Unit 1: Optimisation and PyTorch

Question 1

Deep Learning

Answer 1

A branch of Machine Learning concerned with learning from data using artificial neural networks (ANNs).

Question 2

Neural networks

Answer 2

Mathematical functions that map an array of inputs to one or more outputs.

Question 3

Supervised Learning

Answer 3

Concerned with finding a function that can accurately map a set of input vectors to their corresponding outputs.

Input-output pairs are given and the task is to find a function which correctly maps the inputs to outputs.

Question 4

Clustering

Answer 4

In clustering a measure of similarity, such as distance or density, is used to find clusters of similar items or data points in a given dataset.

Question 5

Dimensionality Reduction

Answer 5

The goal of dimensionality reduction is to find a lower-dimensional representation of data while preserving maximum information about the original dataset.

Question 6

Reinforcement Learning

Answer 6

Concerned with optimal behaviour of an intelligent agent in an environment seeking to maximise a long-term payoff.

The agent in an RL setting does not have access to labeled data.

Instead, the environment rewards or penalises the agents behaviour as it acts and interactis with the environment. The goal of the agent is to find a behaviour policy that maximises its long-term reward.

Question 7

Why deep learning?

Answer 7

Feature engineering is a central problem in machine learning.

Deep learning solves the feature engineering problem by automatically learning representations through the deep layered structure of neural networks.

Question 8

3 Factors leading to the success of deep learning

Answer 8

1. Big data: the representation learning ability of deep learning requires large amounts of data to build accurate generalisable models.
2. Hardware: availability of cheaper computational resources, faster CPUs and general purpose GPUs, parallelisation, and faster network connectivity, has enabled us to train larger and deeper neural network models capable of learning complex learning tasks such as machine translation, object recognition and speech synthesis.
3. Software: advances in techniques and algorithms, and availability of specialised tools and software libraries.

Question 9

Why is image classification hard? (4)

Answer 9

• variations in viewpoint and lighting
• intra-category variation
• shape deformation
• distraction from other objects and scene background