Higher-level cognition

1. Computational Intelligence Higher-level cognition Based on a course taught by Prof. Randall O'Reilly University of Colorado and Prof. Włodzisława Ducha Uniwersytet Mikołaja Kopernika Janusz A. Starzyk

2. So far we have network models: Recognizing objects (sight, hearing, touch). Attention processes, cooperating with recognition. Remembering episodes, general semantic relations, retaining current information. Learning pronunciation and meaning of words. Higher-level functions What's missing? • Goals, striving, planning. • Reasoning. • Emotion. • Making decisions. • Representation of the ego, model of "I", self-knowledge. • Free will? • Social relations, hierarchy of values.

3. What is consciousness? • Consciousness is a process resulting from having sensations and having the belief that you have sensations. Necessary conditions for the emergence of conscious sensations: • model of environmental states vital to the organism; • sufficiently complex relational structure of internal states; • ability to undertake actions on the basis of an analysis of internal states; a specific instance is a talent for "stream of consciousness" narration. • Verification:Let's plan a brain-like system; let's show, that because of its very construction it must claim that it's aware of processes occuring inside it;let's try to challenge its beliefs.

4. CC PFC Biological foundations Damage to the prefrontal cortex (PFC) and the cingulate cortex (CC) influences planning, reasoning, decision making, social interactions, self-knowledge, intentional actions. The PFC and CC don't function in isolation from the rest! Are we machines? Free will, subjectivity and the brain (presentation 3/2007).

5. Working memory, dynamic activations not subject to dispersal. PFC functions Experiments with delayed image matching. PFC holds on for a certain time to information which activates "working memory," this enables a return to executed tasks despite the necessity of performing other actions, monitoring errors, coordinating goal-oriented actions, plans, generating varied actions.

6. It's hard to stop well-trained recognition or action processes (dance, music), eg. in the Stroop effect: we have to name the color of the consequitive words, but we do it slower than for words unrelated to color. Stroop

7. Recognition of words or colors. Stroop model The results are completely consistent with experience. There are many variants of this experiment and possibilities of tracking changes duringmultiple repetitions.

8. Phonological level: nonexistant words don't activate deeper areas. Deep level: phonological and semantic errors (cat – cot, cat - dog), mistakes in sign recognition. Stroop: shapes Superficial observations: new words + semantic access + difficulties with reading exceptions + mistakes in recognition.

9. Stroop: shapes Superficial observations: new words + semantic access + difficulties with reading exceptions + mistakes in recognition.

10. Dynamic categorization Wisconsin card sort test WCST Sorting by color In the WCST, patients have to sort cards according to one of the chosen patterns – shape, color or number of figures A change of the sorting rules is hindered in patients with damage to the frontal cortex

11. Dynamic categorization Sorting by color Wisconsin card sort test Experiments of this type allow us to locate specific types of information processing in the brain

12. Dynamic categorization Sorting by color Wisconsin card sort test

13. Dynamic categorization Wisconsin card sort test: patients with damage to the prefrontal cortex retain the first sorting rule and are not able to change it.

14. Decision about what choice to make Synaptic and dynamic actions Superficial observations: new words + semantic access + difficulties with reading exceptions + mistakes in recognition.

15. Decision about what choice to make – simple test performed on monkeys Dynamic categorization The goal is to choose one of the images, which represents the same 2D figure After the test subject reaches 90% correct answers, the rule is changed and the subject must choose a) a new 2D figure (IDS) b) the same 1D figure (EDR) c) a new 1D figure (EDS) d) reverse the role (IDR) - with reversed roles, we select the figure that we didn't select earlier Superficial observations: new words + semantic access + difficulties with reading exceptions + mistakes in recognition.

16. Orbital PFC: represents features, lateral PFC: abstract dimensions. Or ... Orbital = affective inhibition Lateral = choice based on attention. Activation of the PFC enables a change in function through: Orbital: new features (IDR) Lateral: new dimensions (EDS) Dynamic categorization Duration: synaptic effects with a lack of dynamic activation in the PFC In subjects with orbital damage, it's more difficult to change the rule in figures with the same dimensions (IDR) and in subjects with lateral PFC damage, it's more difficult to change the rule if choosing new figures with different dimensions.

