In case you missed our session on Artificial Intelligence and Machine Learning (AI/ML) at the Insights Associations’ NEXT conference last week, I thought I would share a bit on the blog about what you missed. We had a full room, with some great questions both during and after the session. However, 30 minutes wasn’t enough time to cover everything thoroughly. In the end we agreed on four takeaways:
So today I thought I would share seven additional points about AI/ML that I often get questions on, and then at the end of this post I’m going to share the ‘Cliff Notes’, i.e. I’m going to share just the 2 most important things you really need to know. So, unless you want to geek out with me a bit, feel free to scroll to the bottom.
OK, first, before we can talk about anything, we need to define what Artificial Intelligence (AI) is and isn’t.
AI, and more specifically machine learning (ML) is a term that is abused almost as often as it is used. On the one hand this is because a lot of folks are inaccurately claiming to be using it, but also because not unlike big data, its definitions can be a bit unclear, and don’t always make perfect sense.
Let’s take this common 3-part regression analysis process:
This process, even if automated would not be considered machine learning. However, switch out regression with a machine learning technique like Neural Nets, SVM, Decision Trees or Random Forests and bang, it’s machine learning. Why?
Regression models are also created to predict something, and they also require training data. If the data is linear, then there is no way any of these other models will beat regression in terms of ROI. So why would regression not be considered machine learning?
Who knows. Probably just because the first writers of the first few academic papers on ML refenced these techniques and not regression as ML. It really doesn’t make much sense.
Some ML approaches are binary like SVM (Support Vector Machines), for predicting something like male or female, and others like Decision Trees are multi class classification.
If you are using decision trees to predict an NPS rating on an 11 point scale then that’s a multi class problem. However, you can ‘trick’ binary techniques like SVM to solve the multi class problem by setting it up to run multiple times.
Either way, you are predicting something.
Depending on the approach used, like Neural Nets for instance, training a model can take several days on a normal computer. There are other issues with Neural Nets as well, like the difficulty for humans to understand and control what they are doing.
But let’s focus on speed for now. Of course, if you can apply a previously trained model on very similar data, then results will be very fast indeed. This isn’t always possible though
If your goal is to insert ML into a process to solve a problem which a user is waiting for, then training an algorithm might not be a very good solution. If another technique, ‘machine learning’ or not, can solve the problem much faster with similar accuracy, then that should be the approach to use.
I’ll pick on Neural Nets a bit more, because they are almost a buzz word unto themselves. That’s because a lot of people have claimed they work like the human brain. This isn’t true. If we’re going to be honest, we’re not sure how the human brain works. In fact, what we do know about the human brain makes me think the human brain is quite different.
The human brain contains nearly 90 billion neurons, each with thousands of synapses. Some of these fire and send information for a given task, some will not fire, and yet others fire and do not send any information. The fact is we don’t know exactly why. This is something we are still working on with hopes that new more powerful quantum computers may give us some insight.
We can however map some functions of the brain to robotics to do things like lift arms, without knowing exactly what happens in between.
There is one problematic similarity between the brain and Neural Nets though. That is, we’re not quite sure how Neural Nets work either. When running a Neural Net, we cannot easily control or explain what happens in the intermediary nodes. So, this (along with speed I mentioned earlier) is more of a reason to be cautious about using Neural Nets.
Are all problems best solved with ML? No, probably not.
Take pricing as an example. People have solved for this problem for years, and there are many different solutions depending on your unique situation. These solutions can factor in everything from supply and demand, to cost.
Introducing machine learning, or even just a simpler non-ML based automated technique can sometimes cause unexpected problems. As an example, consider the automated real-time pricing model which Uber used to model supply and demand as inputs. When fares skyrocketed to over $1,000 as drunk people were looking for a ride on New Years eve, the model created a lot of angry customers and bad press.
More on dangers of AI/ML in a bit…
One of the reasons ML is often touted as a solution is because of how much better than humans computers allegedly are. While theoretically there is truth to this, when applied to real world situations we often see a less ideal picture.
Take self driving cars as an example. Until recently they were touted as “safer than humans”. That was until they began crashing and blowing up.
Take the recent Tesla crash as an example. The AI/ML accidentally latched onto an older faded lane line rather than the newly painted correct lane line and proceeded without breaking, at full speed, into a head on collision with a divider. A specific fatal mistake no human would have been likely to make.
The truth is if we remove driving under the influence and falling asleep from the statistics (two things that are illegal anyway), then human accident statistics are incredibly low.
This is an important one. IBM Watson might be able to Google Lady Gaga’s age quickly, but Watson will be completely useless in identifying her in a picture. Machine learning solutions are extremely context specific.
This context specificity also comes into play when training any type of model. The model will only be as good as the training data used to create it, and the similarity to future data it is uses for predictions.
Model validation methods only test the accuracy of the model on the same exact type of data (typically a random portion of the same data), it does not test the quality of the data itself, nor the application of this model on future data other than the training data.
Be wary of anyone who claims their AI does all sorts of things well, or does it with extremely 100% accuracy.
If some of the above points make it sound as if I’m not bullish on machine learning, I want to clarify that in fact I am. At OdinText we are continuously testing and implementing ML when it makes sense. I’m confident that we as an industry will get better and better at machine learning.
In the case of Tesla above, there are numerous ways to make the computers more efficient, including using special paint that would be easier for computer cameras to see, and traffic lights that send signals telling the computer stating “I am red”, “I am Green” etc., rather than having to guess it via color/light sensing. Things will certainly change, and AI/ML will play an important part.
However, immediately after my talk at the Insights Association I had two very interesting conversations on how to “identify the right AI solution”? In both instances, the buyer was evaluating vendors that made a lot of claims. Way too many in my opinion.
If you forget everything else from today’s post, please remember these two simple Cliff Notes on AI:
There’s no way to know whether something is AI or not without looking at the code.
Unlike academics who share everything under peer review, companies protect their IP, Trade Secrets and code, so there will technically be no way for you to evaluate whether something actually is “AI” or not.
However, the good news is, this makes your job easier. Rather than reviewing someone’s code your job is simply still to decide whether the products solves your needs well or not.
In fact, in my opinion it is far more important to choose a vendor who is honest with you about what they can do to solve your problems. If a vendor claims they have AI everywhere that solves all kinds of various needs, and does so with 100% accuracy, run!