Machine Learning and Image Recognition

Question 1

What is machine learning?

Answer 1

The science of programming computers such that they can learn from data without being explicitly programmed to perform the task.

Question 2

What problems is ML good for?

Answer 2

• Where existing solutions require many rules or hand-tuning of parameters
• Complex problems without an existing solution
• Noisy environments
• Insights from large volumes of data

Question 3

Types of machine learning

Answer 3

• Supervised: The system is trained with data and a label/desired output. The goal is to map the data to the desired output.
• Unsupervised: The system is trained only with the data. The goal is to create a representation of the data.

Question 4

How do we build a simple linear ML model?

Answer 4

• Straight line model y = mx + c
• Evaluate the hypothesis at each value of x and find the difference between the hypothesis and the data points
• Find the mean squared error (loss function)
• Find the local minimum on a gradient descent graph
• If the loss doesn’t change then the function doesn’t affect the outcome

Question 5

What is bioimage informatics?

Answer 5

• A subfield of bioinformatics and computational biology
• Using computational techniques to analyse:
  1. High throughput imaging data
  2. Microscopy data for patient tissue samples, cell populations with drug treatments, molecular interactions etc.
• Typical computational tools used are computer vision, machine learning, clustering etc.

Question 6

Cryoelectron microscopy

Answer 6

• Aims to identify candidate molecules in CryoEM data
• Very noisy images
• Each image is the projection of a 3D object
• Images can be used to reconstruct the 3D structure of a protein and possibly the different states

Question 7

High content screening

Answer 7

• Using high throughput microscopy to image cells under different conditions, e.g. genetic/chemical perturbations
• Use informatics methods to extract information from the images and compare between perturbations

Question 8

Example of bioimage informatics in drug discovery

Answer 8

1. Take and label images of cells, image the fingerprints
2. Using an in vitro assay of the same drugs, work out how to map the image fingerprint to the protein activity
3. Work out what the change in protein activity might be
4. More than 200 targets informed, leading to boosted hit rate and diversity for drug discovery projects

Question 9

Digital pathology

Answer 9

Automated diagnostics based on image data:
- Data from digital slides of patient samples and the associated medical metadata
- Disease diagnosis
- Prediction of therapeutic interventions

Question 10

Multi-dimensional image data

Answer 10

Biological image datasets often use multiple markers
- Typically fluorescent reporters for a molecule of interest
- An image may contain multiple channels, each reporting on a different marker
- Each channel represents a marker for the same set of cells

Question 11

Common machine-learning tasks

Answer 11

• Detection of an object of interest
• Classification: assignment of a label (e.g. cell) to an image
• Segmentation: assignment of a per-pixel label to an image
• Feature extraction: image restoration and enhancement

Question 12

Activation functions

Answer 12

• Maps the input to a specific range of values
• e.g. you may wish to only allow positive activations
• Transforms the input to an output that scales monotonically with the input
• These activation function adds non-linearity to the overall function of the network
• Need to be differentiable

Question 13

Convolutional Neural Net

Answer 13

• Convolution: matrix multiplication of image patch with a kernel followed by addition of bias
• Activation function: maps the output of the conv layer to a new range
• Pool: downsamples the output by pooling pixels
• Conv, ReLU, and Pool layers are typically stacked together to form a deep neural network

Question 14

Pooling operations

Answer 14

• Pooling operations reduce the spatial scale
• Image size is halved after every pool
• The next layer of convolutions effectively has a larger receptive field

Question 15

Define ‘loss’ in ML

Answer 15

• The measure of error in the task
• Loss function is the cross-entropy function

Question 16

What is segmentation?

Answer 16

Segmentation is identifying whether a given pixel fits into the cluster/classifier.

• We can set a threshold and everything with a valuer greater will be assigned one state and everything with a value below will be assigned the opposite/different state
• Can take continuous data and make it binary

Question 17

Otsu’s method

Answer 17

The best threshold-finding algorithm. Seeks to minimise intra-class intensity variance.

Question 18

How do convolutional neural nets work?

Answer 18

• Build perceptrons which look at the window of the image and map the output onto an output matrix
• The output matrix is the magnitude of the gradient in the direction that the selection window is moving. It finds the edges of the image.
• The kernel refers to the collection of weights that are being used in a calculation to generate an output matrix.
• Kernels yield different features and can show the magnitude of the gradient in the X or Y direction
• The absolute value of both of the output matrices together gives you the edges of the original image.
• Kernels can be designed to pick out any feature we want.