Christopher Danz

The here mentioned techniques are only a small fraction of potential aids in your toolbox. Having an overview of available methods and current research on Explainable AI is important for data scientists and ML practitioners to explain and interpret model behavior. Applying XAI techniques should become an industry-standard before deploying models, not just since the upcoming of regulations but for the sake of anti-discriminatory behavior of the systems we develop. ML systems are biased and shedding light onto them by applying XAI technique is the only way to ensure we don't roll out biased and discrimnatory models to our customers and the world. Study these methods, keep up with the research and educate your team and customers.

There is a good reason why Lloyd Shapley received a nobel prize. The Shapley values, which were named after him, are a terrific way to explain predictions of ML systems into easy to understand contributions of single features. What makes shapley values stand out and make it so versatile for any ML application is its model-agnostic approach as well as the relatively easy interpretability of the resulting plots. You can apply this technique to any model, from simple linear regression to complex neural nets and get the same kind of coloured plot, explaining how much each feature influenced the prediction of the ML model.

Evaluating the model complexity is a no-brainer since everyone in the industry knows that neural networks are inherently difficult to interpret while a tree-based model offers much more possibilities to explain its decision-making. What is often overlooked though in my opinion, is the evaluation of data and features the model was trained on. High dimensionality, noisy data and irrelevant features can reduce interpretability of your ML system. It is crucial to consider dimensionality reduction techniques or feature selection to simplify the model without significantly sacrificing performance. Only on top of that I think it makes sense to think about simpler model selection and interpretability tools like LIME and SHAP etc.

This is probably the entry point with the lowest entry hurdle into the AI community. Actively participating in discussions, question-answering, problem-solving, knowledge-sharing is a good way to engage with the community and to get in touch with peers of different experience. Like with most social networks, participating on online platforms and forums is about consistency to maximise leverage. The forums basically allow sharing knowledge, solving problems, and even presenting opportunities for collaboration in the AI world and serve as an easy starting point to network in the AI community.

Conferences seem to be a terrific option to network, looking for jobs, keeping up to date with latest research etc.. Since most AI conferences are quite expensive though it is not suitable for everyone unfortunately. I recommend to check out cheap alternatives (ranging from free of charge to a few hundred €/$ entrance fee) to get access. Free conferences usually sell your data (e-mail adress etc.) to third parties which will send you advertisement later on but it can be worth it due to the exposure to experts and companies at the conference. Additionally, consider applying to speak at a conference and present your work / project / journey. It creates a ridiculous amount of new opportunities for you and helps with personal growth.

Meetups are a great and easy way to get in touch and network with a local AI community. Unfortunately though, meetups are only most common in metropolitan cities and sort of underrepresented in smaller cities / rural areas. I have seen many cities where no meetups on AI, data science etc. exist while the size of the city would clearly indicate a sufficient amount of habitants to find likeminded people in the AI community. Therefore, if you're living in a region that does not offer regularly occuring AI meetups, create one on your own! Start the initiative in your own city and be consistent - there are probably more likeminded AI enthusiasts than you would expect. 😊 The reward will be immense!