1
Q
What is learning by gradient descent? Explain the general idea behind it, and the role the error E has in it.
A
- algorithm which aims to minimise the error (E) of the NN by adjusting the model’s parameters
- involves computing gradients of E with respect to these parameters
- the parameters are adjusted in the opposite direction of the gradient to minimise E
- iteratively reduces E, in pursuit of a global minimum
2
Q
Briefly describe what the backpropagation algorithm is, and in which way it relates to gradient descent.
A
- algorithm used to train artificial NN by minimising the error between the predicted output and the target values
- Two main phases: Forward Pass, Backward Pass (backpropagation)
- FP: Input data is fed into the NN, layer-by-layer computations yield the predicted output
- Backpropagation: works backwards through the layers, using gradient descent to minimise the error of the NN by adjusting the model’s parameters at each layer.
- involves calculus to calculate the partial derivatives of the loss function with respect to each parameter (weights and bias)
3
Q
What are the common problems of gradient descent, that may limit its effectiveness?
A
- local minima
- slow convergence
- sensitivity to learning rate
- dependence on initial weight selection
4
Q
Explain the role of activation functions in NN
A
They play a crucial role by introducing non-linearities to the model, which are essential for enabling NN to learn complex patterns in the data
5
Q
What is the purpose of the cost function in a NN
A
Also known as the loss function, it quantifies the inconsistency between predicted values and the corresponding correct values
6
Q
explain the role of bias terms in a NN
A
- bias terms add a level of flexibility and adaptability to the model.
- they “shift” the activation function, providing every neuron with a trainable constant value, in addition to the inputs
7
Q
What is a perceptron
A
an artificial neuron which takes in many input signals and produces a single binary output signal (0 or 1)
8
Q
Explain the differences between Batch Gradient Descent and Stochastic Gradient Descent
A
in BGD, the model parameters are updated in one go, based on the average gradient of the entire training dataset. In SGD, updates occur for each training example or mini-batch.
9
Q
Which gradient descent is preferred for large datasets and why.
A
Stochastic GD is preferred over Batch GD.
- Although BGD usually converges to a more accurate minimum, it is computationally expensive (extremely)
- SGD converges faster and requires less memory. However, updates can be noisy, and it may converge to a local minimum rather than the global minimum
10
Q
Define generalisation
A
The ability of a trained model to perform well on unseen data
11
Q
How can you measure the generalisation ability of a MLP
A
- cross validation
- hold-out strategy (train/test sets)
- consider choice of evaluation measure
12
Q
How can you decide on an optimal number of hidden units?
A
- apply to domain knowledge to estimate a range
- test the model on the range to fine tune the selection
- this may be unfeasible on complex models
13
Q
Explain the difference between two common activation functions of your choice
A
Sigmoid vs TanH
1. Output Range:
- Sigmoid: (0,1): used for binary classification
- tanh: (-1, 1): suitable for zero-centred data
2. Symmetry:
- Sigmoid is asymmetric, biased towards positive values
- tanh is symmetric around the origin (0, 0)
14
Q
What are the problems with squared error as the loss function, give two alternatives
A
There are tricky problems with squared error:
- if the desired output is 1 and the actual output is very close to 0, there is almost no gradient
- alternatives: softmax, relative entropy
15
Q
Define what a Deep Neural Network is
A
- consists of multiple layers that transform the input in a hierarchical fashion
- they typically are feed-forward NN with multiple hidden layers, allowing modelling of complex non-linear relationships
16
Q
Formal definition of overfitting in practice
A
during learning the error on the training examples decreases all along, but the error in generalisation reaches a minimum and then starts growing again.
17
Q
Training data contains information about the regularities in the mapping from input to output.
But it also contains noise, explain how.
A
- the target values may be unreliable
- there will be accidental regularities just because of the particular training cases that were chosen
18
Q
When we fit a model, it cannot tell which regularities are real and which are caused by sampling error. Which regularity does it fit, what is the worst case scenario?
