013).

Psychological contributions

Attention. Attention is a relevant process for goal-directed behaviour Attention processes allow the acquisitionof relevant information while discarding irrelevant stimuli(selective attention; Treisman & Gelade, 1980). They arealso fundamental for the formation of goals formation andtheir internal re-activation. Attention can be divided in covertvs. overt attention (Lavie, Hirst, de Fockert, & Viding, 2004;Rizzolatti, Riggio, Dascola, & Umiltá, 1987), endogenous(‘bottom-up’) vs. exogenous (‘top-down’) attention (Posner& Petersen, 1990).

Overt attention refers to the selection of environment in-formation based on the control of sensors focusing on a par-ticular portion of the environment or stimuli (e.g. the viewof a food pellet regardless of the context where it is found).Covert attention refers to an internal selection process thatcan mentally isolate the relevant parts of an internal repre-sentation (e.g. to internally focus on a certain portion of animagined scene). Both covert and overt attention can supportthe formation of focused representations of states, goals, andmodels of the world.

Regarding the second distinction, exogenous attentionrefers to externally-driven shifts of the attention focus,mostly caused by highly salient stimuli or stimulus changes,while endogenous attention involves a voluntary shift drivenby the agents’ goals (Ognibene & Baldassarre, 2015). En-dogenous attention is very important for goal-directed be-haviour because it allows the internal activate of desiredgoals and also support reasoning processes and planning pro-cesses that rely on a serial scan of mental contents. Theseprocesses strongly rely on top-down driven sensory imagi-nation processes (Mechelli et al., 2004; Seepanomwan, Cali-giore, Cangelosi, & Baldassarre, 2015).

Planning. Given a selected goal, planning processes usethe knowledge on action-outcome contingencies to search a


single action or an action sequence that leads to accomplishthe desired goal (Delatour & Gisquet-Verrier, 2000; Pfeif-fer & Foster, 2013). In humans, this searching process issupported by an ‘internal simulation’ assembling and test-ing actions sequences ‘within the mind’ rather than by actingin the environment in humans (Gerlach et al., 2011). Theseprocesses might also involve non-human animals (Balleine& Dickinson, 1998; Clayton & Dickinson, 2010). Planningprocesses are also supported by bottom-up mechanisms suchas affordances that suggest the possible actions that could beuseful and appropriate for a given condition or object (Gar-barini & Adenzato, 2004; Rounis & Humphreys, 2015; Thillet al., 2013).

Problem solving. Problem solving is a sophisticatedversion of planning, considered a fundamental element of hu-man intelligence in early cognitive-science studies (Simon,1975). In early views, problem solving identifies the a flex-ible search of the correct sequence of actions to pass froman initial state to a final desired state (goal), so is analogousto planning. This view still has an important role in currentartificial intelligence models and cognitive science interpre-tations of human problem solving (e.g. Condell et al. 2010;Waltz et al. 1999), but it overlooks the fact that life prob-lem solving requires not only to find suitable sequences ofactions to achieve the goal, but also to face knowledge gapsregarding the elements of planning, for examples the relevantobjects, their functions, the actions to act on them, and thepossible effects of actions (world model).

Neuro-psychological contributions

Executive functions. The neuro-psychology literaturestudies a set of processes defined ‘executive functions’ (Di-amond, 2013; Miyake et al., 2000) that have a notable over-lap with goal-directed behaviour processes studied in the bi-ological literature but also add useful information. Executivefunctions support a coherent feedback-dependent goal-basedtop-down control of behaviour and rely on three fundamen-tal processes, i.e. working memory, inhibitory control, andcognitive flexibility.

Working memory. Working memory, also studied inpsychology, is a fundamental process for goal-directed be-haviour referring to the capacity of the brain to keep ac-tive information representations for prolongued periods oftime. Without the persistent maintenance of informationby working memory cognition could support only sensori-motor behaviours. Working memory can store informationin a modality-specific form (Fiehler, Burke, Engel, Bien,& Rösler, 2008; Raffone et al., 2014) or in an highly inte-grated multimodal form (Baddeley, 2000; Wolters & Raf-fone, 2008). Working memory is a fundamental process forthe selection and active maintenance of goals and sub-goals,and for supporting planning and problem solving processesthrough the maintenance of representations regarding past or

future world states and their possible causal/temporal links(world models).

Inhibitory control. Working memory processes have asynergistic interaction with inhibition processes, another fun-damental key element underlying goal-directed behaviour.Inhibitory control allowing the focus on the selected goalsby performing a continuous suppression of distractors, suchas other goals, intervening stimuli, and irrelevant actions(Durston et al., 2002).

Cognitive flexibility. The integrated interaction of work-ing memory and inhibitory control allows the emergence ofa third process defined, cognitive flexibility, related to thecapacity to switch behavioural strategy depending on the ex-ternal conditions. For example, in a . The integrated func-tioning of working-memory, inhibitory control, and cogni-tive flexibility allows the emergence of high-order forms ofexecutive functions such as goal monitoring, planning, andproblems solving.

Studies regarding the neural correlate of executive func-tions (Goldman-Rakic, 1996; Hartley & Speer, 2000; Robin-son, Calamia, Gläscher, Bruss, & Tranel, 2014) highlightmany structures that support goal-directed behaviour. Inparticular, working memory is suppo,rted by basal ganglia-cortical loops, hippocampus and high-order prefrontal cor-tices while inhibitory control mostly relies on basal-ganglia,ventromedial prefrontal cortex and anterior cingulate cortex(Durston et al., 2002). Cognitive flexibility relies on a com-plex network involving various portions of basal-ganglia,prefrontal cortex, anterior cingulate cortex, and parietal cor-tex (Leber, Turk-Browne, & Chun, 2008). The monitoring ofthe attainment of sub-goals and final goals relies on frontalportions of prefrontal cortex (Desrochers, Burk, Badre, &Sheinberg, 2015) while the anterior cingulate cortex showsa strong involvement in the detection of the violation of ex-pectations (Benn et al., 2014).

Artificial intelligence contributions

Research on artificial intelligence, that initially wasclosely linked to the study of natural intelligence, gave animportant theoretical foundation to the study of goal-directedprocesses (Russell & Norvig, 2016). A first relevant con-tribution was the one on problem solving introduced above,positing that human intelligence is primarily directed to ac-complish goals by searching suitable sequences of actionsleading from an initial state to a final goal state passingthrough a number of intermediate sub-goals (Simon, 1975).These processes involved the representation of the worldstates as whole elements, such as the configuration of thepieces on the board in the chess game.

Later this approach was developed by the research onplanning that made the search of action sequences more effi-cient by factorising the representation of states into elements(e.g. ‘objects’) and relations between elements (e.g., ‘being


on’, ‘being part of’) suitably expressed with logic symbols(Russell & Norvig, 2016). On this basis, planning systemscould hence focus and reason on fewer relevant elements andrelations at a time.

This thread of research mostly proceeded in parallel withrespect to the research on neural networks (McClelland & thePDPResearchGroup, 1986), which has lately arrived to thegreat success of machine learning and deep neural networks(Goodfellow, Bengio, & Courville, 2017). This success isnow bringing new interest and energies to face the high-levelproblems tackled by planning and problem solving by us-ing approaches based on neural networks (e.g., Baldassarre,2001; Wayne et al., 2018), a trend that is expected to delivergood outcomes also to the research on consciousness (e.g.,see Bengio, 2017).

