The way we develop the next generation of learning software will have a tremendous impact on our life. Prof. Joan...
The way we develop the next generation of learning software will have a tremendous impact on our life. Prof. Joanna Bryson (@j2bryson) from University of Bath dissects this fascinating topic with Azeem Azhar (@azeem), founder of The Exponential view in this incredible podcast.
Intelligence, ethics, and digital transformation. Here’s the scoop:
Artificial Intelligence deep learning and Machine learning are the “Cognitive prosthetics that enable humans to offload information to the cloud“. Through voice recognition, photos and smart devices, we can save a big chunk of our life on the internet. AI and Machine learning give us the power to record but also to access all that data!
Bias in ML models is a very wide and hard problem to eradicate because it usually comes from the data on which the algorithm itself is trained, our data. And our data is no more than a mirror of our society, which is definitely biased. For example, if you train a language model to fill gaps in sentences and you feed it with the string “while the _ goes to the office, the _ cleans the house”, the model will link “man” to “office” and “woman” to “clean” and “house.
That’s not the fault of the engineers who create the model, it’s due to the billion times in human literature where this pattern occurs. In theory, if society would be fair, even ML models would be unbiased, but unfortunately, that’s not the case. And if we want an AI that pushes to a change instead of just mirroring human flaws, we have to figure out how to tweak its beliefs. In order to adjust any unfair bias, this issue is being heavily researched at the moment. If the results are successful, a weakness can be turned into a strength: AI and Machine Learning models can be used to detect, measure, and ultimately remove bias from human-centred processes!
Understanding why a certain action was taken or suggested by a machine learning algorithm is not only important for understanding how moral and ethical guidelines are respected, it’s also been imposed by regulations such as GDPR. For example, AI job-recommendation systems in the US are being criticised for suggesting high-profile roles more frequently to white people than to black people, on top of bias towards men over women seen from AI recruitment tools also used in the US.
Known as Intelligence Augmentation (IA), the idea of IA has been around since the 1950s. Today it is becoming an increasingly used term to describe machines that can mimic human functions like problem solving and learning. Prime examples of its application can be seen in AI applied to AgriTech or Drug discovery.
Just like with other machines (e.g. cars), we need to check the status of an AI model to avoid bad data isn’t being inadvertently introduced! Since these systems are continuing to learn, we need to ensure biases don’t affect our AI models.