A
- Both
- worst case: If the model is very flexible it can model the sampling error really well
19
Q
What does a model having the “right capacity” entail
A
- enough to model the true regularities
- not enough to also model the spurious regularities
20
Q
How to prevent overfitting in NN
A
- limiting number of weights
- weight decay
- early stopping
- combining diverse networks
21
Q
Standard ways to limit the capacity of a neural net
A
- Limit the number of hidden units.
- Limit the size of the weights.
- Stop the learning before it has time to overfit.
22
Q
How to limit the size of the model by using fewer hidden units in practice
A
trial and error
23
Q
What is weight-decay
A
- method for limiting the size of a model
- involves adding an extra term to the cost function that penalises the squared weights, i.e. keeps weights small unless they have large error derivatives
24
Q
What does weight decay prevent, and what does it improve and how?
A
- It prevents the network from using weights that it does not need.
- It tends to keep the network in the linear region, where its capacity is lower.
- This helps to stop it from fitting the sampling error. It makes a smoother model in which the output changes more slowly as the input changes.
- It can often improve generalisation a lot.
25
What is the idea behind early stopping for preventing overfitting
- expensive to train a big model with lots of data
- cheaper to stop adjusting weights once generalisation starts getting worse
- the capacity is limited because the weights have not had time to grow big
26
what hold out strategy is recommended for model selection of a MLP
Strategies which include a **validation set**
27
What is gradient descent
- optimisation technique used to find optimal parameters of a model by iteratively updating them in the direction of the steepest descent of the loss function.
- aims to minimise the error of the model
28
What is the recommended method for the three-way hold-out strategy when training a NN
- **training data:** used for learning the parameters of the model
- **validation data:** used for deciding what type of model and what amount of regularisation works best **(fine tuning)**
- **test data:** used to get a final, unbiased estimate of how well the network works
29
Why NN ensembles?
The average error of a group of predictors is always smaller than the average error of the single predictors (unless the predictors are identical)
30
Briefly explain the steps of k-Fold Cross Validation
- Divide the data into k disjoint subsets - “folds”
- For each of k experiments, use k-1 folds for training and the selected one fold for testing.
- Repeat for all k folds, average the accuracy/error rates.
31
How to achieve Network Ensembling with just one training
Use **Dropout** method
- during training, at each step knock out some randomly chosen connections
- when predicting, **use all connections**. You will need to introduce a **normalising constant** for this to work
- **equivalent to having a very large ensemble of networks**
32
Precautions for applying dropout in practice
- it doesn't always work: **preconditions required**
- Must begin with an oversized net capacity to avoid underfitting
33
What is online learning
Weight updates occur for **each example** during Gradient Descent
34
Discuss the value of the gradient at different error angles, what does gradient descent do at these values and are we satisfied with this?
- the gradient is large where the error is steep, small where the error is flat
- Sometimes we would like to run where it's flat and slow down when it gets too steep. GD does **precisely the contrary**
35
Briefly discuss some of the fixes for the issues of Gradient Descent
Use an **adaptive learning rate**:
- increase the rate slowly if it's not diverging
- decrease the rate quickly when it starts diverging
Use **Momentum**: instead of using the gradient to change the position of the weight, change the velocity of the change
Use **fixed step**: GD decides where to go, but always at same pace
**Normalise the gradient** based on some combination of previous gradients
36
Explain what RL is, and what we want to learn from it
Learning from interaction with an environment to achieve some long-term goal that is related to the state of the environment
- we want to learn how to act to accomplish goals
- given an environment that contains rewards, we want to learn a policy for acting
37
Define a simple RL Setup And Goal
Setup: We have an **agent** which is interacting with an environment which it can affect through **actions**. The agent may be able to sense the environment partially or fully.
