Hosted by IRMA (Information Risk Management and Assurance) Specialist Group of the BCS (British Computer Society).
12/03/19 @ The Chartered Institute for IT, London
Speaker: Mike Small, Senior Analyst at KuppingerCole
Topics of talk:
What AI is
Limitations of AI today
AI in practice today
Areas where AI could be successfully applied
Ethical issues of AI and potential approaches
What Is AI:
Is the attempt to simulate human abilities.
For this we need knowledge. Historically understood via epistemology, mathematics, and science.
Involves robotics and decision making.
Mike will focus moreso on decision making.
Is not new. History stretching back to the 1940s with various attempts to imitate human abilities e.g. chess, speech, etc.
What has changed is the invention of the Cloud. It brings the tech from the 1980s into the modern era by storing more data and making it easily accessible.
Works using rules. These same rules can be misapplied or misinterpreted by people: AI systems can help. AI often seen as less fallible and more trusted.
Anecdote: Salford University has an autonomous car which students regularly, intentionally step in front of in order to make brake. Would not do so if a fellow student were driving.
E.g. social security advice at DHSS – standardising advice by offering consistent suggestions.
E.g. container dock layout planner in Hong Kong – economising on space and efficiently organising the mooring of ships.
Is the business of data analytics. It builds up from statistical systems using algorithms into neural networks and results in features such as computer vision and natural language processing.
Machine Learning: “Iterative convergence on best match to training data”
One trains the system to analyse data with a training data set. Subsequently, just add more data to generate output.
Perceptron as example of an algorithm: a form of maths used to turn data sets/input into output, using weightings and thresholds.
AI is more general than Machine Learning. However, these terms are often interchangeable. Though, it used to be the case that rules-based sytems were called AI, and then systems based on LISP. Now it is Machine Learning.
Mike’s joke/point: systems are called solutions when finished and work; they’re called AI or Machine Learning when unsure – needs a bit of fairy dust to make it sound more interesting/sellable.
Deep Learning aka convolution – layers of processing
Limitations of AI:
Lack of common sense
Neural networks cannot explain themselves
Insufficient training data + bias
Machine Learning can be duped
E.g. exploit facial recognition using glasses (Adversarial Misclassification).
AI in Practice:
Identify anomalies; event analysis and support less skilled analysts.
AI looks not just at data set but also books, websites, social media, etc. to output suggestion to analyst that takes into consideration recent developments and news.
Aside: AI can be used to augment more junior and less skilled roles – requires less of them, gives suggestions and makes them spot aspects/issues that they previously failed to acknowledge.
Trained to spot defects in HD TVs by scanning the surface for pixel damage.
Regulatory Compliance (Finance)
Helps deal with sifting through regulation, particularly new regulation, and relates obligations to company.
Identifies any adjustments needed, especially in light of regulatory change.
Also used to identify insider trading since communication between traders are monitored.
E.g. identify buzzwords and code that can be linked to trader behaviour e.g. “it going to rain in Chicago” and then trader sells shares in Chicago.
Suggestions to organise data differently and thereby improve performance.
Future of AI:
Narrow AI (emerging)
Single tasks, single domain: very accurate, very fast.
Broad AI (disruptive)
Muti task, multi domains: explainable; transfers skills and tasks across domains.
General AI (revolutionary)
Cross domain learning and reasoning; autonomy.
The challenge is public acceptance.
GM crops/Golden Rice as case study of innovative technology curtailed by protests.
Computing may flit from benign to malign in public perception, thus evading risks but also losing out on potential gains and benefits.
Surveillance Capitalism: the focus has changed from money to data. This could potentially face a backlash.
Involves looking at the Context->Consequences->Justification
Popular case study: Trolley problem
Different cultures have different answers.
Can’t sue a car, who do you sue? The designer, developer, builder, owner?
In court, would need to defend the decision made by the car. Would this be clear?
Need to consider:
Benefit v harmfulness
Inclusion v exclusion
Bias v unfairness
Bias towards the culture of the data set
Not all bias is bad, unfair or unintended bias is.
Good v bad behaviour
How do you convince an AI that it has done something wrong?