PSYC 532 Lecture Notes - Lecture 2: Linear Separability, Seymour Papert, Sigmoid Function
23 Jan 2018
School
Department
Course
Professor
Document Summary
Activity units neurons firing rate real numbers active seconds circles. Connectivity connection weights synapses excitation inhibition real numbers long term seconds to years lines. Unit: running a simple program, compute weighted sum of inputs coming from any other units, and outputs a number, a non-linear function of the weighted sum of inputs. Modifying a connection weight reduces the error component. Neurons are sluggish and noisy processing devices, yet we perceive and recognize in about 0. 5 sec, thus the brain must be a parallel processor. Big variety of neuron types, all have dendrites (multiple input) cell body (computation) and one axon (single output) with synapses (transmission device). Computational properties of the brain: robust & fault tolerant (the neuron will function despite inexact inputs, because disfunctionning would need to come from all the input neurons to lead to pb), flexible, approximate, highly parallel, compact and efficient. Network component: units (activation depends on weighted sum of inputs) + connections (positive or negative, weighted).
