Post on 21-Dec-2015
EE1411
Information processing Information processing by the brainby the brain
Janusz A. Starzyk
Computational IntelligenceComputational Intelligence
Based on a course taught by Prof. Randall O'Reilly University of Colorado and Prof. Włodzisława DuchaUniwersytet Mikołaja Kopernika
EE1412
Basic mechanismsBasic mechanismsMicroorganization: basic rules, similar in the whole brain. Macroorganization: diversification and interactions of different areas. On the micro level in the Leabra model we have 6 rules:
EE1413
RulesRules The brain is not a universal computer. Neurons adjusted evolutionally to detect specific properties of
analyzed signals. Compromise between specificity and built-in expectations, and
generality and universality. Compromise between speed of the hippocampus representing
temporal sequences, and slowness of the cortex integrating many events.
Compromise between active memory and control of understanding.
How to build, using neurons, all necessary elements - specific and universal?
Dynamic rules on the macro level: Constraint satisfaction (including internal), knowledge a priori. Contrast reinforcement, attractors, active memory. Attention mechanisms, inhibitory competition.
EE1414
MacrolevelMacrolevel
Neuron-detector layers strengthening/weakening differences. Hierarchical transformation sequences. Special transformations for different signals. Specialized information transfer pathways. Interactions within pathways. Processing and memory built into the same hardware Higher-level association areas. Distributed representations across large areas.Strong feedback between areas causes this to be only approximatedifferentiation, yielding representation invariance, specialization and
hierarchy.
EE1415
Hierarchy and specializationHierarchy and specializationMental processes: the result of hierarchical and specialized
transformation of sensory signals, internal states (categories) and undertaken actions.
Neuron-detector layers process signals coming to them from receptors, strengthening/weakening differences.
Emerging internal states provide interpretations of environmental states - hierarchical processing is necessary to attain invariant representations, despite variable signals, eg. aural (phonemes), or visual (colors, objects).
Transformations and specialized information processing streams stimulate internal representations of categories and provide data for taking action, e.g. motor reactions. Simultaneously, processed information modifies the means of information processing.
EE1416
Distribution and interactionDistribution and interactionSpecialization increases efficiency of activity, but interactions between streams are essential for coordination, acquiring additional stable information on different levels, e.g.. spatial orientation and object recognition.
On a higher level we have heterogenic association areas.
Knowledge linked to recognition (e.g. reading words) is distributed across the whole brain, creating a semantic memory system.
It's similar on a micro and macro level: interpretation of the whole is the result of distributed activity of many elements.
Knowledge = processing,
Program ~ data.
EE1417
Dynamic principlesDynamic principlesWell-known inputs trigger an immediate reaction.New ones may require iterative searches for the best compromise satisfying constraints resulting from possessed knowledge = possible to attain dynamic states of the brain. There exist many local, alternative or sub-optimal, solutions => local context (internal) changes the interpretation.
Time flies like an arrowFruit flies like a banana
Long-term memory is the result of learning, this is synaptic memory.
Active memory (dynamic) is the result of momentary mutual activations of active areas; it's short-term because the neurons get tired and are involved in many processes; this directly influences processes in other areas of the brain.
This mechanism causes the non-repeatability of experiences = internal interpretations, contextual states are always somewhat diverse.
Concentration is the result of inhibitory interactions.
EE1418
General functions of the cortexGeneral functions of the cortex
Brodmann's areas of the cortex
Four cortical lobes and their functions
Various terms used to refer to locations in the brain
EE1419
General functions of the cortexGeneral functions of the cortex Four lobes of the cortex: frontal lobe occipital lobe parietal lobe temporal lobe
The frontal lobe is responsible for: planning, thinking, memory, willingness to act and make decisions, evaluation of emotions and situations, memory of learned motor actions, e.g. dance, mannerisms, specific patterns of behavior, words, faces, predicting consequences, social conformity, tact, feelings of serenity (reward system), frustration, anxiety and stress. The occipital lobe is responsible for: sight, analyzing colors, motion, shape, depth, visual associations
EE14110
General functions of the cortexGeneral functions of the cortex
parietal lobe temporal lobe
The parietal lobe is responsible for: spatial orientation, motion recognition, feeling temperature, touch, pain, locating sensory impressions, integration of motion, sensation and sight, understanding abstract concepts. The temporal lobe is responsible for: speech, verbal memory, object recognition, hearing and aural impressions, scent analysis.
EE14111
Subcortical areasSubcortical areas
Brain stem:
raphe nuclei: serotonin,
reticular formation: general
consciousness.
Midbrain: (mesencephalon):
part of the ventral tegmental
area (VTA): dopamine,
value of observation/action.
Thalamus: input of sensory signals, attention
Cerebellum: learning motion, temporal sequences of motion.
EE14112
Subcortical areasSubcortical areas
Amygdala: emotions, affective associations. Basal ganglia: sequences, anticipation, motor
control, modulation of prefrontal cortex activity,
selection and initiation of new activity. Hippocampus: fast learning, episodic and spatial
memory.
