Why I Am Learning Fast.AI

This really should be titled Why I Am Learning Fast.AI and As A Geographer, You Should Too. However, this title almost takes up my entire allocation of characters on Twitter, so I opted for the shorter version. Still, the point remains. You do not have to even be involved in tech to know Artificial Intelligence (AI), especially Deep Learning (DL), is on track to dramatically transform just about every aspect of our lives. My money says you can find multiple articles in the first section of the newspaper where Artificial Intelligence was either used or is central to the topic being discussed.

Geographers, as Geographic Information Systems (GIS) practioners, we are  well aware of being awash in data. This data typically lands in our laps in one of two forms, raster or vector. Futher, raster data is either oblique (not stright overhead), or NADIR (typically called aerial or stright overhead). In GIS, we are taught to reduce the problem to a manageable size, or at least start with a limited study area to vet our ideas before scaling out to larger areas. This methodology has proven effective for at nearly 50 years, but this is about to change...fast.

This is going to change due to two factors quickly converging - explosion of drones, and explosion of AI. First, drones of all sizes are quickly becoming commonplace. GIS as an industry is wholly unprepared for the velocity and resolution of the data soon to land in our laps. Second, the field of Artificial Intelligence is moving at a pace even making the rest of tech look absolutely glacial in comparison. Four years ago being able to recognize the differerence between dogs and cats in images just over 50% of the time was considered cutting edge. Now you can identify breeds over 97% of the time by completing a how-to tutorial in an afternoon. Simply classifying cats and dogs, though, is only a small tip of the iceberg.

How we as GIS professionals are taught to approch problems, Scaling the problem either geographically or by limiting the scope of the schema (only considering a limited number of fields in vector data), is inherently limiting when using AI, though. Selecting the right fields for analysis requires specific expertise of the topic, and inherently introduces bias. Further, structuring the analysis to a specific area, typically a trade area or a market, this also introduces chance for error since this method of analysis does not take into account variations within the market or trade area. Finally, only focusing on a specific area fails to take into account possble variations outside of the study area.

As GIS practitioners, from day one we have been trained to reduce the problem to a manageable size in this way. GIS data has always been HUGE. Quantifying these relationships introduces even more size to the data. For most of our professional careers there just was neither the storage nor compute, and espcially not the tooling, to handle this massive scale of  data. Consequently, we are conditioned to invent ways to reduce the problem to a manageable size. Artifical Intelligence however, whether Machine Learning or Deep Learning, absolutely loves data. In fact, the more data you throw at it, the more it likes it! This requires a fundamental shift in thinking - how we approach problems.

It also first requires tackling a fundamental challenge, learning how to use some AI tooling, so you can then apply it back to solving Geographic problems by marrying this new tooling to GIS. If you are like myself, I like solving problems. While I really want to understand how things work, I also do not have the time to get another degree before I get something working. This is why I am starting to learn Fast.AI.

Sure, in the comments please feel free to disagree. Heck, I encourage you to. However, I am doing Fast.AI for a few reasons. First, it is designed for productivity first and foremost, which is likely the single largest consideration in my book. Second, a well supported learning course complete with videos, text, references and a GitHub repo is literally part of the website. Third, it is little more than a wrapper around PyTorch, so if I need to get deeper, or recruit one of the four genius data scientists on my team to go deeper, it is not hard to do.

The focus on productivity likely is the single largest reason why I chose to get started wth Fast.AI, and dig my way down into PyTorch if needed. After all, I am learning this to be productive at work, and while I am lucky to work for Esri, an employer encouraging us to learn and expand our skills, Esri still needs me to be productive. Fast.AI is designed to facilitate this.

Since it comes with a learning course built right into the website, this is how I am getting started. Granted, I am, at least in the near term, most interested in using it for tablular data forecasting - covered in lesson three and four. Still, I am starting from the beginning and working my way forward simply due to wanting to learn incrmentally, and become famliar with the structure and capabilities of Fast.AI. Even now, although I am not focusing as much on the raster analysis aspect, watching the videos already sparking some intersting ideas, and this is good.

Finally, as is likely the case for most people, there is always somebody smarter at work. In my case, it is everybody on my team. I work on a team with four data scientists led by a developer who is well versed in Artifical Ingelligence and a sales manager who has spent his entire career in Artificial Ingelligence. As the token GIS practioner on the team, this means I am really good at understanding what Artificial Ingelligence is capable of, and especially how it works in conjunction with GIS. What I want to get better at is plying Artificial Intelligence to be able to apply Artificial Intelligence to Geographic problems much more quickly.

This is where Fast.AI comes in, providing an avenue to explore my ideas and hunches, and as they mature, leaning on the expertise of those more intlligent than I am to double check my assumptions and tweak the tooling. Since built on top of PyTorch, it is not a large task to unpack Fast.AI to acheive much more refined results. Rather, it is an incremental refiniement of a project. This enables me to take advantage of my greatest skill as a GIS practioner, being better at accurately quantitatively modeling Geographic phenomena than anybody on my team, with the abilty to marry this with Artificial Intelligence for truly unprecedented analytical power to unearth insights without the inherent biases introduced when using traditional methods of GIS analysis. This is why I am learning Fast.AI, and as a GIS professional, you should as well.