Goal: the agent tries to maximise the long term reward conveyed using a **reward signal**
38
Explain the differences between Supervised and Reinforcement Learning
- In SL, there's an external "supervisor", which has knowledge of the environment and who **shares it with the agent** to complete the task
- Both strategies use **mappings between inputs and outputs**, but in RL there is a reward function which acts as a **feedback** to the agent
- Supervised learning relies on **labelled training data**
39
Explain the differences between Unsupervised and Reinforcement Learning
- In UL, there is **no feedback from the environment**
- In UL, task is to find the **underlying patterns** rather than the mapping from input to output
40
Characteristics of RL
- no supervisor, only a reward signal
- feedback is delayed, not instantaneous
- time really matters
- Agent's actions have immediate consequences
41
Why is Deep Learning hard
- when networks get deep the gradient **vanishes**
- when a network is untrained, the deeper down a hidden unit is, the more **subtle** its effect on the outputs
- this means it doesn't do much to the error if you change it
42
Briefly discuss the two main categories for Deep Learning solutions
Pre-training:
- stack deep networks layer-by-layer
- make sure each layer represents the previous layer meaningfully before adding another layer
Use Artificial Targets: the real problem is that inner layers don't get gradient, so every now and then use **hard targets**:
- generate some random targets for the layer
- evaluate them all
- use the best one
43
Advantages and disadvantages of pre-training by auto-association
Adv: you can use unlabelled data
Disadvantage (potentially disastrous):
- you aren't considering at all the **property you
want to predict**
- **you compress regardless of the property**. If it's
lossy, the loss can be in the wrong place..
44
Advantages and disadvantages of pre-training without auto-association
Advantage:
- you **compress based on the property you are trying to predict**. If it's lossy, the loss is probably in the right place
Disadvantage:
- can't use unlabelled training data
- shorter training
45
What is clustering
- unsupervised learning: grouping un-labelled data
- find underlying patterns in the data
- large choice of distance functions
- partitioning and hierarchical methods
46
Possible Clustering implementations for connectionist models
k-means can be achieved by using backpropagation in a **non-linear self-associating network**:
- there is one hidden layer with each node representing a cluster centre
- the hidden layer is **hardmax**, therefore only one of the neurons will be activated from the input
- the neurons weight will be adjusted when it is activated **(recomputes cluster centre)**
47
What are the strengths and weaknesses of Clustering compared to Principal Component Analysis (in connectionist models)?
PCA: linear self-associating networks
Clustering: non-linear (hardmax) self-associating networks
- PCA builds **global features** (strength) while clustering builds **local features** (weakness)
- PCA only considers **linear** combinations of the inputs (weakness)
- clustering builds much **stronger features** (strength)
48
State what the difference is between Feedforward and Feedback networks
**Information Flow:**
- In FFN, information flows in one direction
- FBNs have recurrent connections, allows them to maintain and propagate information over time
- Consequently, FBNs can model sequences and time-dependent data
49
Which network architecture (FFN, FBN) do you think are easier to deal with? Justify your choice.
- FFNs are easier to train and are more stable because there are no feedback loops
- FBNs can be more challenging to train due to vanishing gradients
50
Describe Hopfield Networks and Boltzmann Machines
- both are types of Recurrent NN (RNN)
- HNs consists of **binary threshold units with symmetric connections**
- BMs use **binary stochastic units** and incorporate a **probabilistic aspect in the update rule**
51
Discuss the learning process in Hopfield Networks and Boltzmann machines
Hopfield Networks:
- learning involves adjusting the weights to store certain memories or patterns, essentially capturing **second-order interactions**
Boltzmann machines
- learn to generate configurations according to a probability distribution
- involves adjusting the weights based on the **correlation differences in the training and generated data**
52
Do Hopfield Networks and Boltzmann machines tackle similar problems?
- HNs are deterministic, while BMs are probabilistic
- Consequently, BMs can represent **higher-order interactions**
53
Discuss similarities and differences between MLPs and Boltzmann Machines
- both can have hidden layers
- feedforward vs recurrent
- role of hidden units are somewhat similar in both models, aiming to learn complex patterns/structures
- the manner in which HU operate are different: deterministic in MLP, probabilistic in BM