Basal ganglia (striatum, globus pallidus, substantia nigra)Basal ganglia initiate motor activities and the substantia nigra is responsible for controlling learning
EE14113
3 principle brain areas3 principle brain areasPosterior cortex PC – rear parietal cortex and motor cortex; sensorymotor actions, specialization, distributed representations
Frontal cortex FC – prefrontal cortex, higher cognitive behaviors, isolated representations
Hippocampus HC – hippocampus and related structures, memory, rapid learning, sparse representations.
Learning must be slow in order to grasp statistically important relationships, and
to precisely analyze sensory data and control motions, but we also need a
mechanism for rapid learning. Compromise: slow learning in the cortex and rapid learning in the hippocampus.Retaining active information and simultaneously accepting new information in a
distributed system, avoiding interference.
EE14114
Slow/rapid learningSlow/rapid learningA neuron learns conditional
probability, the correlation
between desired activity and
input signals; the optimal value
of 0.7 is reached quickly only
with a small learning constant of
0.005
Every experience is a small fragment of uncertain, potentially useful knowledge
about the world => stability of one's image of the world requires slow learning,
integration leads to forgetting individual events. We learn important new information after one exposure. Lesions of the hippocampus trigger follow-up amnesia. The system of neuromodulation reaches a compromise between stability and
plasticity.
EE14115
Active memoryActive memoryDistributed overlapping representations in the PC can
efficiently record information about the world, but...
having too many associations and connections
decreases the possibility of precise discovery of
information, it can also blur it with the passage of time.
FC – prefrontal cortex, stores isolated representations;
greater memory stability.
Inhibition => active memory must be selective, the effect is a focusing
of attention.
Attention is not a result of the activity of separate mechanisms
connected with the will, it's an emergent process resulting from the
necessity of fulfilling many constraints simultaneously.
EE14116
Cognitive architectureCognitive architecture
Hierarchical structure for sensory data, recurrence in FC, recording
the context.
EE14117
Activity Activity
Parietal cortex: learns slowly, creates extensive, overlapping
representations in a densely connected network.
Dynamic PC states are short-term memory, mainly of spatial relations,
quickly yielding to disorder and disintegration.
Frontal cortex: learns slowly, stores isolated representations, activation
of memory is more stable, the reward mechanism dynamically switches
its activity, allowing a longer active memory.
The hippocampus learns quickly, creating sparse representations,
differentiating even similar events.
This simplified architecture will allow the modeling of many
phenomena relevant to perception, memory, using language, and the
effects of the interaction of different areas.
EE14118
Controlled/automatic actionControlled/automatic actionAutomatic: routine, simple, low level, sensory-motor, conditional reflexes, associations – easy to model with a network.
Controlled: conscious, elastic, requiring sequences of actions, selection of elements from a large set of possibilities – usually realized in a descriptive way with the help of systems of rules and symbols.
Models postulating central processes: like in a computer, working memory with a central monitor, having influence over many areas.
Here: emergent processes, the result of global constraint fulfillment, lack of a central mechanism.
The prefrontal cortex can exert control over the activity of other areas, so it's involved in controlled actions, including the representation of "me" vs. "others", social relationships etc.
EE14119
Other distinctions - consciousnessOther distinctions - consciousness Declarative vs. procedural knowledge
Declarative: often expressed symbolically (words, gestures). Procedural: more oriented towards sequences of actions.
Explicit vs. implicit knowledge
Controlled action relies on explicit and declarative knowledge.Automatic actions rely on implicit and procedural knowledge.
Consciousness => states existing for a noticeable period of time, integrating reportable sensory information about different modalities, with an influence on other processes in the brain.
Each system, which has internal states and is complex enough to comment on them, will claim that it's conscious.
Processes in the prefrontal cortex and the hippocampus can be recalled as a brain state or an episode, can be interpreted
(associated with concept representation).
EE14120
Various potential problemsVarious potential problemsThere are easy things, for which simple models will suffice, and difficult things requiring detailed models.
Many misunderstandings: MLP neural networks are not brain models, they are only loosely inspired by a simplified look at the activity of neural networks; an adequate neural model must have appropriate architecture and rules of learning.
Example: catastrophic forgetting of associations from lists, much stronger in MLP networks than in people => appropriate architecture, allowing for two types of memory (hippocampus + cortex) doesn't have a problem with this.
Human cognition is not perfect and good models allow us to analyze the numerous compromises handled by the brain.
Brains are fairly elastic, although they mostly base their actions on the
representation of specific knowledge about the world.
EE14121
Problem of integrationProblem of integration Binding problem: we perceive the
world as a whole, but information in
the brain, after initial processing,
doesn't descend anywhere. Likely synchronization of distributed
processes. Attention is a control mechanism
selecting areas which should be
active in a given moment. Encoding relevant combinations of
active areas.
Simultaneous activity = dynamic synchronization, partial reconstruction
of the brain state during an episode.
Integration errors happen often.
EE14122
ChallengesChallenges Disruptions: Multi-level transition from one activity to another and
back to the first, or recurrent multiple repetition of the same activity. This is easy for a computer program (loops, subroutines), where
data and programs are separated, but it's harder for a network,
where there is no such separation. PFC and HCMP remember the previous state and return to it. Difficult task, we often forget what we wanted to say when we listen
to someone, sentences are not nested too deeply.
The rat the cat the dog bit chased squeaked.
How and what should be generalized? Distributed representations
connect different features.
Dogs bite, and not only Spot, not only mongrels, not only black dogs...