The Representation Internal-Manipulation (RIM)theory of consciousness: components and manipulation

operations

In this section we present the Representation Internal-Manipulation (RIM) theory of consciousness, in particularthe main ‘components’, or set of functions, that the theoryposits to be at the core of consciousness processes (Figure 1).As the brain works as a whole, these ‘components’ should bethought to have a limited segregation in brain, i.e. to havea relative overlap within the brain structures that implementthem (Caligiore et al., 2019).

(1) Perceptual working memory. This component is re-sponsible for a bottom-up information processing supportingmodal perception (visual, auditive, tactile, somatosensory,etc.) and the formation of increasingly abstract representa-tions, e.g. it supports the processing of visual images fromthe retina and the formation of increasingly abstract repre-sentations, from low-level visual features (edges, corners,etc.) to higher level representations (more complex features,objects, and faces). The same component also supports atop-down information flow moving in the opposite directionand supporting imagination, i.e. the central re-activation ofperipheral perceptual representations. The component, sup-ported in brain by reverberating loops involving sensory cor-tices, has the capacity to implement modal working memo-ries able to maintain active perceptual representations havinga high level of details and possibly several dimensions. Theseactivations also represent the conscious contents of which weare aware.

(2) Abstract working-memory. This component sup-ports the active maintenance of goals and behavioural strate-gies in more abstract multimodal formats. This component,mostly supported by fronto-striatal circuits, forming rever-berating loops with thalamus and basal ganglia, has the ca-pacity of maintaining abstract representations of the worldelements (context, objects, other agents), and of goals andaction plans. These representations can be activated or de-

activated as needed on the basis of the current condition, thegoals to pursue, and the plans, decisions, and thoughts for-mulated to accomplish them.

(3) Top-down internal manipulator. This component,guided by motivationally-charged active goals, activates spe-cific contents within the abstract working memory and theperceptual working memories, in particular by ‘manipulat-ing’ them so to best use the agent’s knowledge to pursuethe currently desired goals. This component is actually fullyembedded into the other three components. In particular, itis implemented by the biases from the motivational compo-nent, and by the selection processes internal to the abstractand perceptual working memories. The ‘component’ in par-ticular relies on the excitatory fronto-parietal cortical path-ways for the top-down guidance, on the disinhibition mecha-nisms of basal ganglia for performing macro selections, andon the local inhibitiory circuits of cortex for performing mi-cro selections. The component allows the conscious flexi-ble manipulation and morphing of externally-driven percepts(e.g., visual or somatosensory representations) and of self-activated representations (imagination). The internal manip-ulator represents the keystone of the RIM theory.

(4) Motivation system. This component, supported inthe brain by sub-cortical emotional systems such as the hy-pothalamus, amygdala, hippocampus, and the nucleus ac-cumbens, furnishes the motivational valence to the represen-tations within the other components. In particular, it supportsthe selection/activation of specific goals within the abstractworking memory, thus forming ‘intentions’, and also sup-ports the selection of other relevant elements, for exampleenhances the activation of representations related to objectsand sub-goals relevant to support intentional behaviour.

The four components support the internal manipulationof representations proposed here to be at the core of con-sciousness. The manipulator performs operations that can begrouped into four classes (Figure 2). We extracted these fourclasses of operations from neuro-psychological and neuro-scientific literature, for example on executive functions, at-tention, the selective processes of basal ganglia, and the hi-erarchical organisation of cortex. We think these operationsare the main means through which conscious processing sup-ports flexible goal-directed planning and problem solving.The four classes of operations are as follows.

(1) Decomposition. This operation, linked to the con-cept of selective attention, allows the separation of wholerepresentations into pieces, thus allowing the agent to focuson single elements (e.g., objects in the world or specific sub-goals) and their parts. This operation supports the focusingon specific portions of representations based on spatial reg-ularities (space-based) or functional/visual regularities (e.g.,object-based).

(2) Composition. A ‘chunk’, or ‘composition’, is agroup of elements glued together by a higher-level seman-tic link (e.g., the eyes, nose, and mouth together form a


Neuro-Functional Elements of the Representation Internal-Manipulation (RIM) Theory

Figure 1. Schema showing the components of the RIM theory of consciousness. The colour gradient (red to blue) indicatesa gradual change of computational functions from those encoding emotional/motivational elements (e.g. goals; fronto-striatalcortex) to those encoding perceptual elements (e.g. visual activations; occipital and dorso-caudal cortex)

face). Neurally, a chunk can be implemented by connectionsthat link the neural representations of the different elementsand/or connections that link them to another neural repre-sentation of the semantic link. The connections allow theactivation of the neural representation of the semantic linkwhen needed, and to use it in place of the representationsof the elements of the composition. Alongside such struc-tural chunks, the composition also allows the formation ofchunks through a functional connectivity, for example basedon the synchronous firing of spiking-neurons encoding theitems of to be chunked. Composition supports the assem-bling of items relevant for many goal-directed processes, forexample the creation of action sequences forming a plan, orto consider wholes useful for problem solving.

(3) Abstraction. This operation allows the activation ofinternal representations that have less details or dimensionswith respect to some original representations. There existmultiple possible abstractions of a pattern depending on theinformation that is preserved or discarded, so the process ofabstraction needs to be biased in a specific direction. Ab-straction is obtained by several features contributing to formmore complex features or to identify the ‘hidden causes’ ofgiven patterns. Abstraction is at the basis of the capacity of

humans to consider specific aspects of objects while ignor-ing others, for example to focus on the colour, or the size, orthe shape of an object. Furthermore, abstraction is stronglylinked to the capacity to create simplified/compact represen-tations of the world.

(4) Specification. Specification represents the inverseoperation with respect to abstraction, in particular it addsdetails and dimensions to original abstract representations.Since specification implies few-to-many mappings, it re-quires that the process of specification is biased towards adesired direction. Neurally, specification is implemented bypassing from abstract to more specific representations, forexample by passing from hidden causes to full representa-tions, for example to imagine some details of the face of aknown person. This operation, together with composition,is thus behind the capacity of adding details to ideas and ofimagining and creating new elements.

Conscious Knowledge Transfer (CKT). The integratedfunctioning of the four operations give rise to a super-ordinate function called here conscious knowledge transfer.CKT is an enhanced version of generalisation, intended hereas the capacity of neural networks to respond in similar waysto similar stimuli; generalisation is in particular based on the


interpolation between previous experiences or the extrapo-lation very proximal to such experiences. For example, anagent might experience an object on various positions at thecentre of the table and on this basis imagine it a bit outsidesuch centre (Nair et al., 2018). Instead, CKT implies thecapacity of extrapolating knowledge far beyond previous ex-periences, e.g. to imagine an object anywhere in a knownspace. Moreover, CKT involves all types of flexible activeconstruction of knowledge that, starting from experience, canbuild fully new knowledge on the basis of the regularities ex-tracted from such knowledge, for example those related tothe physical properties of the environment, and to natural-istic and social knowledge (Lake, Ullman, Tenenbaum, &Gershman, 2017). This construction of knowledge relies onknowledge building processes such as analogy (Gentner &Forbus, 2011), induction, and other types of inference.

