DR.GEEK
Estimation and Learning of generation model
11th-October-2020 For estimation, use conditional probability of hidden layer when input is given Benefits of generated models Input data...
top of page
Search
DR.GEEK
Deep Learning using generation model
10th-October-2020 • In the deep learning method described so far, the behavior of the network is described deterministically • In the...
DR.GEEK
Model of V2
3rd-October-2020 V2 model with sparse RBM - Reproduce realistic V2 neuron response [Ito - Komatsu 04] - Respond to more complex shapes...
DR.GEEK
Visual information processing of the brain
2nd-October-2020 Structure of the visual cortex (ventral cortical visual pathway) - propagation with feed forward - hierarchy: simple...
DR.GEEK
Local contrast normalization
1st-October-2020 Model of Behavior of Visual Field Neuron - Even in Image Recognition - Explanation Based on Natural Image Statistics...
DR.GEEK
Visual information processing of the brain
(29th-September-2020) Gabor wavelet - position / orientation / scale - Topographic map Sparse Coding is here.
DR.GEEK
Improved example of CNN: Tiled CNN
(28th-September-2020) CNN can acquire only simple shift invariance - Local receptive field + tied weights • Tiled CNN - Different filter...
DR.GEEK
Random filter: the importance of architecture
(27th-September-2020) Let the filter be random, learning only the uppermost layer fully-connected layer The architecture is much more...
DR.GEEK
method and pooling case study
(26th-September-2020) "Further data and more early GPUs can gain immediate performance improvement" [Krizhevsky + 12] • Large-scale...
DR.GEEK
Non-NN method and pooling
(25th-September-2020) Familiar with general object recognition - SIFT [Rowe 99] (Descriptor generation) - Bag of Features - Spatial...
DR.GEEK
Convolutional Neural Network (CNN) Cont.
(24th-September-2020) Convolution + pooling is repeated (= Deep CNN) to acquire invariance to various deformations • Learnable - target:...
DR.GEEK
Convolutional Neural Network (CNN)
(23rd-September-2020) Neocognitron's Roots [Fukushima 80] • Backpropagation Application to Learning and Handwritten Character Recognition...
DR.GEEK
Effect of pre-Learning
(22nd-September-2020) • Pre-training prevents excessive learning - Weight obtained by pre-training as initial value • All...
bottom of page