Flashcards
Sem título
Flashcards do conjunto (29)
- 1PCA meansPrincipal Component Analysis
- 2PCA is used in -----Reducing, Finding and Visualizing
High Dimensional Data. - 3PCA applications includes ----Face Recognition
Image Compression
Gene Expression Analysis. - 4Vectors having the same direction are called ---------Eigen Vectors
- 5Their co-efficient are called -------Eigen Values
- 6Steps to perform PCA :Eigen Decomposition of Convariance Matrix
Make a list of the vectors
Sort from high to low
Transform to a new sub space, using largest 2 eigen values. - 7PC1Largest absolute value /
Largest variance. - 8PC2d direction with respect to max. variance
orthogonal to PC1. - 9Mention 2 evaluation methods?Cross-Validation (k-folds),
Hold-Out. - 10Cross Validation uses entire training set whileHold-Out uses half of training set
- 11Cross-Validation is prone to less variations whileHold-Out is prone to more variations.
- 12Cross-Validation has High Computational cost whileHold-Out has low computational cost.
- 13RNN is a ------type of ANN that connections btw the units are in a directed cycle.
They re-use weights and sequence of vectors. - 14LSTM- Long Short Term Memory is a -----Deep learning that avoids the vanishing gradient problem.
It prevents BP errors from vanishing/exploding. - 15Mention 4 Types of Deep Architecture ?Feed-Forward e.g MLNN, CNN
Feed-Back e.g De-CNN
Bi-Directional e.g Boltzman machine
Recurrent e.g RNN - 16Data inside a sequence are -------non - i.d.didentically,
independently,
distributed. - 17Memory is a representation of the -----past.
- 18RNN can be converted into feed-forward network using ------"unfolding over time."
- 19BPTT is used to learn -----. BPTT is an extension of BP.RNN.
- 20DNN layers relate to each other thro multiplication of their derivatives which tends to vanish. How to Solve this ?Proper initialization of matrix.
Use ReLU instead of Tan h or Sigmoid fnc.
Use LSTM. - 21Cell state carries the ------essential information over time.
- 22RNN applications ?Machine Translation,
Visual Ques. & Answering. - 234 Types of Activation Functions ?Step fnc: produces binary results; 0 & 1
Sigmoid fnc: produces values btw 0 & 1
Tan h fnc: produces values btw -1 & 1
ReLU: is d fastest,depends on size of z & x. - 24Solutions for Over-fitting ?More training data.
Reduce the # of feature.
Regularization - kind of smoothing method. - 25Drop-Out ------ignores some of the data and focusing on the important ones.
- 26Ensemble (+3~5% better accuracy) ------combines 2/more different but related analytic models and synthesizing their result into a single score, for more better/accurate prediction of d analysis.
- 27What is k-means ?K-means is an iterative algorithm for classifying data with similar features into a cluster.
- 28What are d 3 steps of K-means algorithm ?1. Randomly assign k cluster centroid,
2. Cluster Assignment step,
3. Move Centroid step. - 29—
Informações
- Descrição
- FlashCards sobre Sem título, criado por Abolanle Sule em 07-11-2017.
- Criado por
- Abolanle Sule
- Visualizações
- 2
- Avaliação