Deep Learning Flashcards

1
Q

what characterises supervised learning

A

labelled data
cost function
reperated epochs

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

what characterises un-supervised learning

A

non-labelled data
no cost function
finds representations of data

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

deep learning what is it

A

usually labelled data
uses multilayered neural networks, either convolutional or perceptrom based

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

What are the key components of a neural network

A

-Some cost function e.g. RMSE
-Forward propagation per layer per node consisting of weight x input + bias
-Following a neuron usually an activation function such as a ReLu or a Softmax
-Backpropagation

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

Why is backpropagation an efficient method of optimization

A

It stores inermediate gradients, allowing for more efficient computations of the partial derivatives at each feed forward stage, this allows for fast optimiazation of a neural network (at least relative to computing the derivatives seperately per variable

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

What is the activation function sigmoid and what are its pros and cons

A

-it is equal to 1/1+e^x and maps values range from any possible value [-infinity,infinity] to range of [0,1]
pros: differentiable and normalisees ranges
cons: hard to compute

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

What is the activation function (tanH) and what are its pros and cons

A

(e^x - e^-x) / ( e^x + e^-x). normalizes values to the range [-1,1]
-differentiable but hard to compute

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

What is the activation function ReLu and what are its pros and cons

A

f(x) = max (0,x). easy to compute but non differnetiable
-avoids saturation issues (i.e. when nodes have weights close to -1 or +1)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

What is the neural radiance field neural network

A

Form of stereo vision
-takes as input an image from different perspectives
-outputs a 3d map of the object
This is an example of a depth image

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

What is the relationship between amount of data and the performance of different ML / AI methods

A

Large neural network performs best with large data but worst with little data
- small neural network 2nd place in both cases
- traditional ML performs better than rest with little data but worst with most data

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

What is the general goal of a convolutional neural network

A

Learning higher-level features from image

Neural networks do this hierarchically i.e. they first learn low level -> mid level -> high level features then train a classifier

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

what is a general structure of a general CNN up to the feature maps

A

Data
Convolution
Activation
Pooling
Feature maps

CNN = multi-layer neural network with:

  • Local connectivity
    –neurone in a layer are only connected to small region of lair before it
  • They share weight parameters across spatial positions
    – Learning shift-invariant filter kernels
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

what parameters do convolutional kernels have

A
  • Kernel Size
  • Stride
  • Padding
  • Dilation:
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

what are non-linear activation functions

A

Leaky ReLu
Parametric ReLu
Exponential Linear Units

Good for when we want to find non-linear relationships

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

what is a pooling layer

A

taking the average or max of a section of kernel to reduce kernel size for better efficiency

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

What is the random wiring strategy for neurel net design

A

Randomly wiring nodes between multiple neurons and testing their efficacy, some examples are Minsky

17
Q

What is a non-random algorithmic approach to neural net design

A

Evolutionary algorithms:
-often out-perform human designed architectures
-keeps good parts of network and discards poor ones with some mutation probability

18
Q

given a set of neural network architectures, we want to find best one. what can we reduce this problem to

A

search problem

19
Q

what are 4 types of regularisation

A

-L2
-L1
-Dropout
-Cross Entropy

20
Q

what is the general form of gradient descent

A

w^(t+1) = w^(t) - lr ( dx ( Loss Function (w^(t))))

21
Q

name 3 optimisation algorithms

A

Adam
RMSProp
AdaGrad

22
Q

name 3 optimisation algorithms

A

Adam
RMSProp
AdaGrad

23
Q

What are some different deep learning algorithms for computer vision

A

CNN:
-classification algorithm, for detection, segmentation, learn hierarchical representations of visual data

R-CNN: (Regional CNN)
Object Detection algorithm combining CNNs. For each region it applies a CNN to localize objects

Fast R-CNN:
-same as R-CNN and runs faster by sharing computation over different regions. this is done by pooling layers across regions

24
Q

What is a hidden layer, convolutional layer and pooling layer

A

Hidden layer:
-region of forward propagation between inputs and outputs with weights and biases
-find reresentations and relationships in a dataset
-in CNN it has convolutional layers, pooling and fully connected

Covolutional Layer:
- convolutional layers are filters with learnable parameters which capture patterns in an image

Pooling Layter:
- Used to downsample spatial dimensions of feature maps produced by Convolutional Layers
-This process effectively retains the most salient features while reducing the spatial size of the feature maps.

25
What is data augmentation
Adding diversity to dataset to improve real world performance. it may include: -scaling, resizing -translations -rotations -gaussian noise ]0colour jittering
26
When is ReLu better than some other activation funcs
When network is very deep: - more efficient - less likely to cause saturation (-1 or +1 only nodes) causing faster convergence - less likely for vanishing gradient problem
27
What is the YOLO algrotithm (simple)
You Only Look Once: -Object detection for classifying multiple objects in an image in real-time -Rather than using sliding window it treats it as a regression problem and attempts to find likely bounding boxes and probabilities using a single nerual net
28
Describe the YOLO algorithm (detail)
-Input image is divided into groups of cells responsible for predicting class probabilities for objects in the box -Pre-definied anchor boxes with different shapes + sizes are used to predict dimensions of bounding boxes -Put image thorugh CNN once which contains cov layer, feature extraction layer, then fully connected layers for regression and classification - Outputs predictions of each bounding box containing an object and co-ordinates of the box and width -Non-Maxima supression to delete overlapping boxes with lower probability -Final output is a set of bounding boxes with estimated probabilities
29
Give some examples of supervised and non-supervised algorithms
supervised: - CNNs -SVM unsupervised: -GANs -Autoencoders -Kmeans