 Our Speaker Bret Dunwoodie  Aly Bandali Thanking Our Speaker |
| Catherine Scheers introduced Bret Dunwoodie, an ENMAX market specialist, who described how artificial intelligence really works. He described three modes of machine learning: supervised, unsupervised, and reinforcement. In supervised learning, many known (labelled) quantities are provided to the AI to train it against expected outcomes. Once the AI has been appropriately trained, it can be used to predict outcomes based on variations of its inputs. For example, Enmax trains AI systems by providing actual gas demand data and past temperatures. AI systems can then predict future gas demand based on forecast temperatures. Supervised learning is used by facial recognition algorithms in which the AI is trained to isolate facial features such as eyes and ears, so that it will be able to identify the features on new faces. Supervised learning can be extended to other applications such as natural language processing and automatic picture captioning. Unsupervised learning follows a different approach in which the AI is not trained or tuned against known data. Instead, the AI simply creates groupings of similar data. Unsupervised learning is helpful in visualizing different groups and is often a good initial step in analyzing data. Customer segmentation in marketing campaigns is a commonly encountered example of unsupervised learning. Reinforcement learning is simply learning through doing and is similar to the way humans learn. The AI is rewarded if it successfully completes a task, while failure is penalized. In this style of machine learning, the AI determines actions to take based on its ability to maximize rewards. |
Reinforcement learning is commonly used by stock selection AI systems.
Mr. Dunwoodie described how many leading AI applications combine all three modes of machine learning to produce powerful predictive capabilities.
In response to a question from the floor, our speaker indicated that the leading AI risks we are likely to face in the coming years are ethical and legal. Ethical issues will arise when deciding what information features should be used to train AI systems. Inputting biased information into AI systems will result in their making biased predictions. Legal issues will arise when responsibility must be assigned to situations involving an AI system. As an example of this, our speaker raised the question of where the legal responsibility would lie when an automated motor vehicle is involved in an accident.
Aly Bendali thanked our speaker for a most thought provoking-presentation.