Accuracy Matters

We’re currently at the peak of the AI hype cycle. Every other day sensational articles are written about the demise of humanity due to the rise of super powered AI.

“Alpha Go has beaten the best human Go player! Alpha Go has now learnt how to play against itself! It now trashes the old version of itself!”

“Facebook’s chat-bots have created their own language! They are conspiring with each other so Facebook had to shut them down!”

“The age of self driving cars is upon us! What about all the truck drivers?”

Dig only a little deeper and you’ll realise how sensationalised these article are. Alpha Go while impressive, is still only tackling a very narrow and constrained problem. Facebook’s chat-bots aren’t really talking to each other. Self driving cars are more than likely further away in the future than we like to think.

Don’t get us wrong - AI is showing great promise and we’re investing heavily in it.

Unfortunately, solving real world problems with AI is not as simple as creating a world beating Go bot. The difference is, that the cost of making mistakes in a game like Go is essentially zero, while the cost of making mistakes in the real world could be really expensive or even extend beyond money. Losing a million games of Go is of no consequence at all but a self driving car making a single mistake could be catastrophic.

The accuracy of AI decisions therefore matters a great deal in solving real world problems. Security is no different and that is why it is our sole focus. To illustrate this, we performed an accuracy benchmark comparing Image Intelligence’s person detection accuracy against various other computer vision platforms. Our accuracy results were best in class. Out of 8,000 random images, we correctly predicted whether the image had a person or not 90% of the time. Google Cloud Vision on the other hand, was only correct about 55% of the time. There is no real world use for an AI that is only correct half the time.

Our commitment to accuracy

Image Intelligence understands the importance of accuracy and are committed to producing the highest accuracy for our customers through these measures:

High experiment velocity

Our models are constantly being trained, tested and deployed to our customers on a weekly basis. Our machine learning pipeline is also constantly streamlined to enable our data scientists to experiment quickly. Quicker experiments leads to quicker learnings, which produces more accurate models sooner.

Frequent testing

Our models are frequently tested with fresh samples of data to detect models that are overfitting or getting stale.

Data collection

Data annotation - the process of labelling each image with descriptions such as person, car, etc - is a tedious and expensive (but necessary) process needed for training deep learning models. The quality of this data is the biggest factor in producing a model with high accuracy. Our team triple checks all data points before using it in our training process.

Guaranteed Accuracy

There are situations where a very high accuracy is not only required, but also depended on. For these applications, Image Intelligence provides Human in the loop (HITL) verification. The addition of human verification, along with all the accuracy measures taken above, provides unparalleled accuracy that you can trust. In fact, you don’t need to blindly take our word for it because HITL verification includes Guaranteed Accuracy - a generous payback scheme should you find mistakes in our results.

This system works in seconds and it is invisible to you. Talk to us for more information about Guaranteed Accuracy.