In this disruptive world we find ourselves in, it’s the AI start-ups that are catching most peoples attention. Whilst all the big global organisations are investing into ML and DL applications, its often the more agile start-ups that are bringing us the new innovations in this space.
Their strengths lie in their creativity, agility, and the fact they are not constrained by corporate structures, giving them the flexibility and freedom to develop projects and applications that might never see the light of day in other organisations. This is one of the primary reasons I like working on projects with AI start-ups; they are dynamic, and you never know what direction the project will take. Because of their agility and creativeness, you find yourself never standing still for long.
However, life for an AI start-up is not easy. Firstly, there has to be a creative genius with a clear vision driving the project and essential funding. Most importantly, two areas are the Achilles heel for AI start-ups. Chances are that you are restricted by the amount of data you can work with (GANS and Transfer models will help), and also by the amount of compute you have available.
The key to any successful AI start-up is to develop and commercially deploy your applications before somebody else does, this really is a classic case of whoever launches products to market fastest will most likely have a long-term advantage over their next rival. So, the restriction for AI start-ups in relation to the amount of compute they have is a serious one (I will cover this again later)
AI start-ups, because of there diversity, will typically be using all the current techniques that we are all very familiar with: Narrow, General (good luck with this one) ML, DL, NLP, reinforcement, CV, robotics, and so. Also, the diversity in the neural networks THAT they are running is again, well... diverse! Just to look at a few: GANS, FF, RBF, DCN,DN, DRN,KN...you get the picture.
But just how big is the UK in this space?
Well, today there is the region of 900 AI start-ups that are known about that are not in stealth mode, and of these, market data suggests around 400 are at an early stage (angel or seed investment), which, when compared to countries in the EU, suggests that the UK is not only a hotspot for AI in general terms, but also for AI start-ups. This is in thanks to the fabulous wealth of talent the UK enjoys from academia. If you look at the work currently being done at UCL, KCL, Oxford, Edinburgh, or any other of the fantastic universities we seem blessed with, the amount of talent the UK has is rife.
So why the constraint on the amount of compute you have?
Firstly, and most obviously, is funding, given the fact that you are a start-up it would suggest that financing compute systems could be a challenge. Investors will continue financing your business as long as they can see a return on their investment, and, more importantly, they can see progress on the experiments you are running.
The other key aspect is what is the right compute platform your business needs. The problem here is that the pace of technology in this space is so quick to evolve that if you blink, you’re behind the game. So, keeping up with technology in itself can be hard (see my article on NVLink as an example), you don’t need to spend a fortune on technology that will seriously impact your business today with minimal costs.
“Developing deep learning models is a bit like being a software developer 40 years ago. You have to worry about the hardware and the hardware is changing quite quickly… Being at the forefront of deep learning also involves being at the forefront of what hardware can do.” - Phil Blunsom, Oxford University and DeepMind
The problem with all this is that Deep Learning is computationally very expensive to do. Let’s have a quick look at AlphaGo:
· AI/GI & Deep Learning: how AlphaGo fits:
AI/GI = machine perception (speech, image, video, gesture, touch...)
+ machine cognition (natural language, reasoning, attention, memory/learning,
knowledge, decision making, action, interaction/conversation, ...)
AGI: AI that is flexible, general, adaptive, learning from 1st principles
Deep Learning + Reinforcement/Unsupervised Learning ?AI/AGI
Computationally this set up is very taxing on the system, and with AlphaZero I read it took around 5,000 first gen tensor units to generate self playing games. Now I know not many AI start-ups will be doing this kind of work, but the point is Deep Learning is only possible due to the power of NVIDIA GPU's, and the need for these GPU's is being increased with everyone's need for faster and faster training times without any accuracy loss. Again, just like it is important for AI start-ups to deploy and launch their services as fast as possible, the same is true for big business.
So maybe the solution for AI start-ups is the cloud. You don’t have the upfront costs associated with purchase, you don’t need any system administrators to run environments, and generally you can get what you need fairly fast and efficiently via the usual suspects. And in some cases, this would be a very logical step to take, however for most of the companies I engage with, the answer would be no.
There are two issues here to deal with. The first and most obviously is the long-term cost of rental vs ownership. In short, I have never seen a rental cloud model that works out cheaper or even comes close to the same cost as ownership. Normally with most of the platforms I deal with, the client could have bought it after 12-18 months of the associated cloud costs, so anything after this point is a pure loss(?)
Secondly, and by far the most important reason, is data security. Imagine as an AI start up or any other type of company for that matter your data is breached and your IP exposed...this would be a disaster for any company but more importantly so for a AI start-up.
Now I'm not saying cloud is bad, of course not. But when it comes to AI based applications, someone developing the security posed by cloud would normally make someone pause for thought.
What to do?
Cloud under the correct conditions can be the perfect answer, but if it isn't there are some options available to you:
Hardware options doesn't have to mean spending £20-200K to get the desired results, contrary to what you have probably been told. In fact, with most of the AI start-ups I deal with, they are completely fine simply running some desktops with bunches of Titan cards. In fact, with the only real limitation being the memory size 12GB (reference image size to fit on all the memory - think medical images), these cards are still delivering fantastic cost versus performance benefits. And most AI start-ups are fine running these types of systems. The trick is to get these systems optimised as best you can with all the supporting software (Ubuntu 16.04.3LTS, Frameworks of your choice, cudnn, dIGITS). A system like this, with, say x2 Titan XP cards and reasonable specification, wouldn’t cost more than £3-5K per box. And systems like this are more than capable for development work, and basic model training. Of course, if you are training models for application deployment, and those models are far more advanced, then you will be looking at the next level of investment.
Is it just GPU compute?
In relation to the compute power question, its not just a simple hardware question though? What I mean is the limitations of the skill sets of the individual. Most of the best Data Scientists are not just great because of their technical ability, its sometimes more to do with their creativity. Looking at a problem and finding new ways to approach it, this in a lot of cases is the deciding factor. However, away from this, the creativity we are really looking at is in the core skill sets of the team. And in this point, there is never a good time to stop learning (we are in the business after all), so AI start-ups who have access to third party companies who provide training can also be a real benefit also. The areas of CUDA GPU programming and OPENACC are very important to make sure the fundamentals are there in place. But also, more advanced tailored workshops that are more specific to your project can be organised as well by certain companies.
Its also good to be aligned with partners that can support you in the minefield of data, and more importantly where to find it or to know how to get it. Data is hard to get, very hard in fact. I would say I get asked this most by start-ups, "how can I get more data"
Well its not so impossible after all, there are companies with vast networks of AI start ups they are communicating with, and they all seem to be saying the same. Enter contractual templates, basically a system where data is shared between organisations. Whilst its not as simple as it sounds, there is a lot of legal exchanges to go through to get to a stage where everyone is happy. However, when it works well, it is very effective, so much so that it is a primary recommendation in the recent independent report 'Growing the artificial intelligence industry in the UK'
So what's the next step?
Well you can contact myself or one of my team (Paul Sansbury email@example.com 02392 322 594, Luke Tiernan firstname.lastname@example.org 02392 322 794) to see how we can assist and support your projects. Its also worthwhile to note, that Novatech has special programmes in place to help support AI start-ups, like special leasing plans, discounted purchase points (similar to education) training available in Deep Learning Frameworks, or CUDA.