9444_1

1b有问题

Supervised, Reinforcement,Unsupervised Learning 区别和概念

Supervised Learning: The system is presented with training items consisting of an input and a target output. The aim is to predict the output, given the input (for the training set as well as an unseen test set).

Reinforcement Learning: The system chooses actions in a simulated environment, observing its state and receiving rewards along the way. The aim is to maximize the cumulative reward.

Unsupervised Learning: The system is presented with training items consisting of only an input (no target value). The aim is to extract hidden features or other structure from these data.

meant by Overfitting in neural networks

Overfitting is where the training set error continues to reduce, but the test set error stalls or increases.

how to avoid Overfitting

• limiting the number of neurons or connections in the network
• early stopping, with a validation set
• dropout
• weight decay (this can avoid overfitting by limiting the size of the weights)

• how Dropout is used for neural networks(train&test)

  During each minibatch of training, a fixed percentage (usually, one half) of nodes are chosen to be inactive. In the testing phase, all nodes are active but the activation of each node is multiplied by the same percentage that was used in training.

  Squared Error, Cross Entropy, Softmax, Weight Decay

  Assume ziz_izi​ is the actual output, tit_iti​ is the target output and wjw_jwj​ are the weights.

  Squared Error: E=12∑i(zi−ti)2E = \frac{1}{2}\sum_i (z_i - t_i)^2E=21​∑i​(zi​−ti​)2

  Cross Entropy: E=∑i(−tilog⁡zi−(1−ti)log⁡(1−zi))E = \sum_i (-t_i\log z_i - (1-t_i)\log(1-z_i))E=∑i​(−ti​logzi​−(1−ti​)log(1−zi​))

  Softmax: E=−(zi−log⁡∑jexp⁡(zj))E = -(z_i - \log\sum_j \exp(z_j))E=−(zi​−log∑j​exp(zj​)), where $i$ is the correct class.

  Weight Decay: E=12∑jwj2E = \frac{1}{2}\sum_j w_j^2E=21​∑j​wj2​

  difference between Maximum Likelihood estimation and Bayesian Inference(supervised)

  In Maximum Likelihood estimation, the hypothesis $h\in H$ is chosen which maximises the conditional probability $P(D\mid h)$ of the observed data $D$, conditioned on $h$.

  In Bayesian Inference, the hypothesis $h\in H$ is chosen which maximizes $P(D\mid h)P(h)$, where $P(h)$ is the prior probability of $h$.

  concept of Momentum( as an enhancement for Gradient Descent)

  A running average of the differentials for each weight is maintained and used to update the weights as follows:

  δw = αδw −ηdEdw\delta w\ =\ \alpha\delta w\ -\eta\frac{dE}{dw}δw = αδw −ηdwdE​

  $$w=w+\delta w$$

  The constant $\alpha$ with $0\le \alpha <1$ is called the momentum.