- John Carpenter
- Originally posted here https://medium.com/towards-data-science/5-reasons-why-ai-will-have-a-larger-impact-than-most-people-expect-e9d7e60dcdd3 *
5 Reasons Why AI Will Have a Larger Impact Than Most People Expect
Nothing gets more clicks than a clickbait article with a leading title like “X reasons why you need to be concerned that Y will happen”. Newspapers have been doing it for years to drag people into their readership numbers. Now online magazines and blogs have really refined the practice. It makes sense since a good title increases readership, increases ad views and increases profitability.
In the past, generating that title was once the work of creative writers and less scrupulous editors. Those titles are now being generated faster using an AI algorithm. See Lars Eidnes article on Generating link titles with RNN (while full of content, that article would benefit from a more titilating title. ‘Buzzfeed did this one thing to their articles and you won’t believe the results!’). Even full articles are now written by AI (Sadly not this one..). Currently AI writing is limited by minimal analysis and covers things like stock reports and sports scores however, in practice there is very little stopping them from writing more in depth analysis. Poetry, songs and films all have been generated by AI programs to varying levels of success.
So we can assume AI will have a deep impact on news organizations and writing. But those are specific use cases for a narrow field. In reality I could have started this article with almost any field or industry and the results would have been the same. AI will fundamentally change numerous organizations and industries not because it is another program but because it is a transformative technology similar to internet, cloud computing and blockchain.
What makes AI and machine learning (ML) algorithms different from ordinary software has much to do with the process in which it solves problems. AI/ML uses simple algorithms to learn and find patterns in data. It doesn’t require any domain knowledge for a field and can be run with almost any data set. It’s a completely flexible technology that can be applied to any existing or new problem, and it doesn’t require deep domain knowledge. AI engineers can write algorithms to identify tumors in MRI scans without having ever visited a medical school or understanding how MRI works. That same algorithm can identify hot dogs or not dogs in images. Same algorithm, slightly different benefits for humanity.
But consider lots of technologies have had an major impact on industries. Ask Blockbuster about streaming video or Nokia about Android. But those technologies either required significant expertise, expense or both to implement. Those transformative technologies were limited to those companies (always companies) that could pull that off. AI/ML is not limited by those same constraints. It’s cheap, accessible and flexible enough to disrupt any industry.
1. AI/ML Algorithms and Techniques are Simple and Easy to Pickup.
As an AI engineer, the scope of the knowledge I need to implement AI/ML can be attained in a 9 month course from many universities or online (thanks Udacity!). The algorithms and techniques are simple and straight-forward, and require little more than slightly above average math skills. The barrier to entry for most programmers or engineers is embarrassingly low. This ensures that as the capability of AI grows, the number of AI engineers available will grow with it. I expect that even these minimal barriers will drop as companies like IBM, Google and Facebook each rush to train more engineers.
*Of course, if you want to be a full-fledged AI researcher expect to spend much, much more time in a classroom. But AI technicians (those who implement) can come from anywhere in the world and don’t require a lot of background skills.
2. Algorithms don’t require domain knowledge
AI/ML is nothing more than the ability to find correlations within data, even if they are not immediately evident. In one example, AI/ML vision applications identify shape, color and texture in images to help classify different parts of the image. This is all done automatically. Nothing is ever coded to separate those characteristics in the image. For the most part researchers aren’t even sure why certain characteristics are isolated. By passing in enough correct classifications (this is a cat/this isn’t a cat) and training the system, AI will learn the features that are characteristic of a cat. There is no requirement to feed any cat attributes into the algorithm.
The most powerful part of this feature is that the domain expertise is no longer limited to a handful of individuals. Work on complicated domains such as medicine or astronomy no longer require deep knowledge of the structure of the data. The engineers can apply the same algorithms to work in any domain that circumvents that knowledge. This opens up many domains to disruption that were traditionally considered expert domains (legal, medicine, engineering)
3. AI/ML Algorithms are Exponential and Shared
Additionally since AI/ML algorithms don’t require any domain specific knowledge or application, you can use knowledge gained in one industry and apply that to many others. This knowledge becomes shared across industries without having to share IP or trade secrets. Algorithms can be passed and shared easily. https://www.kaggle.com/ is a whole community dedicated to opening access to data and models necessary to create and improve AI algorithms.
Take an example from image recognition. AI models for image recognition are widely shared and even compete against each other every year. The models are open and shared throughout the community. What is powerful about these models is that they can be purposed to identify almost any type of image. There are nuances of course but the algorithm required to identify street signs for self-driving cars is essentially identical to the one used for identifying tumors in x-ray images. The work from one model can be shared in all the others.
Even a previous requirement of lots and lots of data is becoming easier with Transfer Learning, researchers are able to leverage existing trained models and apply them to custom situations. Trained on 1.4 million images, a VGG model can be repurposed for identifying dog breeds with relatively few images. So the requirements of processing lots and lots of data to get a good result is drastically reduced.
4. AI is more accessible to more people worldwide.
ML/AI explosion is partially due to the availability of computing resources required to calculate the solutions. Cloud computing allows individuals to access the scale of computing power that was previously unavailable unless you were a large universities, organization or governments. Now a farmer in Japan can train a cucumber sorter using a few dollars worth of resources. Cloud companies are throwing more and more investment into AI. Google TPU has built hardware solely for the purposes of solving AI/ML algorithms faster and more efficiently. Problems that were once only achievable to those with the resources can now be tackled by individuals. Industries that only had a handful of companies that could really disrupt the industry are faced with thousands of potential competitors.
5. AI research is open, fast and accelerating.
AI/ML research is a very hot topic around the world. Organizations are throwing millions into tweaking and updating algorithms. For the time being, these algorithms are all being made open and available in scientific journals and online.
There are standardized tools Tensorflow, Caffe, Keras tools that provide a standardized language for describing AI models and implementing them.
Data is becoming more available and open.
We are now in a situation where there are better models, more data and smarter decisions being made by AI algorithms. These are done at a fraction of the traditional cost and can be done with a limited budget and expertise. This will change a lot of industries.
AI teachers, one per student, that will guide and personalize an education program perfectly suited to their learning style.
AI farms, that will monitor, grow and harvest crops without any human intervention.
AI Shoppers, that will source, negotiate and arrange shipping on everything from groceries to pipelines.
Three simple examples that can easily provide a 10x advantage in their industries. The technology is not that far away from being commercial now….remember a public internet is barely over 25 years old now, and it’s easy to see the impact it has had.