Cognitive Neuroscience and Embodied Intelligence Neurons and Their Connections Based on book Cognition, Brain and Consciousness ed. Bernard J. Baars Janusz A. Starzyk EE141 1 Introduction Neurons did not change much for millions of years The brain can be viewed as a hyper complex surface of neurons. Sensory and motor cortex are viewed as processing hierarchies of neurons. 2 EE141 Introduction A single neuron may have thousands of inputs (dendrites) and one or more outputs (axons). Neurons grow extending their axons and connecting to other neurons in the interconnected structure A bipolar neuron 3 EE141 Neurons’ Growth This growth can be observed in the lab and under stimuli the network can learn a control function 4 EE141 Real and idealized neurons Neurons have been idealized into the classical integrate and fire neuron (right). In this neuron inputs from dendrites are accumulated and if total voltage value exceeds -50 mV it triggers fast traveling action potential in the cell’s axon. Neuron sends its signal by firing spikes from the cell body to terminals synapses. EE141 5 Excitation and Inhibition Classical neurons are connected by excitatory and inhibitory synapses. There are many classes of neurons, neurochemicals, and mechanisms for information processing Many factors determine neuron activity – the sleepwaking cycle, availability of chemicals, and more. 6 EE141 Excitation and Inhibition (cont.) Transmission of signals through axons is assisted by wrapping the axons in Myelinating Schwann cells. The cells improve the conduction velocity of signals. At the breaks known as the nodes of Ranvier, the action potentials are regenerated. EE141 7 A Synapse A spike in the presynaptic cell triggers release of neurotransmitter that diffuses across the synaptic gap and changes potential of postsynaptic cell. Efficiency of signal transmission corresponds to synaptic weigh in network of neurons 8 EE141 Working Assumptions Neurons adds graded voltage inputs until total membrane voltage exceeds -50 mV and then fires. Connections are either excitatory or inhibitory and its strengths is represented by the connection weight. The weight can be normalized between -1 and 1. Artificial neural networks that use simple neuron models can be used for pattern recognition or unknown function approximation. Neurons can form one-way or bidirectional pathways to transfer information from one part of the brain to other. Cortex is a massive 6-layer array of neurons. Arrays of neurons are called maps. 9 Stable collections of neurons form Hebbian cell assemblies EE141 A simple reflex circuit An example of a spinal (knee-jerk) reflex. Sensory neurons pick up the tap and transmit it to the spinal court. An interneuron links the sensory impulses to motor neurons (bypassing higher level brain function) making the leg jump 10 EE141 A simple reflex circuit While reflex circuits can be triggered by outside stimuli, they are normally integrated into voluntary, goal driven activities. In many time this is unconscious and almost automatic. Voluntary goal driven brain mechanisms, are associated with cortex. Sophisticated subcortical activity is also engaged in planning and executing actions. Spinal centers also communicate with higher centers while carrying sensorimotor reflexes and return feedback signals to brain. 11 EE141 Different types of receptors There are several types of receptors, however, they are all similar in structure and function. Sensory nerves have parallel pathways sending sensory information to thalamus and sending back feedback information EE141 90% of neurons go backwards towards the source Most sensory and motor pathways split and cross over the midline of the body 12 Similarities between sensor pathways This image shows the similarities between the different sensory streams. arm vs. leg, high frequency vs. low frequency, and foveal vs. peripheral vision. 13 EE141 Sensory Interactions Sensory regions interact with thalamic nuclei (RTN) Notice similarities between cortical input and output layers 14 in all these senses EE141 Lateral Interactions Lateral This inhibition is used to differentiate between neighboring cells gives better resolution at various levels of sensory perception In retina it helps to spot a tiny point At higher level it helps to differentiate e.g. between ‘astronomy’ and ‘astrology’ 15 EE141 Lateral Interactions Visual demonstration of lateral inhibition Notice that lateral inhibition applies to adjacent black squares, color perception, and even perception of direction 16 EE141 Mapping of the brain Visual quadrants map to cortical quadrants Mapping is observed for various senses 17 EE141 Neuron organization EE141 Neurons organize into layers. The figure below shows a single layer of pyramid neurons at 200 micrometers. 