From the course: Artificial Intelligence Foundations: Machine Learning

Combating bias

From the course: Artificial Intelligence Foundations: Machine Learning

Combating bias

“

- [Instructor] What if we only trained our flower image classification model on images of roses and tulips? What would the model learn from that data? What would we be teaching it? What do you think would happen if we asked the model to identify a picture of an orchid? These are all great questions for illustrating and discussing bias in machine learning, and the specific ways to combat it. To answer those questions, we would be teaching the model that there are only two flowers in the world, roses and tulips. By excluding pictures of other types of flowers, the model would learn that flowers like dandelions, sunflowers, and daisies either don't exist or unimportant. The model would be biased toward roses and tulips, meaning it will always predict a flower to be a rose or a tulip. While this is a simple example, bias that determines whether or not someone will get that small business loan or parole makes the stakes much higher. Bias is a huge impediment to realizing the promise of this life-changing technology. Bias surfaces when predictions made by a model are less favorable to an individual or group, when there's no relevant difference between the group that justifies the prediction. Bias surfaces throughout the machine learning lifecycle and can be mitigated if you know what to look for and the questions to ask. Bias can appear in your data, selected algorithm, or model. Bias shows up in your data if your dataset is imbalanced or doesn't accurately represent the environment the model will be deployed in. In our case, training the model on only two flower species when there are many more in the world causes bias. Outliers and abnormalities in your data can also skew the model and should be handled before training. For example, an image of a flower that is blurry or distorted could confuse the learning algorithm, producing a model that doesn't perform well. When dealing with data, it's important to pair engineers with subject matter experts or SMEs, so problems can be quickly found early on. Bias can also show up in your selected algorithm. We experimented with various algorithms to select the one that performed the best. We saw that XGBoost out-of-the-box outperformed other regression algorithms. You'll need to understand the capabilities of your selected learning algorithm. Is it configurable or not? And how well it fits the problem you're trying to solve. Bias can also show up in your model. Once models are in production, they have a tendency to drift. Drift indicates that the relationship between the target variable and the other variables changes over time. Due to this drift, the model becomes unstable and the predictive power continues to degrade over time. To handle drift in the data, weighting features can be a good option. This is where you call out important features so the model doesn't vary on those. While bias is a cause for concern, there are proven ways to mitigate it. Now that you are aware of how to handle bias, let's look at optimizing a machine learning pipeline.

Contents