AutoML: Opportunities and Challenges

Image source:

My work in machine learning (ML) has shown me how readily AutoML can disrupt markets and drive economic and social transformation. AutoML is essentially automated ML. It is an ML framework which takes in raw data, runs a variety of algorithms (K-Means Clustering, PCA, time series, etc.) to find patterns in and process the data, splits data into folds, then builds hundreds to thousands of iterations of different ML models (DRF, GLM, GBM, ANN, SVM etc.) and runs automated hyperparameter searches to optimise model CV performance. Due to its automated data processing, it is indifferent to the data it intakes and can hence be run on any dataset in any field.

AutoML will have a big impact. Since it can be used anywhere, it will soon be ubiquitous. Current computing power allows it to be used quickly and effectively. As parallel processing power explodes with fast GPU improvement and quantum computing, linear algebra operations used in ML algorithms become exponentially faster; AutoML will become even more viable.

AutoML is also truly democratic and thus easy to adopt and apply. It requires no qualification or data-science expertise to use. Its users must simply input a dataset to see results. It can even analyse results independently using the LIME framework to lend insights into decision-making and important features or relationships. As data generates more value, as one can see with Google and Facebook’s market domination, AutoML can take over data analytics and solution implementation for a fraction of the cost of hiring a data scientist, particularly during this time of shortage.

We are already seeing value generation. In business, AutoML can play a key role in market analysis and customer segmentation. It is already being used to create disruptive products. In financial services, I am seeing how more people can access credit and insurance coverage as ML fintech models can offer more efficient rates than underwriters to more people. There lies significant potential in AutoML use by governments for efficient macro-resource allocation as they generate/access more data. Integrating AutoML delivers higher quality products at lower cost and risk. Its adoption can thus massive economic and social change.

Its primary advantage is its ability to analyse large datasets and make better decisions than a human could in a much faster time. If humans attempt to compete on both accuracy and time, intuition will play a part as analysing millions of rows of numbers in an hour or so is impossible. Consequently, assumptions will be made, and bias will creep in. Hence, a substantial advantage of AutoML is to make unbiased decisions.

Alas, bias finds a way. A dataset is AutoML’s whole universe; it knows nothing besides that, so bias creeps in from the software developer and environment in the initial data generation, training data selection, and final result analysis. For AutoML, bias will result in suboptimal performance. For example, if a loan-giving model working off a skewed dataset finds that people in a specific region, where on average people are 80% likely to pay off a loan, are only 50% likely to do so may charge deny a loan to someone who would have paid it off. Here, both the borrower and lender suffer due to bias, and any positive externalities of investment are lost.

Bias will also cause significant hurdles to AutoML adoption. Take, for instance, Microsoft’s Twitter chatbot Tay, which became racist due to bigotry in the crowdsourced data. As a result, the potential for crowdsourced learning and product improvement was shut down. The same happened with Amazon’s hiring bot, which made hiring more efficient and less prone to human bias, eventually became sexist, because the data it was working off taught it so. As Salim Ismail of Singularity University and EXO outlines, evolution has made us inherently more prone to react to negative stimulus, so whenever algorithms make mistakes, markets or legislative forces cause the project to shut down lest companies lose goodwill. Consequently, substantial positive potential gain is lost due to bias.

Thus, various steps can be taken to remove bias. Software engineers must have a solid ethical grounding so that they can consider the implications of their choices during model building on more than model performance. Diversity in the development environment will also create checks and balances. More perspectives will create more inclusive algorithms.

Fairness incentives need to be created for both companies and individual developers. Tech-literate legislation is required to regulate machine ethics and development. Markets and regulators should place more scrutiny on how decisions are made by machines and should punish breaches more severely.

Given that AutoML is still an emerging technology, more research must be done into the theory behind machine ethics and its implementation and better understanding how ML models make decisions/inferences and the impact thereof.

read original article at——artificial_intelligence-5