Group:
artificial intelligence
Group:
parallel processing
Topic:
actor machines
Topic:
networks of relays
Topic:
Petri net
Topic:
reflex circle
Topic:
self-regulating systems
Topic:
thought is computational
| |
Summary
An artificial neuron net is a network of threshold devices and a firing rule. There are two main designs. The original, by McCullough and Pitts emphasized the logical equivalence of a neuron net to logical propositions. The McCullough and Pitts model was used for the first stored-program computer. The other kind of neuron net was started by Hopfield. Here it is the dynamic behavior of the network that's most important. The network is seen as storing an associative memory, somewhat like a hologram. Stimulation activates the matching memory.
Both styles of neuron nets depend heavily on a representational context. Internally they are a system like any other. So someone needs to identify the correspondence between states and reality. Unfortunately that someone is generally the experimenter. (cbb 5/94)
Subtopic: artificial neurons 
| Quote: an unorganized machine has a large number of randomly connected, similar units (e.g., NAND gates); simple model of nervous system [»turiAM9_1947]
| | Quote: a neuron is a threshold device within a period of latent addition, synaptic delay, refractory period, and inhibition [»mccuWS_1943]
| | Quote: the specificity of a nerve's impulses depends solely on their time and place
| | Quote: the firing of a neuron is equivalent to the assertion of a simple proposition [»mccuWS_1943]
| | Quote: an E-element models a MacCulloch/Pitts neuron; fixed synaptic delay for synchronous operation, thresholds 1,2,3 and absolute inhibition [»vonnJ6_1945]
|
Subtopic: power of neuron nets
| Quote: collective, computational properties can arise from a network of simple neurons and little structure [»hopfJJ4_1982]
| | Quote: acyclic neuron nets and temporal propositional expressions are equivalent [»mccuWS_1943]
| | Quote: Turing machines and neuron nets are equivalent, also Church's lambda-definability and Kleene's primitive recursiveness [»mccuWS_1943]
| | Quote: all of psychology is based on the activity of neuron nets, i.e., on two-valued logic [»mccuWS_1943]
| | Quote: with determination of the neuron net, the unknowable object of knowledge, the "thing in itself," ceases to be unknowable [»mccuWS_1943]
|
Subtopic: state
| Quote: the state of a neuron net determines its future course of action; but given a behavior can not determine the cause [»mccuWS_1943]
| | Quote: the state of a neuron net consists of the afferent stimulation and the activity of each neuron
|
Subtopic: neural nets
| QuoteRef: knigK11_1990 ;;thorough review on connectionist theories and neural networks.
|
Subtopic: Hopfield nets
| Quote: can regard a physical system as a content-addressable memory, if it has many locally stable states; also need to create stable states [»hopfJJ4_1982]
| | Quote: Hopfield's model for content addressable memory consists of neurons that fire with some mean attempt rate [»hopfJJ4_1982]
| | Quote: Hopfield's model differs from Perceptrons by strong back-coupling, emergent computational properties, and asynchronous operation [»hopfJJ4_1982]
| | Quote: Hopfield's networks perform abstract calculations on coded inputs; feature extraction should be already performed [»hopfJJ4_1982]
| | Quote: a neuron net has stable limit points like an Ising model; phase space is dominated by attractors that represent memories [»hopfJJ4_1982]
| | Quote: by raising the threshold can make 0000 represent all unfamiliar states [»hopfJJ4_1982]
| | Quote: memories are gestalts recalled from a subpart; ambiguities resolved statistically
|
Subtopic: training a neuron net
| Quote: learn by altering the structure of a neuron net; e.g., enable a synapse if both sides are simultaneously excited [»mccuWS_1943]
| | Quote: purposive behavior is by system reducing the difference between afferents and neural net activity [»mccuWS_1943]
|
Subtopic: TD-Gammon, self-training neuron net
| Quote: TD-Gammon is a self-training neural network for backgammon that outperforms other programs and, sometimes, human experts [»tesaG3_1995]
| | Quote: TD-Gammon is a multilayer perception network with 40 hidden units and backgammon feature encoders [»tesaG3_1995]
| | Quote: the goal of temporal difference methods is to match the learner's current prediction for a pattern with the next prediction at the next time step [»tesaG3_1995]
| | Quote: with just raw-encoding, TD-Gammon was a strong intermediate after 200,000 training games
| | Quote: human experts use TD-Gammon to evaluate the best move for a position by playing the position to completion several thousand times [»tesaG3_1995]
|
Subtopic: problems with neuron nets
| Quote: a Perceptron classifies all images with one language; e.g., A is stimulated more frequently by the 2nd class [»bongM_1967]
| Quote: a Perceptron has no notion of irrelevant objects or trash; only uncertainty
|
Related Topics
Group: artificial intelligence (14 topics, 509 quotes)
Group: parallel processing (41 topics, 1125 quotes)
Topic: actor machines (2 items)
Topic: networks of relays (27 items)
Topic: Petri net (44 items)
Topic: reflex circle (20 items)
Topic: self-regulating systems (23 items)
Topic: thought is computational
(60 items)
|