18 Visual Maps Neuron connections form various pathways In V1 the upper pathway is sensitive to location ‘where The lower pathway is sensitive to color, shape contrast and object identity ‘what’ EE141 19 Layers have 2-way connections EE141 Neuronal layers have both feed-forward and feedback connections between layers/arrays. Lower levels tend to be sensitive to simpler stimuli, while higher 20 levels respond to more complex stimuli. Sensory and motor hierarchies Sensory and motor systems appear to be arranged in hierarchies with information flowing between each level of the sensory and motor hierarchies. 21 EE141 Ambiguous stimuli EE141 Ambiguous stimuli pose choices for interpretation. It all depends on how the image is perceived and what ever preconceived notions you may have. 22 Hebbian Learning “Neurons the fire together, wire together” Long term potentiation (LTP) and long term depression (LTD) The figure depicts Hebbian learning in cell assemblies. At t1 input is encoded into connection weights. Memory is retained at times t2 & t3. 23 EE141 A simple network Note the combination of excitory and inhibitory connections. 24 EE141 A Three Layer Network Hidden layer makes the network more flexible Backpropagation is used to adjust network weights to match the input to a desired output. 25 EE141 A pattern recognition network EE141 An example of an auto-associative network that matches its output with its input. 26 A self-organizing network Self-organizing networks appear often in biological organisms. A self-organizing network can be used for face recognition. 27 EE141 Neural Darwinism Gerald Edelman proposed that brain is a massive selectionist organ where neurons develop and make connections following Darwinian principle of selection of the fittest. In biological evolution, species adapt by reproduction, mutation that leads to diverse forms, and selection. A similar process occurs in the immune system, where millions of immune cells adapt to invading toxins. 28 Thus selectionism leads to flexible adaptation. EE141 Neural Darwinism Selectionist learning principle was used to train simulated rat to find a hidden platform in a Morris water maze The second figure shows a selectionist robot that learns to play soccer. EE141 29 Symbolic Processing Neural nets can handle both distributed numerical values as well as symbolic expressions. The figure shows proposed by McClelland and Rogers merge between symbolic features and their associations expressed by connections of a neural network Brain uses adaptation and representation to learn the world. 30 EE141 Coordinating Neural Nets Neurons’ activation is coordinated by large-scale rhythms to signify their activities. Epileptic seizures are also caused by slow, intense, regular waves that lead to a loss of consciousness Thus there must be a balance between integration and differentiation. A high density of gamma rhythms has been related to conscious visual perception and understanding of spoken words. Alpha rhythms are associated with an absence of focused attentional tasks. Theta rhythms coordinate hippocampal region and the frontal cortex during retrieval of memories. And delta rhythms signal deep sleep, are believed to group fast neuronal activities to consolidate learned events. 31 EE141 Coordinating Neural Nets This figure illustrates hypothesis how brain rhythms coordinate large number of neuron cells’ firing. Neurons that fire in synch with the dominant rhythm are strengthened by feedback from many other neurons, while those that fire out of 32 synch are weakened. EE141 Summary The basic question in cognitive neuroscience is how the nerve cells work together to perform cognitive functions like perception, memory and action. Models of neurons were developed and used to build functional processing networks. Artificial neural networks and biologically inspired networks are useful to study cognitive processing. Sensory and motor systems are complex hierarchies of neurons organized in two or three dimensional arrays. In vision, touch and motor control arrays of neurons are topographically arranges as maps of the spatial surroundings. Hierarchies are bidirectional pathways, that allow signals to travel up, down and laterally. A major function of downwards pathway is to resolve sensory ambiguities. Lateral inhibition is used to emphasize differences between inputs. 33 EE141