Can an old NonStop brain learn new Deep Learning AI tricks?

In 1984 I wrote a multi-threaded NonStop TAL program as part of the SNAX/HLS product set. It was painfully hard. In 2018, I watched a few hours of videos, implemented a deep-learning algorithm, and got listed on the KAGGLE competition leaderboard[1]. I did this to learn about the human brain. Your organization needs to do this to be competitive in the marketplace.

When invited to write this column, I quickly agreed. It has been a busy time for me since formally leaving the NonStop community and I was eager to reconnect. To all my old friends and colleagues, I send greetings and best wishes.

Prior to starting a high-tech career with Tandem Computers in 1982, I was a PhD student in the biopsychology department at the University of Texas at Austin. I studied brain mechanisms of learning and memory. It was an exciting time and we published significant research[2], but nobody (and I mean nobody) was getting a job. So, I left the lab, met Jimmy Treybig and had a blast for several decades.

Recently, I have reconnected to UT Austin and resumed my study of the awesome power of the human brain. For anyone interested, I have a YouTube seminar series entitled “Your Brain: A Personal Tour[3].” It presents my analysis and finding of the most complex entity in the known universe – your brain. There is a 20-minute overview which I encourage you to view and would welcome your feedback.

Basic knowledge of the brain is expanding at an exponential rate. Billions of dollars are being spent to understand fundamental building blocks of brain organization and operation. Such research is impacting our ability to repair brain systems from neural trauma – but perhaps more importantly – to use brain systems to directly interface to prosthetic devices, allowing quadriplegics to walk again and the limbless to reach again.

I began a search into the world of deep learning to help me progress my quest in the understanding of the human brain. Fundamental research will only take us so far. Brain systems are sufficiently complex to require neural modelling – and that leads to the world of machine learning and deep learning. And my oh my, what a world I found. I was just shocked (…shocked) at how different that world was from the world of programming I knew in the SNAX/HLS days.

The first aspect of this world is the unbounded collaboration with the technical universe. There is an immense set of information, templates, models, forums, video courses and interactive “notebooks.” Since I wanted to study deep learning models – I wanted to get a quick handle on PYTHON. Turns out there are multiple “learn basics of Python in one hour” videos. So, I watched one, and got sufficiently dangerous to begin coding in a Jupyter notebook! I was just flabbergasted at the support tools and models available to help even on “old brain” like mine learn this new task.

I am on the board of MODIUS – a software company implementing machine/deep learning algorithms into our fundamental product set – and it was from our CEO that I learned about KAGGLE. Over 500,000 participants from almost 200 countries actively contribute to this site in order to widen the boundaries of machine learning, predictive analytics, and deep learning. The sense of community is pervasive and the information content is outstanding. If your organization is not making use of this site – you are failing to leverage an excellent resource. The level of support is so outstanding that a dude like me could produce a competitive predictive model (not from scratch – but using an available model on the website) in about four hours. That is how I became listed on the competition leaderboard (…proudly standing at 3387 out of 5422!…;). It is little wonder that ALPHABET acquired Kaggle in 2016.

I also joined the “Deep Learning” meetup groups in Austin. Yes, Austin is a high-tech community. You may not have access to the same groups in your location. But, I bet you will find a lively community if you look. I was surprised. Most surprising was the level of community among the companies that were represented (Walmart, Google, Facebook, and several startups). Knowledge seems to be assumed as a free commodity – if you are willing to work to learn it. Again, this really surprised me – seeing programmers from competitive organizations sharing skills (not IP) to better the capabilities of the overall community.

It was in the Deep Learning Community that I met the first AI model which explicitly modeled the human brain[4]. I was dumbfounded. The model architected a viable solution to several issues in human brain research – and the architecture mapped naturally to the anatomy of the human brain! Awesome. Just Awesome. Of particular note was the fact that agents in this model of the brain DREAM. That is to say that the models runs agents in a “dream world” for lack of a better term. Learning from that dreamworld is then fed back into real-world operation and has a very positive effect. I encourage you to at least skim the paper – I think you will find it fascinating.

Then from the Deep Learning Community I found the FASTAI community[5]. Think of FASTAI as Kaggle on steroids. Their vision is “deep learning accessible to all” – so I decided to challenge that. I dove into the course work – which consists of both video and juypter notebooks. In about 4 hours, I had customized an image classifier to distinguish between our two border collies! While that was important to me – I am not sure that is important to you. But, note well: image classifiers have been patented in FRAUD DETECTION AND PREVENTION[6]. And this is where my old NonStop world bumps into the Deep Learning and Brain worlds.

So much progress has been made to “open up” the NonStop world. It is really impressive. Given that kind of progress, incorporation of the such Deep Learning is a straightforward affair:

• Get your AI programmer dudes to check out Kaggle, FastAI, and perhaps DataRobot (and I am sure there are many others – just pick some) for useful infrastructure and skill information. NB: it does not have to be rocket-level data scientists. If I can do this with my old brain, your brains can do it too!

• Get a modelling test rig up to train and validate any models. You have a range of options from InfiniBand RDMA access aka NSADI (NonStop Application Development Interface) capabilities. You will most likely be in a hybrid mode here with the model being trained on GPU-intense servers (e.g. HPE Apollo environment or another GPU-intensive server).

• Refine and then evaluate using hybrid or virtual system in production. NonStop systems can efficiently perform in either. But care should be exercised in planning and implementation here. It is production, after all.

The world has radically changed since I wrote my first multi-threaded NonStop TAL program. An astonishing level of community and information allows each and every organization to evaluate, implement, and use the value of deep learning and other AI techniques – TODAY.

Yes – this area may seem like brain surgery (and some of it IS – trust me, I know). But the impact of community and new tools enable an order-of-magnitude increase in productivity for everyorganization. Reusable models – trained on YOUR DATA – offer new potential to augment or differentiate products.

So, you can say about deep learning what they say about the brain – Use it or lose it.

All the best to my friends and colleagues, I welcome your feedback and messages.

Cheers,

Steve

Steve.Saltwick@gmail.com

[1] Just look for “Steve Saltwick” in this Kaggle leaderboard <Kaggle>

[2]Gabriel, Saltwick, Miller. “Conditioning and Reversal of Short-Latency Multiple-Unit Responses in the Rabbit Medial Geniculate Nucleus” SCIENCE 189 (26 Sept 1975) pp 1108-1109

[3]Playlist is located <here>

[4]See the GITHUB article <here>

[5]See course material <here>

[6]See representative article <here>