AIBusiness recently interviewed a key developer in Artificial Intelligence, Alex Dalyac, CEO and Founder of Tractable. Tractable are pioneering Deep Learning software developers, focused on fully automated visual recognition with AI.
Alex will be speaking at The AI Summit in London on 5 May, where he will demonstrate how AI technology can learn to interpret imagery as domain experts do, looking at the commercial impact of such AI applications on select use cases.
AIBusiness spoke to Alex to find out his views on AI’s impact on business overall, and hear more about his vision for Tractable ahead of his keynote at The AI Summit.
Alex Dalyac of Tractable
1. How do you believe AI will impact business overall and in what ways?
By enabling computers to perform human tasks, AI will extend the reach of software into new applications, new industrial domains. Service industries like insurance, logistics and retail banking will increasingly become software industries. This will enable these services, which our society relies on, to be performed smoothly, at a fraction of the cost it does today, making them more affordable and reliable.
This expansion will occur with AI software performing tasks too complex and subtle for traditional software to handle, but narrow enough for an AI algorithm to perform. For example, driving a car consists in a sequence of restricted actions: accelerate, decelerate, turn. But driving properly requires careful interpretation of the environment and objects surrounding the car. It’s both a narrow and subtle task that is a good candidate for AI. Interpreting radiology scans is even more narrow: the environment is the scan, the action is the diagnosis. Waitressing, on the other hand, faces a far more complex environment from an AI point of view: a busy cafe has plenty of objects and entities, some of which are slippery, others which move, others that need to be talked to, even ‘pleased’. Tasks involving complex environments are likely to remain human-dominated for much longer,
especially if they require human charm.
But subtle, narrow tasks will become dominated by AI, especially when the scale, speed or accuracy gains from switching to AI are high. Some of the most exciting progress will be when these gains will make a task possible for the first time: for example, being able to analyse so much video footage and react so quickly that you can intervene to prevent adverse outcomes. An extreme example would be the recent terrorist attack at the Brussels airport, whose suspects were filmed on camera in the airport before the detonation.
2. Where are we at the moment in terms of ready-to-implement technology versus wishful thinking?
AI technology can be broken into 4 categories: ready-to-implement old AI, ready-to-implement novel AI, early-stage rapidly-improving novel AI, and moonshot AI.
Old AI concerns tasks that were solved in academia 5+ years ago and applied to industry numerous times, such as predictive analytics. Predictive analytics is a business term so its exact mapping to academic tasks is not clearly defined. If we take the definition as models making predictions from structured data, examples include collaborative filtering: recommending products to users based on similarities between products and similarities between users. Another example of old AI is optical character recognition.
Novel AI concerns tasks that were solved in academia in the past 5 years, which are starting to establish themselves in business applications. Examples include image classification and speech recognition. Novel AI also concerns tasks in which academia is making rapid progress, which we expect to be solved within the next few years. Natural language processing is a notable example. Deep learning is the core technology behind these three examples, and is the technology that we specialise in at Tractable. Another example is autonomous driving, which combines a number of old and novel technologies.
Moonshot AI concerns tasks that academia is still hard at work trying to solve, where no breakthrough has yet been attained. A notable example in biology and genetics is predicting phenotype (how the organism behaves) from genotype (how the organism is programmed) and environment (the setting and other organisms which the organism interacts with).
An important distinction to make within ready-to-implement AI is between AI that requires extra training data from the target task in order to perform adequately, and AI that does not require it. The latter typically occurs when the AI has already received sufficient training data from the target task. At Tractable, we specialise in getting AI that requires extra training data to rapidly adapt to the target task at minimal cost.
All of the examples above belong to the branch of AI referred to as machine learning, where the AI figures out how to solve a task by extrapolating from examples of the task. Another, older branch of AI is symbolic AI, commonly referred to as rule-based AI, a product of which was used to beat chess grand master Gary Kasparov in 1997.
3. What do you think are the main challenges in adopting AI technologies, from machine learning through to image recognition, in business?
As mentioned above, machine learning algorithms that require extra training data from the target task are harder to insert into business workflows, because of the overhead cost of feeding sufficient volumes of training data. This data needs to exist, be operationally and legally accessible, and also needs to be sufficiently clean and curated. At Tractable, we specialise in reducing this overhead cost.
4. Which industries do you believe will be the pioneers in broadly adopting AI technologies?
I expect pioneer industries to be those that score highest on the following three criteria:
– Industries that involve complex yet narrow tasks of great importance
– Tasks that face high scale, speed or accuracy gains
– Industries that produce large quantities of easily accessible training data
Within computer vision, which is one of our specialist fields, I believe these industries will be media, e-commerce, medical imaging and surveillance.
5. What is the key proposition of Tractable in developing an AI-powered business?
Our key proposition is rapidly trainable computer vision for expert tasks. We believe that the greatest impact computer vision can have is not in recognising a large number of simple categories such as cats, tables and cars, but in recognising specific categories that only experts can detect. This requires training on highly specific, challenging images, which our proprietary technology can currently achieve up to 500 times faster. Essentially, we make it economically and practically feasible to implement computer vision AI into business workflows.
6. What new products & solutions can we expect from Tractable in the immediate future?
The cases we are currently focusing on are repair estimating in property & casualty insurance claims, anomaly detection in surveillance cameras, visual inspection in industry, and exploration in oil & gas and mining.
At The AI Summit, Alex Dalyac will be delivering his in-depth keynote, Identifying high value computer vision use cases and tackling the labelled data issue for expert tasks.
The AI Summit is the world’s first event dedicated to Artificial Intelligence for the business world. For more information, and to join us on 5 May at the Four Seasons Hotel, London, visit: theaisummit.com