All these processes allow conscious agents to flexibly‘warp and morph’ all elements of planning, such as contexts,objects, goals/sub-goals, and actions, in order to accomplishthe desired goals in the lack of sufficient knowledge (flex-ible problem solving). For example, a goal might be splitin never-experienced sub-goals trough decomposition; a newfunction of an object might be found by ‘importing’ it fromanother object through abstraction (which allows to see simi-larities), and by linking it to the object by composition; a planmight be searched at different levels of abstractions throughabstraction and specification of objects/states and actions.Overall, these abilities are at the core of the flexible solu-tion of new problems. New problems are challenging exactlybecause their solution requires knowledge that the agent ini-tially does not have, and that has to be built through CKTprocesses. For example, in the classic problem of Duncker(Duncker, 1945), formulated within the psychological litera-ture on problem solving, a participant has to solve a problemof fixing a candle on a wall, and light it with some matches,using some pins in a box. The solution is to pin the cardboardbox of the pins to the wall, and then to put the candle onit. Finding this solution requires to use abstraction to realisethat the box is made of cardboard, and so can be pinned ontothe wall, and to imagine by composition the never-seen com-pound formed by the box pinned to the wall and the candleon it.

Link to the neuroscientific and psychological theories

The components of the RIM theory of consciousness,which are grounded on the fundamental elements of the braingoal-directed system, can be linked to the main aspects ofthe theories of consciousness proposed in the literature andconsidered in the previous sections. This section draws somelinks between those elements to these aspects as examples,but many other links can be easily identified.

The global workspace theory (GWT; Baars, 1997) and theneuronal workspace model (NWM; Dehaene & Changeux,

Figure 2. The four fundamental classes of operations thatcan be performed by the representation internal-manipulatorat the core of consciousness.

2011) suggest that the active maintenance of internal repre-sentations is a fundamental element of consciousness. Thisfunctionality can be related to both the central working mem-ory component of the RIM theory, and the perceptual work-ing memories. The two theories also stress the importanceof the amplification of relevant information by higher ordercortices. In particular, the NWM remarks the importanceof the frontal-parietal system for the top-down selection ofinformation. In the RIM theory, the amplification functionrelies on attention processes involving the top-down corti-cal pathways and also on the selection processes relying onthe basal ganglia-thalamo-cortical loops (Granato & Baldas-sarre, 2019). These selection processes play a crucial role inthe RIM theory as they are at the basis of the internal ma-nipulation of representations. Both theories stress the impor-tance of a consciousness bottleneck, for which only one/fewelements at a time enter the focus of attention. In the RIMtheory, representations can take place at multiple levels ofabstraction, so the manipulator has to select only one/fewlinked representations at a time to avoid interference betweenthem.

The convergence-divergence zones theory (CDZt; Dama-sio, 1989) proposes the existence of a brain multi-level hi-erarchy for information processing, highlighting the funda-mental difference between a abstract and perceptual process-ing and maintenance of information. The RIM theory of con-sciousness also gives a relevant importance to the hierarchi-cal processing and maintenance of information, in particularat the level of abstract working memory and perceptual work-ing memories, to support the key process of imagination.

The GWT and the NWM, and even more the somatic-marker hypothesis linked to the CDZt, assign a notable im-portance to the motivational/emotional processes for guid-ing conscious processes. In the RIM theory, the motivationsystem plays a fundamental role to guide the selections per-formed by the top-down internal manipulator onto the ab-


stract and perceptual working memories. In this respect, themotivation system is at the apex of the hierarchy that ul-timately establishes which information should be attendedand which manipulations of internal representations shouldbe performed.

The integrated information theory (IIT; Tononi, 2008)stresses the features of neural-network and cognitive archi-tectures that can support conscious processes. We hypothe-sise that the circuits posited by the RIM theory might havea high Φ value (Tononi, 2008). In particular, we posit thatin brain the neural assemblies representing the elements rel-evant for goal-directed processes (e.g., in cortex), and theloops supporting their internal selection and manipulation(e.g., based on the cortical circuits and the basal ganglia-thalamo-cortical loops), should present a high-level of speci-ficity and integration.

The sensorimotor theory of consciousness (O’Regan &Noe, 2001) emphasises the relevance for consciousness ofthe sensorimotor interaction that conscious agents engagewith the environment, in particular for the role played by thecausal action-outcome correlations that conscious agents cansense. The closest link that the RIM theory has with this viewis that the theory is a functional theory that posits that con-sciousness is a process having an adaptive utility for agentsas it allows them to produce more adaptive actions. In par-ticular, conscious processes allow the flexible manipulationof internal representations and these support a more effectiveplanning and a more powerful problem solving. The successof these processes requires a close alignment between theagents’ internally manipulated representations and the worldwhere they act, and a tight sensation-action coupling with itfor action success, as claimed by the sensorimotor theory ofconsciousness.

Link to computational and machine-learning models: aresearch agenda

In this section we discuss some initial ideas on how thecomponents of the RIM theory could be implemented incomputational models (Figure 3), while also drafting a pos-sible research agenda to realise them.

A possibility to implement a computational model follow-ing the principles of the RIM theory would be to realise amodel capable of performing planning processes as in clas-sic AI systems, but based on neural networks (Baldassarre,2001). The use of neural networks, vs. symbolic approaches,to implement the system would be a condition necessary toallow the quantitative ‘morphing and warping’ of the internalrepresentations of the system.

Initially, the system could work on the basis of wholestate/observations representations (Russell & Norvig, 2016)as commonly done in RL systems (R. S. Sutton & Barto,1998). This would allow the system to perform simple formsof planning as in Dyna planning (R. Sutton, 1990), i.e. by

using RL run on the basis of an internal model of the worldrather than in the actual real world (Baldassarre, 2003). Thiswould require the system to learn a model of the environ-ment on the basis of experience in the world, either before orduring the accomplishment of desirable goals. Moreover, thesystem might be able to initially perform only one-step plans,involving the processing of the current state, of the goal, andof actions: this system might already allow the study of someelements of the internal manipulation of representations thatis central to the theory (we discuss below a model inspiredby the RIM theory exemplifying this).

It would be important to implement this initial systemthrough the elements of the RIM theory. In particular, thestates of the world should be processed through a generativemodel implementing a perceptual working memory, e.g. forprocessing visual information (we discuss below a model in-spired by the RIM theory exemplifying this). The use of agenerative model would allow the system to abstract infor-mation along the bottom-up information flow and to formneural representations capturing hidden causes of the per-cepts, possibly at different levels of abstraction. The top-down manipulation of these representations would allow themodel to generate (imagine) different percepts. The goal-driven active imagination of different possible input stimulicould allow the creative solution of problems (as in the ex-ample model discussed below). Different generative mod-els could be used to implement this component, such as Re-stricted Boltzmann Machines (RBM; Hinton, 2002), varia-tional autoencoders (VAE; Kingma & Welling, 2013), orgenerative adversarial networks (GAN; Goodfellow et al.,2014).

