CS 182Midterm reviewOutline 1/3•Biology–Neuron mechanics•spiking, neurotransmitters, ion channels, Na+– Development• growth cones, chemical gradients, activity-dependent pruning• Models–McCullough-Pitts model•backpropagation: minimize error by slow steps–triangle node•recruitment learningOutline 2/3• Psych experiments–Stroop task, priming, spreading activation• Brain study–PET, fMRI, EEG, MEG, TMS, lesion, single-unit• Learning– backprop, Hebbian, LTP, calcium, retrograde, blank slate (not)•Colors– language variations and invariants; intermediate representationsOutline 3/3• Categories–basic categories: functional distinction & effects–category structure• Schemas–Image schemas–FrameNet–Regier's model•hand-built feature extraction, learned mappingNeurobiology• Ion pumps–polarized–resting potential–ions: Na+, K+, Cl-• Neurotransmitters–voltage admits Ca+, vesicles to release transmitters–bind to ionotropic receptors– allow ions into postsynaptic cell• different types of ions for different receptors/transmittersNeurobiology 2• Action potential–threshold–all-or-nothing–voltage-gated channels–hysteresis & refractory period• Myelin–electrical transmission–only in mammalsNeural development• Neurons migrate to correct positions•Grow axons with “growth cones” on tips• Axons follow chemical gradients in increasing or decreasing directions, attracting or repelling them– brain areas grow axons generally together• Axons meet other neurons, form synapses–overconnect–prune based on activity– humans not blank slate!The McCullough-Pitts Neuronyj: output from unit jWij: weight on connection from j to ixi: weighted sum of input to unit ixifyjwijyixi = ∑j wij yjyi = f(xi)ti : targetBackpropagation• Minimize squared error–gradient descent–use momentum (constant times previous change)• Minimum error on training set may not be minimum error on all data!–overfitting•memorize training data at expense of generalization–reduce it by checking against a validation set or just stopping earlyTriangle nodes• Triangle nodes–Reciprocal connections to three other units–Activates whenever 2 units activate–Activates all 3 units–Can be implemented as group of McCullough-Pitts neurons•Recruitment learning– Intermediate units activated by multiple sources strengthen connections to sources–Now intermediate unit represents conjunctionHebbian learning• Neurons that fire together, wire together–If presynaptic neuron helps cause postsynaptic neuron to fire, synapse strengthened• Biology–Ca2+ enters postsynaptic cell when synapse active and cell fires–Temporarily makes receptors more sensitive; also slowly adds new receptors–releases nitrous oxide (NO)--retrograde messenger• causes changes in presynaptic neuron, too• releases more neurotransmitterPsych experiments• Measure behavioral differences that indicate processing differences–Stroop task•difficulty in reading words in other color•indicates topic-specific interference in language processing–Word priming•related words make recognizing words faster•timing-dependent effects– opposite words speed up without time to process; slow down processing with more time– explained by slower-onset inhibitionBrain study• PET•fMRI• lesion• EEG• single-unit recording• TMSColors• Biology–3 input signals– neural representations of sums & differences• e.g. yellow = green + red – blue•Psychophysics– quantitative measurements of psychological quantities– e.g. find out how small a color difference subjects can discern• Language– different languages have different groups– certain groupings maintained: no one groups red + green– certain central concepts maintained: no one says turquoise is a good example of “grue” (green-blue color
View Full Document
Unlocking...