Artificial Intelligence (AI) has become embedded in nearly every aspect of our lives. The purchases we make, the people we virtually connect with, even the mechanisms to unlock our phones (if the phone was made in the last four years) are all influenced by AI. That said, should there be a limit to what parts of our lives AI touches? Moreover, how can we be sure AI systems will behave in the way we would expect?
How did we get here?
While AI has broadly existed for the last 70 years, machine learning, and more specifically deep learning, has taken off in the last 10 years due to access to larger datasets and faster computational resources. While the goal is not to provide an in-depth discussion of deep learning, there are a couple important properties to note:
- Deep learning can be used to create models that are extremely accurate on a large number of diverse tasks, such as image recognition, natural language processing, or network analysis. This has led many to adopt deep learning models in settings where they may not always be appropriate.
- Given deep learning’s complexity, deep learning models are notoriously difficult to understand, i.e., you usually cannot confidently argue why a prediction was made. This becomes extremely debilitating when the application domain necessitates understanding, or impacts human lives.
- Even more problematic, deep learning models are often highly susceptible to adversarial attacks -- inputs designed to fool the model into believing the input is something that it is not. Many of these attacks are particularly nefarious as they only require a minor change to cripple the model (think only needing to change ~0.1% of the pixels in an image). This leads to an important insight in regards to deep learning models; often they are simply learning spurious correlation, rather than casual relationships, leading to highly unstable predictions.
These three situations lead to a significant problem in deep learning where complex models are applied to sensitive applications, due to their perceived success, with no ability to understand why a decision is made or if the decision making process is in line with human intuition. This practice raises many red flags in the fields of ethics, especially when a model is directly impacting the life of a human.
The era of AI fairness
As the misuse of deep learning has become a realization, AI fairness has emerged as a field to understand and critique the use of models in sensitive settings. Some aspects of AI fairness include establishing new metrics and datasets to better assess model performance, developing new models able to better handle biased data, and ultimately arguing for best practices when moving a model into production. All of this said, one of the most important AI fairness tenets is to use AI for problems that make sense. Let us consider facial recognition to solidify the significance of this last point.
Facial recognition has become extremely prominent in computer vision with the creation of more powerful deep learning models able to take advantage of images. In fact, it has become so popular, many have moved past simple recognition and instead considered directly predicting attributes and properties of individuals. Some of the properties considered have included an individual’s sexual orientation, as well as a person’s likelihood to commit a crime. Without mechanisms to properly determine how a decision is made, we often see that models are taking advantage of spurious, unethical, and sometimes blatantly wrong facial attributes to make decisions when post-hoc analysis is performed in response to a harmful prediction. That said, whether the model picks up particular facial attributes or not, it is important to take a step back and consider an important question: “Why would I believe facial features to be indicative of these properties?”
Repeating the harmful mistakes of the past
The belief that physical attributes are indicative of non-physical properties, such as being more likely to commit crime, alludes to a dark point in history where scientific racism worked to differentiate based on observable characteristics. In modern 21st century science, notions of inherent inferiority based on skin tone, eye shape, cranial size, etc., have largely been relegated to late 1800s pseudo-scientists. If this is the case though, why do deep learning applications creep up that seem to be based on similar hypotheses? One issue stems from the black-box nature of deep learning which has allowed the view of data objectivity, and the model simply extracts insights from that data, to absolve the model creators of harm. However, to make this argument fails to recognize the historical and systematic biases that dictate many of our data generation processes, such as issues of over-sentencing or redlining.
I believe deep learning does have the power to change the world, hopefully for the better. I advocate for recognizing that the supposed success of deep learning is not without fault, and these faults can produce significant societal harm if not properly vetted. Continuing to push for research that bridges social science, computer science, and political science in regards to AI is, in my opinion, the only way to safely and fairly integrate deep learning into society.
Comments