Deep learning - Conceptual understanding and applications
•
91 likes•75,587 views
This document provides an overview of deep learning, including conceptual understanding and applications. It defines deep learning as a deep and wide artificial neural network. It describes key concepts in artificial neural networks like signal transmission between neurons, graphical models, linear/logistic regression, weights/biases/activation, and backpropagation. It also discusses popular deep learning applications and techniques like speech recognition, natural language processing, computer vision, representation learning using restricted Boltzmann machines and autoencoders, and deep network architectures.
Report
Share
More Related Content
What's hot
What's hot (20)
Viewers also liked
Viewers also liked (15)
Similar to Deep learning - Conceptual understanding and applications
Similar to Deep learning - Conceptual understanding and applications (20)
More from Buhwan Jeong
Recently uploaded
Recently uploaded (20)
Deep learning - Conceptual understanding and applications
- 1. Deep Learning Conceptual Understanding & Application Variants ben.hur@daumkakao.com The aim of this slide is at conceptual understanding of deep learning, rather than implementing it for practical applications. That is, this is more for paper readers, less for system developers. Note that, some variables are not uniquely defined nor consistently used for convenience. Also, variable/constant and vector/matrix are not clearly distinct. More importantly, some descriptions may be incorrect due to lack of my ability.
- 2. Deep Learning (DL) is, in concise, a deep and wide neural network. That is half-true.
- 3. Deep Learning is difficult because of lack of understanding about — artificial neural network (ANN) — unfamiliar applications.
- 4. DL — ANN — Signal transmission between neurons — Graphical model (Belief network) — Linear regression / logistic regression — Weight, bias & activation — (Feed-forward) Multi-layer perceptron (MLP) — Optimization — Back-propagation — (Stochastic) Gradient-decent
- 5. DL — Applications — Speech recognition (Audio) — Natural language processing (Text) — Information retrieval (Text) — Image recognition (Vision) — & yours (+)
- 6. Artificial Neural Network (ANN)
- 7. ANN is a nested ensemble of [logistic] regressions.
- 8. Perceptron & Message Passing
- 10. 5 5 y = x Message passing
- 11. y = Σ x y = Σ wx y = Σ wx + b Linear regression
- 12. y = sigm(Σwx + b) Logistic regression Σwx?Σwx
- 13. Activation function — Linear g(a) = a — Sigmoid g(a) = sigm(a) = 1 / (1 + exp(-a)) — Tanh g(a) = tanh(a) = (exp(a) - exp(-a)) / (exp(a) + exp(-a)) — Rectified linear g(a) = reclin(a) = max(0, a) — Step — Gaussian — Softmax, usually for the output layer
- 14. Bias controls activation. y = ax b y = ax + b
- 16. Linearly non-separable by a single perceptron
- 17. Input layer Hidden layer Output layer http://info.usherbrooke.ca/hlarochelle/ neural_networks/content.html Interconnecting weight Bias Activation function
- 20. {x, y} h=a(wx+b) y'=a2(w2h+b2) l = y - y'w2 & b2w & b w2 R & b2 RwR & bR y' Back-propagation {x, y} —> feed-forward <— back-propagation
- 21. Reinforced through epoch ANN is a reinforcement learning.
- 23. Stochastic Gradient Descent The slide excludes details. See reference tutorial for that.
- 24. ANN is a good approximator of any functions, if well-designed and trained.
- 25. Deep Learning
- 26. DL (DNN) is a deep and wide neural network. many (2+) hidden layers many input/hidden nodes
- 27. DEEP WIDE Large matrix Many matrix Image from googling
- 28. Brain image from Simon Thorpe Other from googling
- 29. Many large connection matrices (W) — Computational burden (GPU) — Insufficient labeled data (Big Data) — Under-fitting (Optimization) — Over-fitting (Regularization)
- 30. Unsupervised makes DL success.
