June 30, 2020

How Tiny AI Can Secure the Future of Artificial Intelligence

Importance of Tiny AIImportance of Tiny AI – The most common kind of algorithmic problem is ‘green computing’ which basically involves systems that directly manage temperature, energy and power as an important computational resource. Green technologies are known to have unique non-energy related properties that can easily trump previous technologies.

All of this points to one fundamental question: Computing algorithms are getting smarter on a daily basis, but are they getting any greener?

Apparently, not for the most part – and this is a significant problem in the technological atmosphere. Due to this problem, researchers can be seen working hard to look for new ways to develop smaller algorithms and we’re going to see how Tiny AI is the solution.

The Hidden Environmental Cost of AI

AI has had many breakthroughs over the last couple of years – such as Deep Learning. This breakthrough has the ability to power a number of AI systems that could deliver with high accuracies. With the help of Deep Learning, algorithms have the ability to scan a medical image, identify a tumor or to translate literature from hundreds of languages or even to navigate vehicles through complex traffic patterns.

There’s no doubt about it – AI is getting accurate, but there is a hidden environmental cost of this accuracy. Researchers from the University of Massachusetts Amherst resealed a study that helped shine some light on how energy intensive it can be to train an algorithm. According to this study, training a single algorithm could consume upto 5 times the lifetime of all the carbon emissions out of an average car. This was also equivalent to around 300 round-trip flights in between San Francisco and New York.

It seems as if the AI race may have lost track on the much needed focus on energy efficiency. A research scientist from the Allen Institute, Roy Schwatz, also suggested in his paper titled Green A.I. that artificial intelligence researchers should work in a manner that they make energy efficiency a strong evaluation criteria for algorithms – alongside accuracy and other typical measures. Similarly, another article quoted Schwartz to have said that the world doesn’t want to reach a state where artificial intelligence will have become the biggest contributor to global warming.

This is exactly where Tiny AI comes in.

What is Tiny AI?

Tiny Ai is basically an effort of the AI community to reduce the overall sizes of their algorithms – especially in the case of those that require huge computational power and datasets to work. Researchers of Tiny AI are developing methods that will not only decrease the size of AI models but they will also accelerate inference while maintaining consistently high levels of accuracy. These methods are  known as distillation methods and they can be used to scale down models by factor that may be as high as 10 times.

With a much smaller algorithm, users will not have to send over data to the cloud. Instead, these programs could easily be deployed on the device itself. Take BERT for instance .BERT was created by Google as a pre-trained language model – or simply put – it helped users write. While this may seem simple enough, BERT has the ability to understand your written words in terms of their context – and not just via programmed grammar and vocabulary. Due to this, BERT could make better writing suggestions to help users finish their sentences.

Now while BERT is a very helpful model, it is also very large. When MIT reviewed this technology, they found that BERT consisted of 340 million parameters of data! Other than this, one session of training BERT would consume as much energy as would be sufficient for an average American household for 50 days!

This is also why BERT became a great target for Tiny AI researchers.

BERT becomes TinyBert

In the most recent attempt, Huawei researchers had claimed that they were successful in reducing BERT 7.5 times while also improving its overall speed 9.4 times. Ironically, they named this new model TinyBERT but more seriously, how much better was this new version? Well, the original authors claimed that TinyBert was able to achieve as much as 96% of the original’s performance.

This example shows you how as technology evolves, Tiny AI will play a significant role in moderating its environmental footprint. Even existing services such as cameras, voice assistants and the likes will not have to constantly transfer data to the cloud if they are treated with Tiny AI.

The Importance of Tiny AI

Simply put, Tiny AI can make it possible for the tech community to deploy any complex algorithm from an edge device. For instance any layman could conduct medical image analysis using their smartphones. They may even be able to partake in autonomous driving without the help of a cloud. What’s more, with so many of these possibilities being limited to your average edge devices, users will also be able to improve on data security and privacy.

Now this doesn’t necessarily mean that cloud centers will become outdated. Instead, they will be used for very high-performing computing algorithms such as for DNA analysis. To do so, cloud systems will have to deal with huge amounts of data in a matter of hours. Again, Tiny AI will be able to help here by making hyper-efficient AI systems. Here’s  how:

Tiny Data

In other words – smarter data usage. This would involve data reduction techniques with the help of surrogate modeling. Other than this, data processing can be assisted by AI while compression strategies are deployed (such as network pruning).

Tiny Hardware

Due to the technological advances in nanotechnology, Tiny AI could help produce new architectures, new materials and new structures with the help of 3D integrated systems.

Tiny Algorithms

All of the above could be done with the help of energy-efficient processing for edge or extreme edge devices. AI algorithms could easily be delivered ‘on-chip’ to make technological endpoints meet.

The possibilities are limitless – and more importantly – they all adhere to a greener future.

Final Word

It is no longer feasible for your sales reps to wait a month to find out what they are earning and how they are achieving compared to their goals. With the right sales performance management solutions, you can elevate your sales performance management in 2020 and beyond.

by Bobby J Davidson

We love our company and we love what we do.  Check out the ‘Why Percento‘ page to learn more: Love of Technology and Business!  As the President of Percento Technologies International, I provide day-to-day leadership to the company’s senior management and I am personally involved in the strategy, business development and sales activities of the firm.

The company was founded in 1999 with the purpose of providing a one call source for organizations in need of Enterprise IT Consulting and Management.  We also provide a line of products in the boutique Cloud Server space with a touch of high-end website strategy consulting and design services.   We personalizes the IT Service experience with a team approach, working with clients from diverse sectors of industry, including energy services, financial, legal, entertainment, healthcare, hospitality, retail and general and/or corporate business.  percentotech.com/contact