One of the biggest issues to consider is that understanding and testing the behavior of a deep learning model is much more difficult than with most other code you write. With normal software development you can analyze the exact steps that the software is taking, and carefully study which of these steps match the desired behavior that you are trying to create. But with a neural network the behavior emerges from the model’s attempt to match the training data, rather than being exactly defined.

This can result in disaster! For instance, let’s say we really were rolling out a bear detection system that will be attached to video cameras around campsites in national parks, and will warn campers of incoming bears. If we used a model trained with the dataset we downloaded there would be all kinds of problems in practice, such as:

• Handling nighttime images, which may not appear in this dataset
• Ensuring results are returned fast enough to be useful in practice

A big part of the issue is that the kinds of photos that people are most likely to upload to the internet are the kinds of photos that do a good job of clearly and artistically displaying their subject matter—which isn’t the kind of input this system is going to be getting. So, we may need to do a lot of our own data collection and labelling to create a useful system.

This is just one example of the more general problem of out-of-domain data. That is to say, there may be data that our model sees in production which is very different to what it saw during training. There isn’t really a complete technical solution to this problem; instead, we have to be careful about our approach to rolling out the technology.

Out-of-domain data and domain shift are examples of a larger problem: that you can never fully understand the entire behaviour of your neural network. They have far too many parameters to be able to analytically understand all of their possible behaviors. This is the natural downside of their best feature—their flexibility, which enables them to solve complex problems where we may not even be able to fully specify our preferred solution approaches. The good news, however, is that there are ways to mitigate these risks using a carefully thought-out process. The details of this will vary depending on the details of the problem you are solving, but we will attempt to lay out here a high-level approach, summarized in <>, which we hope will provide useful guidance.

Where possible, the first step is to use an entirely manual process, with your deep learning model approach running in parallel but not being used directly to drive any actions. The humans involved in the manual process should look at the deep learning outputs and check whether they make sense. For instance, with our bear classifier a park ranger could have a screen displaying video feeds from all the cameras, with any possible bear sightings simply highlighted in red. The park ranger would still be expected to be just as alert as before the model was deployed; the model is simply helping to check for problems at this point.

The second step is to try to limit the scope of the model, and have it carefully supervised by people. For instance, do a small geographically and time-constrained trial of the model-driven approach. Rather than rolling our bear classifier out in every national park throughout the country, we could pick a single observation post, for a one-week period, and have a park ranger check each alert before it goes out.

One of the biggest challenges in rolling out a model is that your model may change the behaviour of the system it is a part of. For instance, consider a “predictive policing” algorithm that predicts more crime in certain neighborhoods, causing more police officers to be sent to those neighborhoods, which can result in more crimes being recorded in those neighborhoods, and so on. In the Royal Statistical Society paper , Kristian Lum and William Isaac observe that: “predictive policing is aptly named: it is predicting future policing, not future crime.”

Part of the issue in this case is that in the presence of bias (which we’ll discuss in depth in the next chapter), feedback loops can result in negative implications of that bias getting worse and worse. For instance, there are concerns that this is already happening in the US, where there is significant bias in arrest rates on racial grounds. According to the ACLU, “despite roughly equal usage rates, Blacks are 3.73 times more likely than whites to be arrested for marijuana.” The impact of this bias, along with the rollout of predictive policing algorithms in many parts of the US, led Bärí Williams to : “The same technology that’s the source of so much excitement in my career is being used in law enforcement in ways that could mean that in the coming years, my son, who is 7 now, is more likely to be profiled or arrested—or worse—for no reason other than his race and where we live.”

A helpful exercise prior to rolling out a significant machine learning system is to consider this question: “What would happen if it went really, really well?” In other words, what if the predictive power was extremely high, and its ability to influence behavior was extremely significant? In that case, who would be most impacted? What would the most extreme results potentially look like? How would you know what was really going on?