17. Attention: activation PFC: sets the direction, VTA functions as an adaptive critic, causes changes in the PFC and a reward. IDE/ED model Duration: synaptic effects with a lack of dynamic activation in the PFC

18. IDR/EDS model Duration: synaptic effects with a lack of dynamic activation in the PFC

19. Model and experience Duration: synaptic effects with a lack of dynamic activation in the PFC

20. PFC tasks • Dynamic working memory: PFC neurons retain their activation, allowing for simultaneous fulfillment of several functions. • Monitoring/evaluation of action results: monitoring errors associated with the dopaminergic reward mechanism. • It seems that exactly this information becomes conscious, because evaluations must be accessible to all areas of the brain in order to influence correction processes. • Competition between goals causes a stoppage of one action and an activation of a different action • Flexibility: continuation of the same answer after a change in the task • Fluency: Difficulty in generating a variety of responses • Executive control: Problems with goal-directed planning and coordination

21. Uniform approach based on activation • Central frontal mechanisms: • Dynamic working memory: PFC neurons retain their activation, allowing for simultaneous fulfillment of several functions. • Monitoring/evaluation of action results: monitoring errors associated with the dopaminergic reward mechanism. • Flexibility: continuation of the same answer after a change in the task • Fluency: Problem with new categories of results – requires support from above in order to overcome existing categories • Executive control: Guarding and actualizing plans/goals in time, avoiding dispersal of attention

22. Automatisms and control Dynamic working memory: PFC neurons retain their activation, allowing for simultaneous fulfillment of several functions.

23. Automation of actions Formation of new quasistable brain states during learning => neural models. Learning requires reinforcement of desired actions, observation and evaluation of complex brain states. Connecting current actions with remembered results of similar actions requires evaluation and comparison, and then emotional reactions, which will liberate neurotransmitters (dopamine) as a reinforcement signal, increasing the speed of learning of neural modules Working memory is essential in a process this complex. Errors should be remembered, especially the bitter taste of defeat. There is no transfer from conscious to unconscious! There is only a (conscious) process of evaluation necessary for reinforcement. Learning: first, conscious actions engage the whole brain; in the end, automatic, subconscious, localized actions.

24. Where is the mind? Central Paradox of Cognitive Science: How, from the counting of impulses by neurons, do structure, symbols, meaning, sense, impressions, and emotions arise… i.e. the world of the mind? • Philosophical problems: psychophysical problem, problem of the quality of sensation, consciousness, semantics and syntactics, many thought experiments ... • Technical problems: • How to reconcile the integrity of the mind with distributed processing (binding problem)? • What are the conditions for the formation of sensations? • Psycho-logos, logic of the psyche, has very few general laws. • Lack of a good model connecting the neuro level and the psyche

25. Neurophenomenology F. Varela, hypothesis of neurophenomenology: phenomenological description of the structure of experience and its counterpart in experimental sciences mutually specify themselves. Extensive synchronizations of active brain areas are a neural basis for the experience of time. A sequence of synchronizations and relaxations is the proper functional level for consciously experienced time. Temporary events are represented in the form of a trajectory in the phase space(neural states). Dependencies between synchronized areas of the brain are nonlinear. This explanation is used for temporal objects (events) (layer 1). Deeper layers of the experience of time are disclosed by phenomenological analysis: felt passage of time (layer 2) and temporality itself (layer 3), in which is embedded the experience of the passage of time and events in time. These deeper layers can be explained properly by landscapes in phase spaces (2) and by the open character of emotion and disposition (3).

26. Why do qualia exist? • He sniffs it and carefully nibbles it. • The taste cortex sends out RFC, a request for comment. • Memory is distributed, the sought associations are in the whole brain. • RFC appears in working memory (WM) on the level of global brain dynamics. • WMis not large, it holds only several images (7±2in humans). • There arise resonance states activating memory trails. • The strongest dominates: bad associations! poison! spit! • A strong physiological reaction begins – perception serves action • The episodic state of WMis remembered in long-term memory. • A rat has various "feelings”for various tastes, likes to be tickled. If a rat could talk, how would he describe this episode? • The results of nonsymbolic, continuous processes, eg. distinguishing tastes, are remembered and associated with the organism's reactions: qualia! Let's imagine a rat sniffing food. In a fraction of a second he must decide: eat or spit?

