by Dmitris Vlitas, Head of AI at DPL
Today, we often come across statements like ”Artificial Intelligence (AI) is the new electricity”.
This statement draws an analogy between AI and the time, one hundred years ago, when electricity was invented and people were excited by, although not completely clear about, the possibilities it provided for both society and business. The rise of electricity transformed every major industry. It is very difficult to imagine households, communications, health care and practically every single domain of our life now without it. In very recent years AI has advanced to a point where we see a surprisingly clear path transforming every major industry.
Everyone is very familiar with lucrative applications of AI such as speech recognition, machine translation or self driving cars. The current research on AI focuses on the next steps of these applications. For instance AI will produce well trained models that will read and understand large volumes of text. They will be able to read, summarise and answer questions accurately, comparably to human level performance. As a consequence every company is faced with the challenge to stay agile as it navigates the new normal.
AI is the theory and development of computer systems able to perform tasks previously thought to require human intelligence. It boils down to a software consisting of a family of algorithms which helps machines to learn. Machine Learning (ML), is the procedure of providing our machines with large amount of data and letting them learn by themselves. Inspired by the way that we, humans learn, ML algorithms repeatedly process available data, so that they can make accurate predictions or classifications.
Traditionally every company has a development team, which receives some input data set with the corresponding output results. The team analyses the data, designs algorithms based on a prefixed set of rules and then tests the outcome of this model against the given outputs. In the new era of ML, we create algorithms that are able to learn the rules that govern the relations between the given inputs and outputs (labels). The ML algorithms are not given any directions or constraints . They are fed the available data and then independently develop the ability using Mathematics, to detect patterns and make the best possible approximation of the mapping of input-output data.
Therefore without human aid or interference, machines are able to quickly and efficiently handle vast amounts of data and learn how to map inputs to outputs. Here is a typical example of a ML model:
Take a data set that contains labeled images of either lemons or apples, a ML algorithm can be trained on these data to be able to distinguish a picture of a lemon from that of an apple. In the process of training, which has no human input the algorithm learns to look mainly for round-red fruits, in the case of apples. The algorithm itself realises that the shape and colour are the two very important features for the classification task that has to be performed.
In many tasks the performance of a trained ML model surpasses that of human performance levels. AI drives incredible amounts of economic value. Almost all of the economic benefit created by AI today is through the above learning process based on input-output data.
Everyday examples where this occurs include loan approvals and online advertising. In the first case ML algorithms get fed with data describing client characteristics and outputs, informing whether clients were able to pay their debts or not. With such data sets, the machine learns from the financial history and becomes an expert in evaluating new loan applications.
In online advertising, ML algorithms are provided with records of consumers habits, so they can easily identify customers with similar preferences and make appropriate suggestions. Amazon's recommendation system is probably the most recognisable application of this. What this approach does is to allow companies to build "intimate" relations with their customers by better understanding them through the effective analysis of data.
Typical success stories of using ML are Man Group and Bank of America. Man Group Plc. built a system that evolved autonomously, finding money making strategies humans had missed. The results were startlingly good. By 2015 artificial intelligence was contributing roughly half of the profits in one of Mans biggest funds.
Consumers today are accustomed to seamless mobile experiences provided by apps like Uber and Airbnb and want better banking experiences. To this end Bank of America introduced Erica, the new digital assistant. Erica uses artificial intelligence, predictive analytics and cognitive messaging to help customers to make payments, check balances, save money and pay down debt. As systems like Erica integrate deeply into our lives, we should expect the rise of fully synchronised voice banking, payments and commerce.
There are three important factors that are responsible for today's blossoming of AI.
First it is the availability of data. Daily we generate somewhere in the region of 2.5 quintillion bytes of data and we are continuing to improve the ways of capturing this data effectively.
Secondly, our computational power has dramatically increased over the last decade. Graphic processing unit (GPU) together with central processing unit (CPU) allow for parallel processing that greatly accelerates computing.
Finally, algorithmic innovation has never been so rapid and it now allows us to be able to process vast amounts of data with great scalability.
These three factors permit us to leverage data for effective machine learning. Here lies one of the main causes responsible for the outstanding performance of ML algorithms. They can learn from massive amounts of data, recognise patterns, combinations of patterns, in a speed much greater than us humans. This is the main reason why ML models can achieve results that surpass human level performance on certain tasks.
Besides the supervised methods of learning (input-output data), we have succeeded in transferring learning between domains. For instance, back to the data set containing labelled images of either lemons or apples, we may have algorithms that have been trained to classify images of those two fruits.
This trained algorithm can then be utilised in other domains where the amount of data available is significantly less. So the algorithm trained to identify lemons and apples can easily be used to to distinguish x-rays images of adults from those of children.
Even though today we are beginning to experience an AI shared economy with open resources and software available to everyone, it is the data that builds defensive business. With the techniques necessary to train a ML model becoming widely accessible the limiting factor is the data, or lack thereof in specific business cases. It is critical for companies to start to build these data assets. Once they have begun to build these assets, through use of their products, visitors to their sites, interactions with their customers, for example, the data will grow and this will enable better ML modelling. Making it far more useful.
That said, it is worth noting that just the use of ML algorithms does not guarantee a company's best use of the new technology. Below are some thoughts on what the key steps should be to begin to take real advantage of AI in your business.
Firstly, it is critical for businesses to remember that the very first prerequisite is strategic data acquisition. It is essential to get the right data in the right format.
This should then be followed by establishing a unified data policy and ensure that this is followed. This will enable every data science team and engineer to have access to the right data.
Build a centralised AI team, which communicates with all the other units in the company. Scientists from the AI team then can address the individual needs of each unit.
Remember, AI is not a finished product. It is an autonomous self learning process that has to be customised to the organisation that makes use of it. It needs Data Scientists who can figure out the necessary input-output data, fine tune the ML algorithms and eventually produce well trained models. It is necessary for a company to be alert and understand the abilities and limitations of AI.Then it should decide how AI could be of service to its needs.
That said, in this process certain products of pre-trained models can be helpful. The company has to have a clear vision and the capacity to navigate its own route in the AI ecosystem. Your business can absolutely learn from AI companies in the market, but make sure you are not burdened by them. Each business has unique characteristics and the AI needs to be flexible and agile to meet the ever evolving needs.