Getting Into AI: 5 Lessons from a Product Manager

Since starting out as a product manager working with cutting-edge AI and Deep Learning technologies, I’ve gathered some useful tools and information. I wanted to share the knowledge for those getting started in this field, whether as product managers or in other roles at tech companies who want to understand more about the underlying technologies.

1. First, make sure you understand the basics

If you don’t have a technical background, your first step will be to learn more about AI. You should get to know what is behind the terms that are being discussed. While AI can be as simple as an algorithm, machine learning, and deep learning are more complicated.

Machine learning and deep learning are both types of AI. In short, machine learning is AI that can automatically adapt with minimal human interference. Deep learning is a subset of machine learning that uses artificial neural networks to mimic the learning process of the human brain.

I compiled two of my favorite resources for learning more about all of the above here. Deep Lizard is a great channel that breaks down many deep learning concepts such as GANs and CNNs in a way that I found really easy to understand. Depending on how deep you would like to go, Andrew Ng’s courses on Neural Networks and Deep Learning were my source for understanding supervised vs. unsupervised learning, loss functions, parameters, weights, and biases + forward and backward propagation.
A basic understanding of linear algebra will help here, but you can pick up the basics through this course as well.

2. Follow leading companies

Tencent, Alibaba, Adobe, nvidia, OpenAI, MegVii, SenseTime, Google, and Facebook are all working on amazing AI , ML and DL technologies. Many companies have made their models open source, for various reasons, such as improving their own technologies and collecting data. It is definitely changing the game, and a key part of the current atmosphere in development. You can look for information on specific frameworks like Tensorflow, or set a google news alert for companies like OpenAI.

3. Get to know the field and the players

To work with or stay up to date with the latest academic research on AI, get to know the key players. For computer vision, for example, this includes:

• The International Conference on Computer Vision
• Computer Vision and Pattern Recognition Conference (CVPR) from the CVF and IEEE
• Institute of Electrical and Electronics Engineers (IEEE)
• The Computer Vision Foundation (CVF)

4. Stay up to date on hot new papers

If your team is looking for new methodologies to explore, you can find inspiration and examples in academia. On Google scholar, you can search for a relevant topic like inpainting. For up-to-date technologies, filter the results for the last couple of years. When you find a good paper from one of the above conferences with lots of citations, you can set a Google scholar alert for citation of that paper and for the author as well as follow the authors on Github, following and starring the repository if code was included. In addition, if you click on the citations (meaning all the papers that cited the first one you clicked on), you can filter for new papers again and then do the same with all of these for even more hot new research.

Make sure that your team is checking out the papers and trying to see that they are not cherrypicking photos to make the results seem better than they are. Often, when these results are tested on real-world cases, the results are less than impressive.

Deeplearn.org will do much of the work I described above for you, in the hot papers section, and this site will give you just papers with code if you are looking to jump straight to testing the models without writing any new code.

5. Always keep your users in mind

Exciting new papers often show almost unbelievable results. You might see an inpainting model that can fill in a realistic photo from the basis of an almost all-black photo — it’s astonishing. However, when testing on high-quality images that users will want to edit in a photo editing app for example, or images with many faces, more detail, etc., you will often end up disappointed. Academic research is not thinking about your specific end user, so you must keep them in mind whenever approaching this discipline.

The large companies that come out with amazing AI technologies are often able to do so because of their greatest asset — large datasets (and the financial resources for training large models). While it can be very expensive to train these state-of-the-art models, you can take a page from the large companies’ books and leverage your unique datasets to create a better experience for users. If you already have traffic, and you know what kind of photos your users will end up using the AI model on, then you can ask users for consent to train on their images, and you will have a custom model that fits your needs. Make sure to respect their privacy and always deal carefully with users’ information.