27. More on sensations What happens with the taste of ice cream? Taste buds send information the whole time; the brain processes it, but qualia disappear after a short time. Why? Working memory completes many objects and if it has no resonance with the taste cortex then there is no sensation. • Adaptive resonance: ascending (senses=>conceptions) and descending (conceptions=>senses) information streams create self-agreed reverberations, temporary states of the brain/mind. • Resonance states are “clothed”: they contain associations, memory trails, actions, all included in one state – very differently than in the case of abstract states of registers of a Turing machine. Long-term memory (LTM) is huge, around 1014synapses. Working memory (WM) is the actualization of several states of LTM, a dynamic phenomenon.

28. When does the system have sensations? • Working memory (WM), based on a dynamic model of a recurrent neural network, should include information allowing for the recreation of the state of all subsystems. • Long-term memory, allowing for the recreation of working memory states • Ability to differentiate between different types of continuously changing states of WM; "differentiate" means to associate with different types of actions, symbols, commentaries. • A mechanism for activating associations remembered in the long-term memory structure, and adding this information to states of WM. • Possibility of actively commenting on states of WM: words, actions. • Distinguishing ‘I’ and ‘other', categorizing values of states of WM from the point of view of the goals (survival) of the system. • Instincts and drives transmitting the general orientation to the system. Every system able to evaluate its working memory states must claim that it has sensations and is aware of them! Associative mechanisms are sufficient for this. Minimal requirements to build such a system:

29. External perspective Qualia have to exist in brain-like structures: Sensations depend on the actions of cognitive mechanisms of the brain; that's why habituation or intense concentration removes qualia, even if the information is accessible to some areas of the brain. Qualia require proper transformation and interpretation of information reaching the brain, eg: segmentation of the visual scene, which can be seen clearly in tests on blindness to changes; without interpretation there are no sensations. The secondary sensory cortex is responsible for interpretation; lesions change qualia, causing eg. sensory asymbolia, you feel something but you don't know what – the sensation doesn't have normal properties (it's similar with emotions, which are harder to interpret). There is no sensation of pain without interpretation of pain signals. Visual sensations: the sensation of color requires specific states of V4 areas analyzing colors, damage to this area causes the disappearance of color sensations, awaken and in dreams.

30. Memory and sensations Memory is necessary for the interpretation of brain states: Qualia should change under the influence of agents which influence memory. Perceptual training influences the way in which we receive sensations; remembering new sounds/tastes/objects changes qualia. New qualia also appear in dreams. Why doesn't tying your shoelaces have a taste? Only episodic memory creates resonances, procedural memory doesn't. Instances of bad or untypical interpretation of information by the brain lead to numerous strange sensations and behaviors, like: synesthesia – mixing of sensations of different senses; blidvision – residual sight without consciousness; hemispatial neglect, inability to remind oneself, and ignoring half an area or half one's body; body dysmorphia – suffering because of one's body; phantom limbs controlled by mirror reflections; and many others.

31. Ending How does the brain create the mind? Or how do neurons create our thoughts, behaviors, emotions etc?

32. Foundations Neurons: sensors, activation/inhibition/charge loss, threshold value, impulses Networks: transformations, image completion, amplification, attractor dynamics, constraint fulfillment Learning: modeling and goals Recognition/Attention: transformations, spatial object pathways Memory: weights vs. actions, brain cortex, hippocampus, frontal cortex Language: transformations, integration of spelling, semantics and phonetics Higher recognition functions: transformations based on activation, threshold values

33. How does the brain create the mind? Models allow us to understand phenomena: oriented stripes of light I V1, specialized memory systems, division of tasks in reading, effects of damage Models deal with complexity: image recognition, semantics Models are transparent, they unfold psychological constructions: memory mechanisms - disconnection/inhibition vs. activation Models can be controlled: Models lead to a uniform approach:

34. Summary Psychology changes completely thanks to neurocognitive methods! Most explanations presented until now by psychologists don't really explain much, postulating mechanisms which don't exist. Example: categorization and inverse frequencies from the point of view of neural models and from the point of view of psychology. Attractor neural networks and concept formation in psychological spaces.

35. Remaining Problems Connections: Temporal control: Missing brain regions: Scaling: Error signals: Models are too simple/complex: Models can accomplish everything/not much:

36. What now? What do these models say about nature vs. nurture? How would these models reflect individual differences? Would the ideal model be exactly like the brain? Would we be able to predict behavior with complete certainty? Is there room for free will? Can these models represent emotions? Experiences? Consciousness? Do these models dispose of God?