- 31. Unsupervised pre-training — Initialize W matrices in an unsupervised manner — Restricted Boltzmann Machine (RBM) — Auto-encoder — Pre-train in a layer-wise (stacking) fashion — Fine-tune in a supervised manner
- 32. Unsupervised representation learning — Construct a feature space — DBM / DBN — Auto-encoder — Sparse coding — Feed new features to a shallow ANN as input cf, PCA + regression
- 36. visible hidden
- 38. if k <= n (single & linear), then AE becomes PCA. otherwise, AE can be arbitrary (over-complete). Hidden & Output layers plays the role of semantic feature space in general - compresses & latent feature (under-complete) - expanded & primitive feature (over-complete)
- 40. Contractive auto-encoder Regularized auto-encoder Same as Ridge (or similar to Lasso) Regression
- 41. Hinton and Salakhutdinov, Science, 2006 Deep auto-encoder
- 42. (k > n, over-complete) Sparse coding
- 43. X Y
- 44. RBM3 RBM2 RBM1 Stacking Training in the order of RBM1, RBM2, RBM3, …
- 45. Representation Learning Artificial Neural Network+ - Sparse coding - Auto-encoder - DBN / DBM
- 46. Stochastic Dropout Training for preventing over-fitting
- 49. Deep architecture is an ensemble of shallow networks.
- 50. Key Deep Architectures — Deep Neural Network (DNN) — Deep Belief Network (DBN) — Deep Boltzmann Machine (DBM) — Recurrent Neural Network (RNN) — Convolution Neural Network (CNN) — Multi-modal/multi-tasking — Deep Stacking Network (DSN)
- 51. Applications
- 52. Applications — Speech recognition — Natural language processing — Image recognition — Information retrieval Application Characteristic - Non-numeric data - Large dimension (curse of dimensionality)
- 53. Natural language processing — Word embedding — Language modeling — Machine translation — Part-of-speech (POS) tagging — Named entity recognition — Sentiment analysis — Paraphrase detection — …
- 54. Deep-structured semantic modeling (DSSM) BM25(Q, D) vs Corr(sim(Q, D), CTR)
- 55. DBN & DBM
- 56. Recurrent Neural Network (RNN)
- 57. Recurrent Neural Network (RNN)
- 59. Convolution Neural Network (CNN)
- 60. C-DSSM
- 61. CBOW & Skip-gram (Word2Vec)
- 62. Deep Stacking Network (DSN)
- 63. Tensor Deep Stacking Network (TDSN)
- 65. Google Picture to Text
- 67. In various applications, Deep Learning provides either — better performance than state-of-the-art — similar performance with less labor
- 68. Remarks
- 69. Deep Personalization with Rich Profile in a Vector Representation - Versatile / flexible input - Dimension reduction - Same feature space ==> Similarity computation
- 71. Hard work — Application-specific configuration — Network structure — Objective and loss function — Parameter optimization
- 73. Yann LeCun —> Facebook Geoffrey Hinton —> Google Andrew Ng —> Baidu Joshua Benjio —> U of Montreal
- 74. References Tutorial — http://info.usherbrooke.ca/hlarochelle/neural_networks/content.html — http://deeplearning.stanford.edu/tutorial Paper — Deep learning: Method and Applications (Deng & Yu, 2013) — Representation learning: A review and new perspective (Benjio et al., 2014) — Learning deep architecture for AI (Benjio, 2009) Slide & Video — http://www.cs.toronto.edu/~fleet/courses/cifarSchool09/slidesBengio.pdf — http://nlp.stanford.edu/courses/NAACL2013 Book — Deep learning (Benjio et al. 2015) http://www.iro.umontreal.ca/~bengioy/dlbook/
- 75. Open Sources (from wikipedia) — Torch (Lua) https://github.com/torch/torch7 — Theano (Python) https://github.com/Theano/Theano/ — Deeplearning4j (word2vec for Java) https://github.com/SkymindIO/deeplearning4j — ND4J (Java) http://nd4j.org https://github.com/SkymindIO/nd4j — DeepLearn Toolbox (matlab) https://github.com/rasmusbergpalm/DeepLearnToolbox/ graphs/contributors — convnetjs (javascript) https://github.com/karpathy/convnetjs — Gensim (word2vec for Python) https://github.com/piskvorky/gensim — Caffe (image) http://caffe.berkeleyvision.org — More+++
- 76. FBI WARNING Some reference citations are missed. Many of them come from the first tutorial and the first paper, and others from googling. All rights of those images captured, albeit not explicitly cited, are reserved to original authors. Free to share, but cautious of that.
- 77. Left to Blank