Hinton机器学习与神经网络

机器学习祖师爷之一Geoffrey Hinton大神于2012年主讲录制。课程内容以深刻睿智，鞭辟入里见长。课程深入介绍了Machine Learning中的神经网络的方法，人工神经网络在语音识别和物体识别、图像分割、建模语言和人类运动等过程中的应用，及其在机器学习中发挥的作用。

• 课时1：Why do we need machine learning (13分14秒)

• 课时2：What are neural networks (8分30秒)

• 课时3：Some simple models of neurons (8分23秒)

• 课时4： A simple example of learning (5分38秒)

• 课时5：Three types of learning (7分37秒)

• 课时6：An overview of the main types of network architecture (7分8秒)

• 课时7：Perceptrons (7分56秒)

• 课时8：A geometrical view of perceptrons (6分24秒)

• 课时9：Why the learning works (6分2秒)

• 课时10：What perceptrons can not do (4分38秒)

• 课时11：Learning the weights of a linear neuron (11分55秒)

• 课时12：The error surface for a linear neuron (5分3秒)

• 课时13：Learning the weights of a logistic output neuron (3分56秒)

• 课时14：The backpropagation algorithm (11分51秒)

• 课时15：How to use the derivatives computed by the backpropagation algorithm (9分49秒)

• 课时16：Learning to predict the next word (12分33秒)

• 课时17：A brief diversion into cognitive science (4分26秒)

• 课时18：Another diversion_The softmax output function (7分20秒)

• 课时19：Neuro-probabilistic language models (7分52秒)

• 课时20：ways to deal with large number of possible outputs (12分16秒)

• 课时21：Why object recognition is difficult (4分40秒)

• 课时22：Ways to achieve viewpoint invariance (5分58秒)

• 课时23：Convolutional neural networks for hand-written digit recognition (16分1秒)

• 课时24：Convolutional neural networks for object recognition (17分44秒)

• 课时25：Overview of mini-batch gradient descent (8分22秒)

• 课时26：A bag of tricks for mini-batch descent (13分15秒)

• 课时27：The momentum method (8分42秒)

• 课时28：A separate, adaptive learning rate for each connection (5分44秒)

• 课时29：rmsprop_divide the gradient (11分38秒)

• 课时30：Modeling sequences_brief overview (17分23秒)

• 课时31：Training RNNs with backpropagation (6分23秒)

• 课时32：A toy example of training an RNN (6分14秒)

• 课时33：Why it is difficul to train an RNN (7分43秒)

• 课时34：Long term short term memory (9分15秒)

• 课时35：Modeling character strings with multiplicative connections (14分35秒)

• 课时36：Learning to predict the next character using HF (12分24秒)

• 课时37：Echo state networks (9分37秒)

• 课时38：Overview of ways to improve generalization (11分44秒)

• 课时39：Limiting size of the weights (6分22秒)

• 课时40：Using noise as a regularizer (7分31秒)

• 课时41：Introduction to the bayesian approach (10分49秒)

• 课时42：The bayesian interpretation of weight decay (10分52秒)

• 课时43：MacKays quick and dirty method of fixing weight costs (3分31秒)

• 课时44：Why it helps to combine models (13分10秒)

• 课时45：Mixtures of experts (13分15秒)

• 课时46：The idea of full bayesian learning (7分27秒)

• 课时47：Making full bayesian learning practical (6分44秒)

• 课时48：Dropout an efficient way to combine neural nets (8分35秒)

• 课时49：Hopfield Nets (13分1秒)

• 课时50：Dealing with spurious minima in hopfield nets (11分2秒)

• 课时51：Hopfields Nets with hidden units (9分39秒)

• 课时52：Using stochastic units to improve search (10分24秒)

• 课时53：How a boltzmann machine models data (11分44秒)

• 课时54：The boltzmann machine learning algorithm (12分15秒)

• 课时55：More efficient ways to get the statistics (14分48秒)

• 课时56：Restricted boltzmann machines (10分54秒)

• 课时57：An example of contrastive divergence learning (7分14秒)

• 课时58：RBMs for collaborative filtering (8分16秒)

• 课时59：The ups and downs of backpropagation (9分53秒)

• 课时60：Belief nets (12分35秒)

• 课时61：The wake-sleep algorithm (13分14秒)

• 课时62：Learning layers of features by stacking RBMs (17分34秒)

• 课时63：Discriminative fine-tuning for DBNs (9分40秒)

• 课时64：What happens during discriminative fine-tuning (8分39秒)

• 课时65：Modeling real-valued data with an RBM (9分56秒)

• 课时66：RBMs are infinite sigmoid belief nets (17分11秒)

• 课时67：From principal components analysis to autoencoders (7分57秒)

• 课时68：Deep Autoencoders (4分10秒)

• 课时69：Deep autoencoders for document retrieval and visualization (8分19秒)

• 课时70：Semantic hashing (8分50秒)

• 课时71：Learning binary codes for image retrieval (9分37秒)

• 课时72：Shallow autoencoders for pre-training (7分2秒)

• 课时73：Learning a joint model of images and captions (9分5秒)

• 课时74：Hierarchical coordinate frames (9分40秒)

• 课时75：Bayesian optimization of neural network hyperparameters (13分29秒)