An abstract working memory component could store per-sistent representations of the pursued goal, sub-goals, andother elements relevant for planning (and problem solving,see below). This component could be implemented througha recurrent neural network such as a long-short term memory(LSTM; Hochreiter & Schmidhuber, 1997) or transformers(Vaswani et al., 2017). The use of the latter model could sup-port an easier implementation of the internal attention selec-tive processes needed to manipulate internal representations,via the manipulator system, thanks to the internal ‘attentionheads’ of such model.

An internal manipulator component would allow the se-lection of elements needed for planning such as suitable rep-resentations of the initial state, goals, and actions to perform.This component could learn to select the internal representa-tions of the system on the basis of reinforcement learning(RL) processes. Two strategies might be followed to thispurpose. A first strategy, now successfully used in severalworks (e.g., Levine, Finn, Darrell, & Abbeel, 2016; Mnih etal., 2015), would use an ‘end-to-end’ RL strategy where thereward is used to train by RL the outer action-oriented partof the model, and then to use a supervised gradient-based


Representation Internal-Manipulation (RIM) Implementation Framework

Figure 3. Possible key elements of a RIM theory usable as a framework to build specific computational models. The graphshows the possible functional components of the theory, and some of the possible machine learning systems that might be usedto implement them.

algorithm to train all layers of the system, here including theinternal manipulation processes, down to the layers close tothe input. A second strategy, that we are exploring at themoment, is inspired by the functioning of brain where re-ward signals (e.g., based on dopamine or noradrenaline neu-romodulators) directly reach, and affect the learning of, the‘deep layers of the system. This strategy specifically impliesto use reward signals to learn to select by RL not only therepresentations of the actions to perform in the environment,but also the internal representations of the system that affectthe selection of such actions (we are at the moment develop-ing some models implementing this idea).

A motivation system would furnish the reward signals toguide the learning of the manipulation processes of the sys-tem. These rewards might be generated by different possi-ble mechanisms. They might just be standard rewards, asthose commonly used in RL (R. S. Sutton & Barto, 1998), orthey might be represented by ‘pseudo-rewards’ indicating ifa desired goal has been achieved. Alternatively, they mightbe represented by ‘intrinsic rewards’ (Baldassarre & Mirolli,2013; Barto, Singh, & Chentanez, 2004), or goal-related ‘in-trinsic pseudo-rewards’, generated by intrinsic motivationmechanisms (Santucci, Baldassarre, & Mirolli, 2016).

A further step of this research agenda might aim to im-plement a multi-step planning process allowing the systemto accomplish goals that require a sequence of actions, ratherthan one action, to be accomplished (Wayne et al., 2018).

Another fundamental step would be to enhance the sys-tem to allow it to do the ‘morphing and warping’ of inter-nal representations advocated by the RIM theory to activelybuild lacking knowledge elements that are necessary to solvechallenging problems that go beyond the need for a ‘simple’planning. Crucially, the system should be able to ‘search’solutions to problems not only by assembling sequences ofactions, as in planning, but also by changing the internalrepresentations of the key elements of planning on the ba-sis of the four internal-manipulation operations of decompo-sition, composition, abstract, and specification. Generativemodels are relevant for this purpose as they show to have anotable potential to generate novel imagined representationsby manipulating the internal representations capturing rele-vant hidden causes of learned stimuli (Bau et al., 2019; Klys,Snell, & Zemel, 2018).

An important means of validation of the principles of theRIM theory would be to show their capacity to solve prob-lems that require not only to accomplish goals by planning(and possibly also to autonomously acquire the knowledgeon the world to support such planning) but also to solve prob-lems that require elements that have not been previously ex-perienced/learned by the agent. The system might for ex-ample be required to modify internal representations of ex-perienced conditions, objects, and actions to form new onesallowing the solution of new problems.

The ideas above led us to implement two computational


models aiming to implement some of the principles of theRIM theory. A first model (Baldassarre, Lord, Granato, &Santucci, 2019; see Figure 4) uses decomposition (based onovert and covert attention processes) to segment an overallgoal into a number of sub-goals that the agent is able tosolve. In particular, a goal assigned to the system requiredthe creation of a certain configuration of objects in space(based on a simple custom simulator of simple 2D colouredcircle/triangle/rectangle ‘objects’). The agent could accom-plish the overall goal by decomposing it into sub-problemsthanks to the independence between the sub-goals: each sub-goal could be accomplished by re-arranging the single objectin the position and state indicated by the overall goal. Themodel could then accomplish each sub-goal through one-step planning operations involving the performance of single(hardwired) actions.

A second ‘companion’ model (Granato & Baldassarre,2019) implements some initial mechanisms to perform ab-straction and specification. The model reproduces the be-haviour, and errors, of human participants of the Wisconsincard sorting test (WCST; Berg, n.d.; Heaton, Chelune, Tal-ley, Kay, & Curtiss, 1993), a standard neuropsychologicaltest used to study cognitive flexibility. The model is ableto actively apply abstraction on objects (cards to sort), thusextracting either their shape, or colour, or size, and then tosearch correspondences between objects, with respect to oneof such dimensions, to solve the WCST. The model is in par-ticular able to employ specification to imagine objects thatare formed by specific attributes, for example to imagine asmall, red, circle. These processes are based on a restrictedBoltzmann machine that perceives cards and, with its ‘hid-den units’, learns (for now based on a supervised algorithm)to encode the different shape/colour/size attributes of the ob-jects depicted on them. The different possible attributes ofthe cards are then used to compare them along different di-mensions and to solve the WCST by inferring the unknownrule to match the cards through a RL algorithm.

Manipulation agency and the subjective experience ofconsciousness

The nature of subjective experience is a widely debatedissue within the philosophical literature. Many philosopherspropose different ideas to explain the uniqueness of sub-jective experience such as the ‘zombie argumentation’ ofChalmers (Chalmers, 2003), for which any physicalist ex-planation of subjective experience is deemed to fail, thered room mental experiment of Searle (Searle, 2004), warn-ing against the possibility of processing information withouthaving a deep grasp of the process, or the list of fundamen-tal features of qualia, i.e. private subjective sensations, ofDennet (Dennett, 1988).

Some scientific theories of consciousness point to whatshould be the fundamental elements of subjective experience,

for example through the concept of ‘qualia space’, a compu-tational model capturing the functioning of brain thalamo-cortical networks at the basis of consciousness (Tononi,2008); the drawing of a close link between visual experienceand embodied sensorimotor loops (O’Regan & Noe, 2001);the emotional dimension of conscious experience (Damasio,1996); or the dynamics of a complex distributed frontal-parietal brain network proposed within the GWT (Baars etal., 2013; Seth, Baars, & Edelman, 2005) and NWM (De-haene & Naccache, 2001). While we think that each of theseproposals highlights a fundamental aspect of the subjectiveelements of conscious experience, it seems that they still re-main unclear and that the RIM theory of consciousness pro-posed here might contribute to better understand them.

