Most machine learning algorithms and artificial intelligence (ML/AI) libraries have been around for over a decade. But easy access to ML/AI libraries and large training datasets is new. The tech industry has come a long way since we worked on our first AI project for a client in 2009. Back then, we were a small startup team of <10 people. Even though it’s only the beginning of 2018, Oursky has seen a doubling of inquiries about ML/AI and I want to take time to reflect on how AI has picked up in the past few years. The accessible libraries free up entrepreneurial developers to think more about business problems and applications to create new AI-powered products.
How we Got into AI
Oursky’s first machine learning contract was to develop character recognition software to streamline identity verification back in 2009. We experimented with neural networks, decision trees, and some OCR libraries before we ultimately went with neural network solutions for accuracy. Back then, we had to implement most of the algorithm from scratch since it was not really common.
The machine learning techniques we experimented with have been around for decades, but the tech industry was constrained by computing resources. Some types of machine learning, such as convolutional neural networks (https://en.wikipedia.org/wiki/Convolutional_neural_network) have been applied since the 80s. For example, LeNet-5, developed in 1998, helped some banks recognize hand-written numbers on checks that were digitized as 32×32 pixel images. To work on more complex classification tasks, more and more convolutional layers are needed, so this technique was restricted by the availability of processing power at the time.
That’s all changed in the past couple of years. The hardware has caught up, and it’s opened new opportunities for investment and commercialization.
Why is AI Picking up now?
To put the potential of AI into perspective, investment in AI/ML in 2017 alone had already overtaken all of the investment in big data combined. According to the Financier, investment will reach $37 billion by 2025, with China already leading the way as the second largest AI ecosystem after the US.
AI is picking up because the technology ecosystem has come together:
- availability of opensource machine learning libraries like CNTK by Microsoft, Theano, or Google’s TensorFlow
- increased processing power (servers and consumer-grade processors) that helps commercialization
- cloud infrastructure is now common for storing large quantities of business data (potential to train models)
Companies are approaching AI differently from 5 years ago
When we first started ~10 years ago, only a handful of clients dared to try AI technologies. These clients usually had specific problems they wanted to solve from generating chords from music files to identifying various zipper pullers styles from an image to match with inventory in a database. These days Oursky has been getting more enquiries from IT or business departments in large companies that want to use AI, but do not have a clear picture of how types of AI can be applied to their existing applications.
Many people also realize what AI can do, but are not sure about how to apply it to their business. While Oursky is not specialized in business consulting, we are now in a position to recommend AI solutions that are sometimes more cost-effective for business operations in the long-term for clients. For example, businesses can think about target areas where AI can help internal business operations (such as sorting data or fraud detection) or improve front-line customer engagement (such as real-time product matching based on photos).
Developers should experiment with new AI libraries to generate business ideas
Another thing is that machine learning research is ongoing. The extent of applications for neural networks was simply not yet fully understood 10 years ago. The generative adversarial network, which produces great visual images, is now having wider applications when adapted for 3D model generation and FinTech, and the possibilities for other applications seem endless. Transfer learning, reapplying a trained AI model to a new context, has also become more widely understood and is being adopted for applications such as Google AutoML.
Now, there are also many more AI related products, usually in the form of APIs and libraries, available for common applications so that developers don’t even need to train a neural network for a specific purpose. An example is an app we built to help clients record their team’s emotions, or identify objects from images that leveraged robust public APIs and pre-trained networks. Our team tested facial detection for our company’s previously developed IoT door lock in 48-hours by leveraging opensource libraries. Only 2 colleagues worked on that Christmas hackathon project. That’s how accessible ML/AI has become.
Business applications need to understand what types of AI are available
It’s important to understand the different parts of AI before thinking about its applications. Virtual assistants, for example, have several different AI components in order to provide a good service. Virtual assistants need AI components to convert voice to text, process the text as natural language processing (NLP), perform an action if required, and convert from text back to speech. It may also have predictive abilities, which also count as machine learning.
Many AI components have already been incorporated into businesses. Everything from targeted ads on social media platform, item and price recommendations for sellers listing items on Amazon, to stickers on Snapchat and Instagram are AI-powered features. Some exciting things are also happening on area like algorithmic trading based on sentiment with neural networks, medical imaging like the Tencent-backed Voxelcloud, industrial robotics, Baidu’s Apollo for driverless cars, or a chatbot for banks in Singapore.
Using machine learning no longer requires hiring a large team of developers. Any developer with a buddy who has worked with AI can make use of opensource libraries to create new products and start experimenting (of course, we encourage open-sourcing).
We’re excited to work on and grow with the global ML/AI development community. Feel free to ping Oursky on Twitter if you’ve got a great library to share! We also look forward to sharing the new AI-powered products we’re working on this year, so stay tuned to our blog! If you have any ideas you’d like to run by us, feel free to get in touch!