Neural Nets How the brain achieves intelligence 1011

Neural Nets How the brain achieves intelligence 10^11 1 mhz cpu’s

Concerns • Representation – What is it – What can it do • Learnability – How can it be trained • Efficiency – Of Learning – Of Learned concept

Weka’s Neural Net Output on Iris Node 0 <-> Iris-Setosa Threshold -3. 50 Node 3 -1. 00 Node 4 9. 07 Node 5 -4. 10 Node 1 <-> Iris-versicolor Threshold 1. 06 Node 3 3. 89 Node 4 -9. 76 Node 5 -8. 59 Node 2 <-> Iris-viginica Threshold -1. 00 Node 3 -4. 21 Node 4 -3. 62 Node 5 8. 80 Node 3 Threshold 3. 38 sepallength 0. 90 sepalwidth 1. 56 petallength -5. 0 petalwidth -4. 91 Node 4 Threshold -3. 33 sepallength -1. 11 sepalwidth 3. 12 petallength -4. 13 petalwidth -4. 07 Node 5 Threshold -7. 49 sepallength -1. 21 sepalwidth -3. 53 petallength 8. 40 petalwidth 9. 46

Representation : Feed-Forward Neural Net • • DAG of perceptrons Leaf nodes take inputs Outputs node yield decisions Architecture: no one knows how to build them. • Weights: trained by “hill-climbing”; slow and guarantee of only local optimum.

Representational Power • Any boolean function can be represented in disjunctive or conjunctive normal form. • Disjunctive = “or” of “anded” features. • Since perceptron can learn “or” and “and”, 2 -layer network can represent any boolean function.

Neural Nets Work • Disease diagnosis: 90% accurate on prostate cancer prediction • Handwritten Character Recognition (5 -layer net) 99% accurate • Net. Talk: 80 hidden units, 28 inputs. 78% accuracy. Sounds like child learning to talk.

Summary • Neural nets can do multiple classes and regression • Training is slow • Performance is fast and high quality • No one knows how to create architecture • Neural nets tend to be incomprehensible