Based on the RIM theory, a conscious agent intentionallycontrols the representation internal-manipulator to create aninternal simulated reality at will. This manipulation has anotable power as it relies on the four flexible primitive op-erations of manipulations illustrated above that can trans-form and warp internal representations in arbitrary ways.This power allows the agent to also internally simulate it-self, e.g. based on images involving other similar agents:these might be possibly abstracted along various dimensionsand then added, by composition, some features related tooneself, including elements of own cognition (Fernandez-Duque, Baird, & Posner, 2000).

An key point of our view is that, based on mechanisms asthose operating for the actions exerted on the external world,conscious agents have a sense of agency for the actions ex-erted through the internal manipulator on the internal realitygenerated by imagination. We call this manipulation agency.

The internally simulated and manipulated realty relies onimagination that involves also the lowest-level perceptualrepresentations. The representations that the internal manip-ulator acts on are largely overlapped to those activated by theperception of the real world: this activation thus produces avivid perceptual and emotional experiences similar to thoseproduced by experiences in the real world. These processesare at the basis of the human intuition that we have an exclu-sive private access to our conscious thoughts, a fundamentalfeature of the concept of Qualia (Dennett, 1988, 2001).

To summarise, the RIM theory posits that the subjectiveexperience of consciousness relies on the operations of thegoal-driven representation internal-manipulator that allowthe agent to intentionally create, and powerfully manipu-late as desired, an internal reality that possibly includes el-ements of the agent itself and has the perceptual and emo-tional vividness of the outer world.

The concept of manipulation agency could also be usedto propose a possible ‘consciousness scale’. Increasing lev-els of manipulation agency could characterise three statesof: phenomenal consciousness, access consciousness, andmanipulative consciousness (Figure 5). The three levels


Figure 4. Key elements of previous computational models that implemented some elements of the RIM theory. (Left) Themodel that can perform a decomposition of some types of goals into sub-goals based on decomposition (Baldassarre et al.,2019). (Right) The model that performs selected types of abstraction and specification of input images (Granato & Baldassarre,2019).

of manipulation agency could characterise different neuro-cognitive states accompanying three typical mental process-ing modes. In particular, an unexpected event could cause thetransitory perception accompanied by a low level of controland manipulation agency. Mind-wandering, a common brainmode leading to explore sequences of conscious thoughts,and accompanying the performance of routine automatic ac-tivities or causing a mental shift from a specific tasks, oftenhappening without awareness (Schooler et al., 2011), mightbe characterised by a medium level of manipulation agency.At last, a high attention focus characterising states of mindas those achieved with mindfulness could be accompaniedby the highest level of control on own cognition and henceof manipulation agency.

Our proposal could also clarify the difference betweenpseudo-hallucinations (which do not mimic real perceptionand are perceived as unreal), and hallucinations (percep-tion in the absence of external stimuli, having the qualityof real perception, Telles-Correia, Moreira, & Goncalves,2015). In particular, different alternations of manipulationagency could explain the presence of insight in pseudohallu-cinations, and its lack in true hallucinations. These states arealso reported in (Figure 5) to indicate the levels of manipula-tion agency possibly corresponding to them. The scale couldalso characterise dreams, involving uncontrolled imaginationduring REM sleep, and lucid dreams, involving controlledimagination during sleep (Stumbrys, Erlacher, Schädlich, &Schredl, 2012).

Figure 5. A scale of consciousness based on the concept ofmanipulation agency.

Conclusions

We introduced the representation internal-manipulation(RIM) theory of consciousness. The theory views consciousprocesses as playing a key adaptive function for flexible plan-ning and problem solving. The theory pivots on the idea thatthe processes of consciousness support the internal manipu-lation of representations of all aspects of goal-directed pro-cesses, from the representations of objects, to those of goals,actions, and plans. This manipulation relies on four cardi-nal operations (decomposition, composition, abstraction, and


specification) that allow conscious representation manipula-tion to build new knowledge on the basis of previously ac-quired experience, a process called here conscious knowl-edge transfer.

We discussed how the RIM theory integrates various as-pects of some main theories of consciousness, in particularby linking them to different elements of goal-directed pro-cesses at the basis of the theory. We also discussed prelimi-nary ideas on the possible computational implementation ofsome elements of the theory based on current machine learn-ing systems, and we briefly reviewed initial computationalmodels implemented to capture some of the theory ideas.

Finally, we used the theory to propose a new interpretationof the subjective elements of conscious experience by relyingon the new concept of manipulation agency. This indicatesthe sense that humans have of being able to manipulate asdesired, in powerful ways, an internal imagined reality thatpossibly includes oneself and that is vividly sensed and feltas the outer world.

References

Anderson, M. L. (2003). Embodied cognition: A field guide. Arti-ficial intelligence, 149(1), 91–130.

Aust, U., Range, F., Steurer, M., & Huber, L. (2008). Inferentialreasoning by exclusion in pigeons, dogs, and humans. Ani-mal cognition, 11(4), 587–597.

Baars, B. J. (1997). In the theatre of consciousness. globalworkspace theory, a rigorous scientific theory of conscious-ness. Journal of Consciousness Studies, 4(4), 292–309.

Baars, B. J. (2005). Global workspace theory of conscious-ness: toward a cognitive neuroscience of human experience.Progress in brain research, 150, 45–53.

Baars, B. J., Franklin, S., & Ramsoy, T. Z. (2013). Globalworkspace dynamics: cortical "binding and propagation" en-ables conscious contents. Frontiers in psychology, 4.

Baars, B. J., Ramsøy, T. Z., & Laureys, S. (2003). Brain, consciousexperience and the observing self. Trends in neurosciences,26(12), 671–675.

Baddeley. (2000, Nov). The episodic buffer: a new component ofworking memory? Trends Cogn Sci, 4(11), 417–423.

Baldassarre, G. (2001). Planning with neural networks and rein-forcement learning (PhD Thesis). Computer Science Depart-ment, University of Essex, Colchester, UK.

Baldassarre, G. (2003). Forward and bidirectional planning basedon reinforcement learning and neural networks in a simu-lated robot. In M. Butz, O. Sigaud, & P. GÃ c©rard (Eds.),Anticipatory behaviour in adaptive learning systems (Vol.2684, pp. 179–200). Berlin: Springer Verlag.

Baldassarre, G. (2011). What are intrinsic motivations? a biologicalperspective. In A. Cangelosi et al. (Eds.), Proceedings of theinternational conference on development and learning andepigenetic robotics (icdl-epirob-2011) (pp. e1–8). New York,NY: IEEE. (Frankfurt am Main, Germany, 24–27/08/11)

Baldassarre, G., Caligiore, D., & Mannella, F. (2013). The hierar-chical organisation of cortical and basal-ganglia systems: acomputationally-informed review and integrated hypothesis.

In G. Baldassarre & M. Mirolli (Eds.), Computational androbotic models of the hierarchical organisation of behaviour(p. 237-270). Berlin: Springer-Verlag.

Baldassarre, G., Lord, W., Granato, G., & Santucci, V. G. (2019).An embodied agent learning affordances with intrinsic mo-tivations and solving extrinsic tasks with attention and one-step planning. Frontiers in neurorobotics, 13, 45.

