Artificial Intelligence Flashcards
(21 cards)
What is AI?
Artificial intelligence uses computers and machines to mimic the problem-solving and decision-making capabilities of the human mind [IBM]
What is machine learning?
A branch of AI and computer science which focuses on the use of data and algorithms to imitate the way that humans learn, gradually improving its accuracy [IBM]
Why is AI suddenly so available?
- Increased computer power is such that training and using AI models has become commercially viable, e.g. cloud computing, GPU acceleration
- A number of models have been developed that show promise
- Commercial products have become available
- Public are gradually accepting AI as a positive development
What are some applications of AI?
- Automatic image segmentation
- Automatic optimisation of radiotherapy planning parameters
- Identification of images with abnormal pathology
- Highlighting areas of interest in images
- Prioritising images for further human review
- MRI “acceleration”
- Patient scheduling and bed management
- Early-warning about changes in equipment performance
- Generating synthetic patient data
What are the general classes of AI models?
- Image or data class - which group something belongs to
- Regression - predicting outputs for a given input
- Object detection - whether something is present, its structure or dimensions
- Generative AI - producing document summaries, compensating missing data
What is the Turing Test?
Measures the maturity of AI systems - if the computer’s output is indistinguishable from that of a human, then it passes the test
What is a Reverse Turing Test?
Distinguish between AI and humans
How does an AI model work?
Input information, algorithm performs processing and gives an output, which then learns through an (optional) feedback loop. Without the feedback loop, the model should always produce the same outputs for the sasme inputs
How do you train an AI model?
Input the training data, the model is stimulated, gives initial outputs and learns from it, updating the model. Model hyperparameters are saved. Cloud resources may be required. Multiple learning cycles may be required, using the same data.
What an epoch?
A complete cycle when training an AI model
What is an ROC curve? And what do the kinks represent?
Receiver Operating Characteristic curve, true positive vs false positive
Kinks in the curve represent decision thresholds
What does the area under the curve of a graph of false vs true positive rates represent?
Measure of the overall performance
When are clinical scientists involved in AI?
- Developing new models (rare)
- Updating or repurposing models
- Identifying and specifying applications for AI
- Training models
- Implementing and testing models
- Developing QA programme
Clinical/scientific/corporate user of AI
What is bias in AI?
An inclination to treat one group of people more favourably than another, especially in a way which is considered unfair
How do you overcome bias in AI?
- Need access to sufficient breadth of training data so that model can accommodate all anticipated patients
- Test pre-trained model using data representative of local populations
- Include abiluity to re-train or refine training of model using local data
4, Local datasets need to incldue sufficiently detailed clinical inofrmation to be meaningfully used in training
What are issues involved with the use of AI?
Bias - can one model accommodate ALL patients
Humans are adept at recognising situations not previously encountered and adapting accordingly
May need to use different models for different groups of people (e.g. children, infants)
Care required around use of GDPR special category data
Involvning the right peoplr
collating the correct size and comprehensive data
Who has access to the data?
Who is responsible? e.g. for the decision making
De-skill humans
What GDPR considerations are there around AI?
- Lawful, fair and transparent - how do we explain AI decision making?
- GDPR states that the data subject shall have the right not to be subject to a decision making based soleyl on automated processing (unless contractualor consent)
What IRR17 considerations are around AI?
Equipment (including software) delivering or controlling exposures
Operator - “person”
What is LLM?
Large Language Model
Deep learning models trained on vast amounts of unlabelled text data, usually “encoder-decoder”, in which the model refines itself (self-supervised training)
Generative Pre-trained Transformers (GPTs)
What is an “encoder-decoder” model?
Model which predicts a sequential output from a sequential input
What is a GPT?
Generative Pre-trained Transformers
