Feature Nets and Word Recognition PDF

Preview

Summary

Feature nets in terms of reading and language ...

Description

1. Feature Nets and Word Recognition 2. The Design of a Feature Net a. Feature nets A system for recognizing patterns that involves a network of detectors, i. with detectors for features as the initial layer in the system. ii. There is a network of detectors organized in layers 1. The bottom layer is concerned with features 2. As we more up in the network, each layer is concerned with larger scale objects a. This is similar to bottom up processing b. Activation level A measure of the current status for a node or detector. i. ii. Activation level is increased if the node or detector receives the appropriate input from its associated nodes or detectors; activation level will be high if input has been received frequently or recently. iii. Activation is like the cells, when one detector is energized, others related in the system will start to energize as well. The activation level increases until it reaches a response threshold 1. Recency and frequency will determine how much it takes for a detector to be activated c. Response threshold i. The quantity of information or activation needed to trigger a response ii. When the detector will fire 3. The Feature Net and Well-Formedness a. Bigram detectors i. Hypothetical units in a recognition system that respond, or fire, whenever a specific letter pair is in view 4. Recovery from Confusion a. This is part of the process where you are unfamiliar with one of the detectors, it is solved by the bigram detectors stage though 5. Ambiguous Inputs 6. Recognition Errors a. Occurs because the network is biased, inevitably favoring frequent letter combinations over infrequent ones. b. network operates on the basis of “when in doubt, assume that the input falls into the frequent pattern. c. bias will pull the network toward errors if the input happens to have an unusual spelling pattern, but (by definition) these inputs are less common in your experience. i. Hence, the network’s bias necessarily helps perception more often than it hurts. 7. Distributed Knowledge a. Locally represented i. Not stored in a particular location or built into a specific process

b. Distributed knowledge information stored via a distributed representation i. ii. look at the relationship between their levels of priming, and we also need to look at how this relationship will lead to one detector being more influential than the other c. Efficiency versus Accuracy i. errors as the price you pay in order to gain the benefits associated with the net: If you want a mechanism that’s able to deal with unclear or partial inputs, you simply have to live with the fact that sometimes the mechanism will make mistakes...