AI – From Not Working to Neural Networking

Last time I wrote a blog about the Second Machine Age. It was based on the book with the same name by Erik Brynjolfsson and Andrew McAfee.The question remains, after many a false dawn is AI ready to go mainstream? In this blog I’ll look at the many changes in mathematics, hardware and in software that are propelling this Second Machine Age. My previous piece was more insurance focused but if you will let me indulge, let me “Geek Out” for this piece, then I will promise you I will go back to good ol’ Loss Ratios, NCOR, Coverages etc. soon.

Anyway let’s get cracking!

<Geek>

Alan Turing in the 1950 paper ‘Computing Machinery and Intelligence’, proposed the question ‘can machines think?’ Six year later at Dartmouth College the term Artificial Intelligence appeared in the computing lingo. During the proceeding 60 years it has vacillated from being hype, and a flickering promise to over optimism but the common denominator has been about under delivering. In short all of it would point to answering Alan Turing’s big question with a big “No”. Fast forward to 2016, what has changed? Is it still hype? Is AlphaGo’s win over Lee Sedol just clever marketing? Have we gone from avoiding the term (cue Expert Systems in the 80’s) to just about every man and his dog talking about it across every industry? Have we actually gone from “Not working to Neural networking”?

In my humble opinion the overwhelming answer is that we are at an inflection point for AI. I think we are still in the early stages of this, so today I want to look at how we got to this inflection point. Primarily it comes together as a series of developments in Deep Learning, Explosion of Data (Volume, Variety, Velocity and Veracity), Computational Hardware and Open Source AI Software. Let me expand on these in order of importance.

1) Deep Learning – As part of the 2012 ImageNet challenge, a team from University of Toronto, led by Geoff Hinton, made a break through using a biology inspired technique called Deep Learning, mimicking the way our brain operates. These were the Convolution Neural Networks (CNNs – if you’re inspired read about here). This technique took the image labeling from 72% to 85% and more recently took it to 96%. The average human labels at a success rate of 95%!! Suddenly we were getting Neural Networks that struggled to be 4 layers deep, to having 150+ layers from Microsoft. Deep Learning takes a lot of the “expertize” and “tacit” knowledge out of the equation which was traditionally used in more pattern or regression based models and turns it into a game of data and computational power.

2) Explosion of Data – It doesn’t take an expert to understand that pretty much everything is digital these days and the more digital input and outputs we have the more powerful AI systems get. The versatility of Deep Learning with data is that it can be applied to any Vertical AI* problem, as long as it has enough inputs and outputs. It is independent of the complexity or the domain knowledge required to solve problems (e.g. the years of experience to price commercial insurance or the years of studying that goes into diagnosing cancers), but more a function of the input and output data. As they say Deep Learning systems do not discriminate against the “color of your collar”…more on that in my next blog AI impact on Labor Markets.

*For the purposes of this I have described Vertical as narrower problems where we can use Supervised, Unsupervised or Reinforcement learning to solve few use cases. Horizontal AIs like Siri, Cortana and Alexa are heading down the Artificial General Intelligence (AGI) path.

3) Computational Hardware – Graphics Processing Units (GPUs) have been developing for 30+ years. The impact they have on AI is that their parallel computational ability has reduced training iterations from months to days.

NVIDIA the company “formerly” known as powering the gaming revolution, recently grew 15% on the partnership with Tesla. They have recently released the NVIDIA Inception Program, which provides both access to the SDKs as well as early release hardware. NVIDIA is currently reducing training of Deep Neural Networks 50 fold in 3 years (well ahead of Moor’s Law). It will not be that long when Google, Microsoft, Amazon and IBM release their own chips. In fact at Google I/O they announced the Tensor Processing Unit (TPU).

4) Open Sourcing AI Software – there is so much choice in AI software, that there is actually a Periodic Table for it. However in my opinion two major changes fired up the rise of AI since 2013 (post Hadoop). First Google’s TensorFlow and secondly Watson Developer Cloud enabled companies with little to no data science experience to jump start their AI capabilities. More recently Kaggle has enabled to even outsource your data science practice while keeping your engineering practice. Before you say that will never happen in insurance here it is from Allstate and a more left field one from State Farm

It is arguable that there are many other contributors including education being a factor. Coursera and Udacity have been wildly popular, but I see this at a distant number 5.

As you can see we have 4 areas of development expanding at faster than Exponential pace, and it is anyone’s guess as to the speed they will impact your industry. One thing that is guaranteed is that a Vertical AI will impact Financial Services and specifically Insurance within the next 2 years in a big big way.

“Two years, that is ages away, I want to know where we are at now!!” Well I think Siri said it best in the following 3 screens (of which 2 of them are real – you can guess which ones!!). Essentially we are crossing the bridge (pardon the pun) where AI can actually be very useful, and even saves lives.

Before Artificial Intelligence -Before
AfterNowSiri
Umm…Artificial Intelligence SIRI

As you can see, we are still not quite at a stage AI can handle everything

</Geek>

I would really welcome your thoughts on if you think I got the ranking right or if you think I am completely missing a factor for AI actually working. You may actually disagree to the point that AI is actually working in your industry? Tell me if it is and what you think it will take to make it work and when you think it will happen?

That ends my “Geek out”, next week let’s talk about Jobs. No, not Steve but the impact of AI on Labor markets. The week after I will also discuss impact on Insurance!

lakshan
Author
lakshan
Lakshan is an experienced global executive that has worked across technology, venture capital, insurance, wealth management, construction, manufacturing and mining.

He has worked in corporates across the globe and has rounded it off with a MBA and a couple of Exec Programs, but please don't hold that against him, as he is busily unlearning everything he learnt over the last 20 years to stay relevant for the next 20 years 🙂

As CTO, he is bringing in exponential technology that will define the next 10 years into the Intellect SEEC products. His current projects revolve around AI and Blockchain.

He holds global patents in several technologies and is an investor and advisor to numerous Fintech startups.

He is a sports fan, music aficionado and animal lover and his claim to fame is that he has trained his pet bulldog Mortimer to obey him whenever the bulldog wants to 🙂
Previous Post
2MA Insurance

THE SECOND MACHINE AGE : Underwriting 3.0

Next Post

AI does not care about the color of your collar!

@ 2018 Intellect Design Arena Inc. | 20 Corporate Place South, Piscataway, NJ 08854, USA.