‘Artificial intelligence’ (AI) and ‘machine learning’ are sneaking into every part of our daily lives.
How do we tell the realizable use cases available now from what’s projected?
In 1850, Michael Faraday released his seminal paper on electromagnetic induction. However, some 100 years after Faraday’s paper, that electricity become widely adopted, standardized, and ultimately reached a level, which meant it was superior to other technologies,
such as steam.
The same can be said of machine learning and AI. The term artificial intelligence was coined by John McCarthy in 1956 during his work at MIT. There has been steady, incremental development, but it wasn’t until 2015 with Google DeepMind’s Alpha Go defeating
the chess world champion, Li Ke, in 2015, that we saw some exponential gains. What was special about this victory was that the machine actually devised a move that had not been considered by expert human players. More recently we’ve seen Google’s Duplex mimicking
a human-like voice.
The harsh reality of these examples is that they are all machine learning-based, and while new algorithms may have been designed and larger data sets pushed through them, the premise still remains that these are narrow use cases.
This does not mean we are years away from an AI revolution, but it does mean there is a set of criteria, which should drive investment in this area. The ‘do nothing’ option was ruled out several years ago, so action must be taken.
But, how do we target investment and ensure we are unlocking real business value? Here are 8 actions to consider:
1. Train all algorithms on YOUR data. While there may be a set of algorithms available, this must be trained on the data in which it will operate. Vendors claiming otherwise should be shown the door.
2. Select a narrow use case. The world’s largest technology companies are spending billions trying to automate relatively simple tasks for humans. A broad “artificial intelligence for everything” approach will only lead to misguided investment.
3. Categorize your problems based on simple, mathematical principles. There is no need for a PhD in Mathematics to solve the problem of classification, ranking, numeric prediction, or optimization. 90-95% of the challenges faced today will fall into
the first three buckets.
4. Remember not all algorithms are equal. An algorithm that has been trained on a specific use case over time will be more accurate than a generic classifier or tool. Be sure to check how long the algorithms have been used on the specific problem
you’re trying to solve.
5. Decide what are the value levers. Everyone wants to have their cake and to eat it too, but what is the most important or top two value levers – faster, cheaper, more accurate? Challenge your team to come up with these and quantify
6. Do away with small-scale proofs-of-concept. Although useful for building management support, these are often too small with respect to the data sets to give any real indication. As a rule of thumb, training data sets tend to be 15-20% of the total
data. Pilots of this size tend to be a better measure of future benefit.
7. Don’t be afraid to pick-and-mix technologies from an ecosystem of partners. With many mature integration technologies available, it is now possible to select specialized algorithms for specific automation tasks.
8. Build a capability. Building algorithms that pull value levers (for example, reduce response time) is a competitive advantage for your company and they should be treated as assets. Ensure your in-house data science team does not become a machine
learning shop and is focused on giving users and customers the tools to deliver better outcomes for your business
The question of IF there should be an investment in machine learning-based technologies has already been made.
Delay now will only lead to future challenges. However, the question of WHAT to invest in is still open. Now is the time to address it.
External | what does this mean?