Baldassarre, G., Mannella, F., Fiore, V. G., Redgrave, P., Gurney,K., & Mirolli, M. (2013). Intrinsically motivated action-outcome learning and goal-based action recall: A system-level bio-constrained computational model. Neural Net-works, 41, 168-187.

Baldassarre, G., & Mirolli, M. (Eds.). (2013). Intrinsically mo-tivated learning in natural and artificial systems. Berlin:Springer. (Cost 91.62 euros, pp. 458, 82 illustrations, 55illustrations in color)

Balleine, B. W., & Dickinson, A. (1998). Goal-directed instru-mental action: contingency and incentive learning and theircortical substrates. Neuropharmacology, 37(4-5), 407–419.

Barto, A. G., Singh, S., & Chentanez, N. (2004). Intrinsicallymotivated learning of hierarchical collections of skills. InJ. Triesch & T. Jebara (Eds.), International conference on de-velopmental learning (icdl2004) (pp. 112–119). New York,NY.

Bau, D., Strobelt, H., Peebles, W., Wulff, J., Zhou, B., Zhu, J.-Y.,& Torralba, A. (2019). Semantic photo manipulation witha generative image prior. ACM Transactions on Graphics(TOG), 38(4), 59.

Bechara, A., & Damasio, A. R. (2005). The somatic marker hy-pothesis: A neural theory of economic decision. Games andeconomic behavior, 52(2), 336–372.

Bengio, Y. (2017). The consciousness prior. arXiv preprintarXiv:1709.08568.

Benn, Y., Webb, T. L., Chang, B. P. I., Sun, Y.-H., Wilkinson, I. D.,& Farrow, T. F. D. (2014). The neural basis of monitoringgoal progress. Front Hum Neurosci, 8, 688.

Berg, E. A. (n.d.). A simple objective technique for measuringflexibility in thinking. Journal General Psychology(39), 15–22.

Berridge, K. C. (2004, Apr). Motivation concepts in behavioralneuroscience. Physiol Behav, 81(2), 179–209.

Blaisdell, A. P., Sawa, K., Leising, K. J., & Waldmann, M. R.(2006). Causal reasoning in rats. Science, 311(5763), 1020–1022.

Borghi, A. M., & Cimatti, F. (2010, Feb). Embodied cognitionand beyond: acting and sensing the body. Neuropsychologia,48(3), 763–773.

Caligiore, D., Arbib, M. A., Miall, C. R., & Baldassarre, G. (2019,February). The super-learning hypothesis: Integrating learn-ing processes across cortex, cerebellum and basal ganglia.Neuroscience and Biobehavioral Reviews, 100, 19–34.

Chalmers, D. J. (1995). Facing up to the problem of consciousness.Journal of consciousness studies, 2(3), 200–219.

Chalmers, D. J. (2003). Consciousness and its place in nature.Blackwell guide to the philosophy of mind, 102–142.

Chiappe, D., & MacDonald, K. (2005, Jan). The evolution ofdomain-general mechanisms in intelligence and learning. JGen Psychol, 132(1), 5–40.


Cisek, P., & Kalaska, J. F. (2010). Neural mechanisms for interact-ing with a world full of action choices. Annu Rev Neurosci,33, 269–298.

Clayton, N. S., & Dickinson, A. (2010). Mental time travel: Cananimals recall the past and plan for the future? Encyclopediaof Animal Behavior, 438–442.

Condell, J., Wade, J., Galway, L., McBride, M., Gormley, P., Bren-nan, J., & Somasundram, T. (2010). Problem solving tech-niques in cognitive science. Artificial Intelligence Review,34(3), 221–234.

Damasio, A. R. (1989). The brain binds entities and events by mul-tiregional activation from convergence zones. Neural Com-putation, 1(1), 123–132.

Damasio, A. R. (1996, Oct). The somatic marker hypothesis andthe possible functions of the prefrontal cortex. Philos TransR Soc Lond B Biol Sci, 351(1346), 1413–1420.

Deci, E. L., Vallerand, R. J., Pelletier, L. G., & Ryan, R. M. (1991).Motivation and education: The self-determination perspec-tive. Educational psychologist, 26(3-4), 325–346.

Dehaene, S., & Changeux, J.-P. (2011, Apr). Experimental and the-oretical approaches to conscious processing. Neuron, 70(2),200–227.

Dehaene, S., Kerszberg, M., & Changeux, J. P. (1998, November).A neuronal model of a global workspace in effortful cogni-tive tasks. Proceedings of the National Academy of Sciencesof the United States of America, 95, 14529–14534.

Dehaene, S., & Naccache, L. (2001). Towards a cognitive neuro-science of consciousness: basic evidence and a workspaceframework. Cognition, 79(1), 1–37.

Delatour, B., & Gisquet-Verrier, P. (2000). Functional role of ratprelimbic-infralimbic cortices in spatial memory: evidencefor their involvement in attention and behavioural flexibility.Behavioural brain research, 109(1), 113–128.

Dennett, D. C. (1988). Quining qualia. In Consciousness in modernscience. Oxford University Press.

Dennett, D. C. (2001). Are we explaining consciousness yet? Cog-nition, 79(1), 221–237.

Desrochers, T. M., Burk, D. C., Badre, D., & Sheinberg, D. L.(2015). The monitoring and control of task sequences in hu-man and non-human primates. Front Syst Neurosci, 9, 185.

Diamond, A. (2013). Executive functions. Annu Rev Psychol, 64,135–168.

Duncker, K. (1945). On problem solving (No. 58). American Psy-chological Association.

Durston, S., Thomas, K. M., Yang, Y., Ulug, A. M., Zimmerman,R. D., & Casey, B. (2002). A neural basis for the devel-opment of inhibitory control. Developmental Science, 5(4),F9–F16.

Fantz, R. L. (1964, Oct). Visual experience in infants: Decreasedattention to familiar patterns relative to novel ones. Science,146(3644), 668–670.

Felleman, D. J., & Van Essen, D. C. (1991). Distributed hierarchi-cal processing in the primate cerebral cortex. Cereb Cortex,1(1), 1–47.

Fernandez-Duque, D., Baird, J. A., & Posner, M. I. (2000). Exec-utive attention and metacognitive regulation. Consciousnessand cognition, 9(2), 288–307.

Fiehler, K., Burke, M., Engel, A., Bien, S., & Rösler, F. (2008,

Feb). Kinesthetic working memory and action control withinthe dorsal stream. Cereb Cortex, 18(2), 243–253.

Garbarini, F., & Adenzato, M. (2004, October). At the root of em-bodied cognition: cognitive science meets neurophysiology.Brain and cognition, 56, 100–106.

Gentner, D., & Forbus, K. D. (2011). Computational models ofanalogy. Wiley Interdisciplinary Reviews: Cognitive Science,2(3), 266–276.

Gerlach, K. D., Spreng, R. N., Gilmore, A. W., & Schacter, D. L.(2011, Apr). Solving future problems: default network andexecutive activity associated with goal-directed mental sim-ulations. Neuroimage, 55(4), 1816–1824.

Goldman-Rakic, P. S. (1996, Oct). The prefrontal landscape: impli-cations of functional architecture for understanding humanmentation and the central executive. Philos Trans R Soc LondB Biol Sci, 351(1346), 1445–1453.

