We hear a lot about Machine Learning. However, the case studies of Machine-Learning-in-action can seem slightly esoteric.
As an example, a few years ago DeepMind’s AlphaGo was the first computer program to beat a human player at the game Go and accomplished this feat using Machine Learning1. More recently, tools like GPT3 from OpenAI can take virtually any text and continue the text in a similar pattern - also based on Machine Learning.2
For data scientists and technologists, these examples demonstrate remarkable advances in Machine Learning. But in ‘corporate world’ it can be hard to see the relevance.
The State of ML
So, how ready is ML for corporate adoption? The short answer: fully ready and most-if-not-all companies should be looking for opportunities to apply ML to their businesses.
Algorithms versus Applications
For many years, Machine Learning discussions were primarily about ML Algorithms. These have complex-sounding names like K-Nearest Neighbors, Random Cut Forest and Latent Dirichlet Allocation. And the complexity is real: picking the right algorithm and implementing it correctly - such as training the model and defining the ‘hyper-parameters’ - continues to be the domain of ML specialists and data scientists.
What’s changed more recently is the explosion in the number of ML Applications. These are ML algorithms applied to solve specific problems. As opposed to choosing and then implementing the right algorithm, these applications provide a fully managed service for the problem area.
Some examples of ML Applications from Amazon Web Services (AWS):
- ‘Translate’ provides text translation from / to over 60 languages,
- ‘Transcribe’ converts audio into text for over 30 languages,
- ‘Forecast’ provides time-series forecasting based on historical data, and
- ‘Rekognition’ enables easy image and video analysis
In each case these applications make use of underlying ML algorithms however they abstract-away a lot of the complexity of managing ML models.
A recent example: My team were interested in looking at customer - contact centre interactions through chat. Could we provide realtime feedback to contact centre agents on the sentiment of chat messages? And could we highlight whether agent responses were seen as having a positive, neutral or negative sentiment?
In less than a week a team of 3 people had implemented a proof-of-concept stitching together four AWS services. AWS ‘Comprehend’ was at the heart of the solution providing instant sentiment analysis.
Engineering versus Data Science
Deploying ML Applications as opposed to ML Algorithms changes the skillsets required. As opposed to needing data scientists and ML specialists, implementing ML Applications requires engineering talent. There’s still complexity but this is ‘data-wrangling’ complexity as opposed to algorithm complexity3.
In other words: there’s still work to be done, particularly in manipulating the training and input data, but it requires less Machine Learning specialist skills and knowledge.
Corporate Adoption of ML: Examples
Some practical examples of companies adopting ML:
- DeepMind, the same company that won at Go, applied its ML approach to Google’s data centres. The result was a reduction in their data centre cooling bill of 40%4.
- Tabnine, a software company, uses a version of OpenAI’s GPT2 tool to make software development easier by auto-completing code5.
- HSBC, a retail bank, created a more accurate forecast of the cash needed in its 1,200 ATMs in Hong Kong. Previously the bank manually created forecasts which could result in ATMs running out of cash. The new ML-based tool uses live ATM data and factors-in seasonality, holidays, public events, and recent withdrawal trends to calculate how much money is needed and where. Better forecasts result in 15% fewer trips to refill ATMs and a saving of $1M a year6.
Corporate Adoption of ML At Scale
In the 1990s, ‘Six Sigma’ was a popular set of techniques to improve quality - particularly for manufacturing plants. ‘Black Belts’ were people trained-up in the Six Sigma approach and applied Six Sigma tools to specific projects, supported by ‘Green Belts’. ‘Master Black Belts’ and ‘Champions’ were responsible for applying the approach across the whole organisation.
Machine Learning is ready for a similar approach: companies need groups of people trained-up to spot business problems that can be readily improved using ML techniques. These people need to be supported by teams, including engineers, that can implement the relevant ML applications. And company leadership should be reporting on the number of ‘use-cases’ found for ML - in the same way companies report on ‘re-engineering’ costs saves.
With recent developments of ML Applications, Machine Learning is firmly out of the laboratory and is ready for the prime time.
For companies this means there are big opportunities for business improvement through adopting ML. Corporate world needs to find a scale approach that enables ML opportunities to be systematically identified and grasped.
I’ve consistently used ‘Machine Learning’ and not ‘Artificial Intelligence’ in this discussion. ML is a subset of the broader domain of AI and - in my view - ML now has tools that are ready for wide adoption for many companies.