Monday, 19 August 2019
Artificial Intelligence is arguably the single biggest technological disruption we have ever seen. Today, everyone is talking about the impact AI will have on the future of business, and specifically, object detection.
What is AI? It’s hard to define exactly, but I think the truest definition is this; AI is a collection of computer systems brought together to imitate human behavior in some way. But even more than this, artificial intelligence has the capacity to outperform humans on an increasingly wide variety of complex tasks.
Machine Learning is one of the most exciting branches of AI. Significant advances in this space over the last few years have made this technology more accessible than ever, with real and robust solutions offering business an advantage unlike we’ve seen before.
Recent times has seen Think180 explore the implications of AI to create better solutions for our clients. Having significant experience in mathematics and computer science meant I took responsibility for our explorations. Machine learning however is such a broad discipline, there is always more to learn – this is part of the challenge and the enjoyment! Using Python, OpenCV and TensorFlow, I delved into the intriguing technology of machine learning, starting with the increasingly popular methodology of deep learning.
How does deep learning work? It uses a connected web of artificial ‘neurons’, inspired by the neural network of the human brain, to enable machines to be ’trained’ as opposed to the traditional approach of being ‘instructed’ programmatically. I decided to experiment using deep learning for object identification.
I started with cats and dogs. After downloading around 25,000 images of cats and dogs I passed them through a simple model (neural network) several times. Could my machine identify a cat from a dog? Turns out yes. With surprisingly high accuracy of 80%. I expect with some more training and tweaking of the model this number could be pushed higher.
I moved on to handwritten numbers. Training made several passes on around 10,000 images. The accuracy of this model was 99% – far higher than the cat vs dog model, the nature of the images being trained on has a significant bearing on its accuracy.
Excited by these results, I moved onto something bigger. Object detection could be used for a practical purpose; to identify street signs. For this exercise I trained a more complex model on around 280 images of 6 different classes of signs. Training consisted of over 4000 passes on the selected images and took several hours to complete. To capture some testing data I drove down the road at 40km an hour with my phone out the window, capturing on video 12 various street signs including permit zones, parking signs and speed signs. The model identified every sign in the video correctly bar one – the missed sign was most likely due to a change in lighting or angle.
So where to from here? Well, we now have proof of concept that machine learning can be used for object detection. In the case of street signs, it’s a matter of efficiency. Instead of manually recording what street signs are where, councils can potentially use machine learning to do that for them. Not only identifying what type of sign it is, but its exact GPS coordinates.
Imagine how else this technology might be applied. AI research is leading to breakthroughs in image recognition, speech recognition, vehicle automation and so much more. It is increasingly being used in market research and sentiment analysis with some of the most exciting implications around its use in healthcare. Whilst it is not an exact science, AI really is a technological revolution.
Think180 is enthusiastic to continue exploration of machine learning, implementing it into our service offerings to come up with even easier solutions for our clients. If you are looking to move your business to explore Intelligent Technologies including AI, reach out to us here.
‘Head of AI’