Goodale, M. A., & Milner, A. D. (1992, January). Separate visualpathways for perception and action. Trends Neurosci, 15(1),20–25.

Goodfellow, I., Bengio, Y., & Courville, A. (2017). Deep learn-ing. Boston, MA: The MIT Press. Retrieved from http://www.iro.umontreal.ca/~bengioy/dlbook

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley,D., Ozair, S., . . . Bengio, Y. (2014). Generative adversarialnets. In Advances in neural information processing systems(pp. 2672–2680). (10 Jun 2014)

Granato, G., & Baldassarre, G. (2019). Goal-directed top-down control of perceptual representations: A computationalmodel of the wisconsin card sorting test. In Conference oncognitive computational neuroscience. (13-16 September2019, Berlin Germany)

Haggard, P. (2017). Sense of agency in the human brain. NatureReviews Neuroscience, 18(4), 196.

Hartley, A. A., & Speer, N. K. (2000). Locating and fractionatingworking memory using functional neuroimaging: storage,maintenance, and executive functions. Microscopy researchand technique, 51(1), 45–53.

Heaton, R. K., Chelune, G. J., Talley, J. L., Kay, G. G., & Curtiss,G. (1993). Wcst: Wisconsin card sorting test. PsychologicalAssessment resources.

Hills, T. T. (2006, Jan). Animal foraging and the evolution of goal-directed cognition. Cogn Sci, 30(1), 3–41.

Hinton, G. E. (2002). Training products of experts by minimizingcontrastive divergence. Neural computation, 14(8), 1771–1800.

Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory.Neural computation, 9, 1735–1780.

Hutto, D. D. (2005). Knowing what? radical versus conserva-tive enactivism. Phenomenology and the Cognitive Sciences,4(4), 389–405.

Keller, A. (2014). The evolutionary function of conscious infor-mation processing is revealed by its task-dependency in theolfactory system. Front Psychol, 5, 62.

Kingma, D. P., & Welling, M. (2013). Auto-encoding variationalbayes. arXiv.

Klys, J., Snell, J., & Zemel, R. (2018). Learning latent subspaces invariational autoencoders. In Advances in neural informationprocessing systems (pp. 6444–6454).

http://www.iro.umontreal.ca/~bengioy/dlbook

http://www.iro.umontreal.ca/~bengioy/dlbook


Koch, C., Massimini, M., Boly, M., & Tononi, G. (2016). Neuralcorrelates of consciousness: progress and problems. NatureReviews Neuroscience, 17(5), 307.

Kumaran, D., & Maguire, E. A. (2007, Aug). Match mismatch pro-cesses underlie human hippocampal responses to associativenovelty. J Neurosci, 27(32), 8517–8524.

Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J.(2017). Building machines that learn and think like people.Brain and Behavioural Sciences, 40, 1–72.

Lavie, N., Hirst, A., de Fockert, J. W., & Viding, E. (2004, Sep).Load theory of selective attention and cognitive control. JExp Psychol Gen, 133(3), 339–354.

Leber, A. B., Turk-Browne, N. B., & Chun, M. M. (2008, Sep).Neural predictors of moment-to-moment fluctuations in cog-nitive flexibility. Proc Natl Acad Sci U S A, 105(36), 13592–13597.

Levine, S., Finn, C., Darrell, T., & Abbeel, P. (2016). End-to-endtraining of deep visuomotor policies. Journal of MachineLearning Research, 17(39), 1–40.

Lisman, J. E., & Grace, A. A. (2005). The hippocampal-vta loop:controlling the entry of information into long-term memory.Neuron, 46(5), 703–713.

Mannella, F., Gurney, K., & Baldassarre, G. (2013). The nucleusaccumbens as a nexus between values and goals in goal-directed behavior: a review and a new hypothesis. Frontiersin Behavioral Neuroscience, 7(135), e1-29.

McClelland, D. E., James L. andRumelhart, & the PDPResearch-Group. (1986). Parallel distributed processing: explorationsin the microstructure of cognition (Vol. 1). Cambridge,MA:The MIT Press.

Mechelli, A., Price, C. J., Friston, K. J., & Ishai, A. (2004, Nov).Where bottom-up meets top-down: neuronal interactionsduring perception and imagery. Cereb Cortex, 14(11), 1256–1265.

Meyer, K., & Damasio, A. (2009, Jul). Convergence and divergencein a neural architecture for recognition and memory. TrendsNeurosci, 32(7), 376–382.

Mirolli, M., Mannella, F., & Baldassarre, G. (2010). The roles ofthe amygdala in the affective regulation of body, brain, andbehaviour. Connection Science, 22(3), 215–245.

Miyake, A., Friedman, N. P., Emerson, M. J., Witzki, A. H., How-erter, A., & Wager, T. D. (2000, Aug). The unity and diver-sity of executive functions and their contributions to complex"frontal lobe" tasks: a latent variable analysis. Cogn Psychol,41(1), 49–100.

Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness,J., Bellemare, M. G., . . . Hassabis, D. (2015, February).Human-level control through deep reinforcement learning.Nature, 518(7540), 529–533.

Nair, A., Pong, V., Dalal, M., Bahl, S., Lin, S., & Levine, S.(2018). Visual reinforcement learning with imagined goals.In The second lifelong learning: A reinforcement learningapproach workshop (llrla2018 at faim2018). (Stockholm,Sweden)

Newman, S. D., Carpenter, P. A., Varma, S., & Just, M. A. (2003).Frontal and parietal participation in problem solving in thetower of london: fmri and computational modeling of plan-ning and high-level perception. Neuropsychologia, 41(12),

1668–1682.Ognibene, D., & Baldassarre, G. (2015). Ecological active vi-

sion: four bio-inspired principles to integrate bottom-up andadaptive top-down attention tested with a simple camera-armrobot. IEEE Transactions on Autonomous Mental Develop-ment, 7(1), 3–25.

O’Regan, J. K., & Noe, A. (2001, Oct). A sensorimotor accountof vision and visual consciousness. Behav Brain Sci, 24(5),939–73; discussion 973-1031.

O’Reilly, J. X., Schüffelgen, U., Cuell, S. F., Behrens, T. E., Mars,R. B., & Rushworth, M. F. (2013). Dissociable effects ofsurprise and model update in parietal and anterior cingulatecortex. Proceedings of the National Academy of Sciences,201305373.

O’Regan, J. K., Myin, E., & Noë, A. (2005). Sensory conscious-ness explained (better) in terms of corporality and alertingcapacity. Phenomenology and the Cognitive Sciences, 4(4),369–387.

Panksepp, J. (1998). Affective neuroscience: the foundations of hu-man and animal emotions. Oxford: Oxford Unversity Press.

Passingham, R. E., & Wise, S. P. (2012). The neurobiology ofthe prefrontal cortex: anatomy, evolution, and the origin ofinsight (Vol. 50). Oxford: Oxford University Press.

Paus, T. (2001, Jun). Primate anterior cingulate cortex: where mo-tor control, drive and cognition interface. Nat Rev Neurosci,2(6), 417–424.

