Prof. Greg Francis 1/5/11 PSY 200: Intro. to Cognitive Psychology 1 Purdue University Neural learning PSY 200 Greg Francis Lecture 08 A problem with virtual reality. Purdue University Networks As we saw last time, a network of neurons can have very complicated behavior The behavior depends on the connections between cells How do those connections get established? 2 3 1 5 4 w45 Purdue University Hebb’s rule If two neurons are active simultaneously, then they strengthen the connection between them Signals from the environment change the properties of the network Purdue University A “simple” model A cell’s activation is on or off (one or zero) Cell connections (weights) are reciprocal Cells update activations one at a time Cell activations are calculated with the rule aw aw aiij jij j=>≤∑∑1 00 0 if if Purdue University Simplified learning Initially, all connections are zero wij = 0 Hebb’s rule cells that are simultaneously active develop positive weights (excitation) an active cell develops negative weights with inactive cells (inhibition) Demonstration Purdue University Self-organization A network of this type does not need an intelligence to set the connection weights The network self-organizes in response to stimulation It can remember things it has previously experienced It can interpret new information on the basis of things it has previously learnedProf. Greg Francis 1/5/11 PSY 200: Intro. to Cognitive Psychology 2 Purdue University Learning This may not be the same type of learning you do when you study for school but it is important just the same Consider the length of your arm to catch and throw objects your brain must know exactly your arm’s length but the length of your arm changes as you age! » And depends on unknown environmental factors Purdue University Hand-eye coordination vision info arm info vision info arm info Purdue University Coordination and learning We do not know the exact nature of the network involved in this coordination but we know it continually modifies part of itself to match up with the current situation This is actually a good design feature, because the brain cannot know in advance every detail of the eye-hand system Purdue University Virtual reality Using computer graphics to convince the body it is someplace other than it really is Useful for architects, designers surgeons, pilots entertainment Purdue University Cameras Enhance visual perception Night vision for helicopter pilots Purdue University Cameras Enhance visual perception MRI overlaid on actual image of brain for surgeon » highlight tumor Avoid other brain regions » fasterProf. Greg Francis 1/5/11 PSY 200: Intro. to Cognitive Psychology 3 Purdue University Problem The network coordinating eye-hand systems, adjusts itself Extended use of the computer cameras makes the user adapt so his eyes are where the cameras are! Purdue University Problem After taking the cameras off, it takes some time to adapt back Eye-hand coordination is off Could be a problem for surgeons and pilots! Purdue University Other adaptations Inverted prisms Fortunately, the adaptations return to normal pretty quickly Kind of like the feeling you get after roller-skating Purdue University Conclusions Learning in neural networks changing connections relatively simple rules Much of our perceptual and motor behavior is based upon this type of continuous learning It’s not clear if more cognitive learning is similar Purdue University Next time Neural networks for visual perception brightness color form Why we see color
View Full Document
Unlocking...