Pennartz, C. M. (2017). Consciousness, representation, action:the importance of being goal-directed. Trends in cognitivesciences.

Pezzulo, G., Rigoli, F., & Chersi, F. (2013). The mixed instrumentalcontroller: using value of information to combine habitualchoice and mental simulation. Front Psychol, 4, 92.

Pfeiffer, B. E., & Foster, D. J. (2013). Hippocampal place-cellsequences depict future paths to remembered goals. Nature,497(7447), 74.

Posner, M. I., & Petersen, S. E. (1990). The attention system of thehuman brain. Annu Rev Neurosci, 13, 25–42.

Raffone, A., Srinivasan, N., & van Leeuwen, C. (2014, May). Per-ceptual awareness and its neural basis: bridging experimen-tal and theoretical paradigms. Philos Trans R Soc Lond BBiol Sci, 369(1641), 20130203.

Raffone, A., Srinivasan, N., & van Leeuwen, C. (2015, Feb). Rapidswitching and complementary evidence accumulation enableflexibility of an all-or-none global workspace for control ofattentional and conscious processing: a reply to wyble et al.Philos Trans R Soc Lond B Biol Sci, 370(1661), 20140315.

Redgrave, P., Prescott, T. J., & Gurney, K. (1999). The basal gan-glia: a vertebrate solution to the selection problem? Neuro-science, 89(4), 1009–1023.

Ressler, N. (2004). Rewards and punishments, goal-directed be-havior and consciousness. Neuroscience & BiobehavioralReviews, 28(1), 27–39.

Rizzolatti, G., Luppino, G., & Matelli, M. (1998, April). The orga-nization of the cortical motor system: new concepts. Elec-troencephalography and clinical neurophysiology, 106(4),283–96.

Rizzolatti, G., Riggio, L., Dascola, I., & Umiltá, C. (1987). Reori-enting attention across the horizontal and vertical meridians:


evidence in favor of a premotor theory of attention. Neu-ropsychologia, 25(1), 31–40.

Robinson, H., Calamia, M., Gläscher, J., Bruss, J., & Tranel, D.(2014, Jan). Neuroanatomical correlates of executive func-tions: a neuropsychological approach using the examinerbattery. J Int Neuropsychol Soc, 20(1), 52–63.

Rossano, M. J. (2003, Oct). Expertise and the evolution of con-sciousness. Cognition, 89(3), 207–236.

Rounis, E., & Humphreys, G. (2015). Limb apraxia and the "af-fordance competition hypothesis". Front Hum Neurosci, 9,429.

Roy, M., Shohamy, D., & Wager, T. D. (2012, Mar). Ventromedialprefrontal-subcortical systems and the generation of affec-tive meaning. Trends Cogn Sci, 16(3), 147–156.

Russell, S. J., & Norvig, P. (2016). Artificial intelligence: A modernapproach (Third Edition ed.). Harlow, UK: Pearson Educa-tion.

Santucci, V. G., Baldassarre, G., & Mirolli, M. (2016). Grail:A goal-discovering robotic architecture for intrinsically-motivated learning. IEEE Transactions on Cognitive andDevelopmental Systems, 8(3), 214–231.

Schooler, J. W., Smallwood, J., Christoff, K., Handy, T. C., Reichle,E. D., & Sayette, M. A. (2011). Meta-awareness, perceptualdecoupling and the wandering mind. Trends in cognitive sci-ences, 15(7), 319–326.

Schultz, W. (2002, October). Getting formal with dopamine andreward. Neuron, 36(2), 241–263.

Searle, J. R. (2004). Mind: a brief introduction. Oxford UniversityPress.

Seepanomwan, K., Caligiore, D., Cangelosi, A., & Baldassarre, G.(2015). Generalization, decision making, and embodimenteffects in mental rotation: a neurorobotic architecture testedwith a humanoid robot. Neural Networks, 72, 31–47.

Seth, A. K., Baars, B. J., & Edelman, D. B. (2005). Criteria forconsciousness in humans and other mammals. Conscious-ness and cognition, 14(1), 119–139.

Simon, H. A. (1975). The functional equivalence of problem solv-ing skills. Cognitive psychology, 7(2), 268–288.

Stumbrys, T., Erlacher, D., Schädlich, M., & Schredl, M. (2012).Induction of lucid dreams: A systematic review of evidence.Consciousness and Cognition, 21(3), 1456–1475.

Sutton, R. (1990). Integrated architectures for learning, planning,and reacting based on approximating dynamic programming.

In Proceedings of the seventh international conference onmachine learning (Vol. 216, pp. 216–224).

Sutton, R. S., & Barto, A. G. (1998). Introduction to reinforcementlearning (Vol. 135). MIT press Cambridge.

Telles-Correia, D., Moreira, A. L., & Goncalves, J. S. (2015).Hallucinations and related concepts-their conceptual back-ground. Frontiers in psychology, 6, 991.

Thill, S., Caligiore, D., Borghi, A. M., Ziemke, T., & Baldassarre,G. (2013). Theories and computational models of affordanceand mirror systems: An integrative review. Neuroscienceand Biobehavioral Reviews, 37, 491-521.

Tononi, G. (2008, Dec). Consciousness as integrated information:a provisional manifesto. Biol Bull, 215(3), 216–242.

Tononi, G., Boly, M., Massimini, M., & Koch, C. (2016, Jul). In-tegrated information theory: from consciousness to its phys-ical substrate. Nat Rev Neurosci, 17(7), 450–461.

Treisman, A. M., & Gelade, G. (1980). A feature-integration theoryof attention. Cognitive psychology, 12(1), 97–136.

Tsujimoto, S., Genovesio, A., & Wise, S. P. (2011). Frontal polecortex: encoding ends at the end of the endbrain. Trends incognitive sciences, 15(4), 169–176.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L.,Gomez, A. N., . . . Polosukhin, I. (2017). Attention is allyou need. arXiv preprint arXiv:1706.03762.

Verdejo-García, A., Pérez-García, M., & Bechara, A. (2006, Jan).Emotion, decision-making and substance dependence: asomatic-marker model of addiction. Curr Neuropharmacol,4(1), 17–31.

Waltz, J. A., Knowlton, B. J., Holyoak, K. J., Boone, K. B.,Mishkin, F. S., de Menezes Santos, M., . . . Miller, B. L.(1999). A system for relational reasoning in human pre-frontal cortex. Psychological science, 10(2), 119–125.

Wayne, G., Hung, C.-C., Amos, D., Mirza, M., Ahuja, A., Grabska-Barwinska, A., . . . Lillicrap, T. (2018). Unsupervised pre-dictive memory in a goal-directed agent. arXiv.

White, R. W. (1959, Sep). Motivation reconsidered: the concept ofcompetence. Psychol Rev, 66, 297–333.

Wolters, G., & Raffone, A. (2008, Mar). Coherence and recurrency:maintenance, control and integration in working memory.Cogn Process, 9(1), 1–17.

Yin, H. H., & Knowlton, B. J. (2006, June). The role of the basalganglia in habit formation. Nature Reviews Neuroscience,7(6), 464–476.

representation internal-manipulation (rim): a neuro ... · putational concepts, have been...

Documents