Advertisment

3 Key Preparation and Precautions for AI Deployment in 2019

AI Deployment 2019: With the increased global focus on digital transformation, digital operations and interactions have become an integral part of our

author-image
CIOL Bureau
New Update
Key Differences between AI and Machine Learning - AI Deployment 2019: AI Can’t Do for Your Business

By Tarun Dua, MD, E2E Networks

Advertisment

AI Deployment - With the increased global focus on digital transformation, digital operations and interactions have become an integral part of our modern high-tech lives. The last phase of technology adoption was primarily based on rule engines and rule-based systems which delivered tremendous business value by eliminating human error. Increasingly, the edge of these systems was staffed to handle egregious exceptions that couldn’t be handled by even complex rule-based systems. The rule-based systems collected a lot of data which was fed to domain-specific business analytics software to help the business analysts and management to gain valuable insights to drive the business forward. These rule-based systems became even more meaningful to organizations when they could be modified by domain experts using either domain-specific languages or the user interfaces to add additional rules for both operations as well as generation of new insights.

Rules+Data == Answers

AI is the radical new approach compared to the rule-based systems in mimicking the learning ability of domain experts via highly parallel processing machines and AI model generation software like tensorflow, caffe2 etc. to add to newer rules based on analysis of datasets generated in the past.

Advertisment

Data+Answers == Rules

AI software that deals with business data has primarily involved dealing with structured text inputs from Databases as well un-structured text generated by humans to learn and generate models with a feedback loop to make them learn more based on additional data.

AI-based machine learning in voice/images and video is opening up new avenues for digital transformation of industries and society in ways we can’t fully imagine today. For instance, restaurants/e-commerce stores can receive orders via natural voice conversations, handle support queries or returns/complaints. Toll Booths, Parking buildings can use image recognition to directly charge fees from the identified vehicle owners. Municipalities can clean up cities by having crowd-sourced images on social media with geo-tagging to identify garbage dumps or identifying crimes against air quality hotspots in a city. AI-based facial/object scanning can improve security and queues at airports and secure installations move faster.

Advertisment

At present, 15% of the enterprises have already adopted the AI in their business operations and this number is slated to rise up to 45% in the next 12 months (Source: Adobe). Gartner further predicts that artificial intelligence will soon merge with our physical world and provide highly customized solutions through digital devices and platforms.

Preparations for AI introduction

Initial goals for the introduction of AI into existing business processes should be focused on developing a constantly iterative Machine Learning model which can deliver new capabilities or increase speed and efficiency of existing business processes.

Advertisment

Setting up data pipelines for secure transfer of data to Deep Learning systems might involve a high-level understanding of laws, agreements with customers to be analyzed and understood by implementation teams internal or external. For instance, a Voice Recognition model being implemented for an Indian Call Center ( with Indian customers) to work with plain old telephone systems ( POTS ) would either require massive local technology investments or a complete re-invention of data pipelines to operate out of the cloud without falling foul of outdated regulations. Similarly, any organization would need to ensure that data pipelines leveraging the AI infrastructure capabilities delivered via the cloud do not result in violations of governance standards or customer agreements or laws applicable to any organization.

Artificial Intelligence demands cloud infrastructure based on massively parallel processing cores that happen to be conveniently something that GPUs were already doing so they were rapidly co-opted into AI deep learning and model training. Implementation must be prepared for feeding both historical datasets and newly generated datasets to be in a constant feedback loop to train the AI models using deep learning algorithms that fast using massively parallel compute pipelines based on GPUs or newer hardware processors.

Precautions for AI implementation

Advertisment

AI has the capability to deliver beyond human capabilities for every industry and all business sectors. However, every advanced technology has its own limitations. In a research conducted by International Data Corporation (IDC) in 2017, in the U.S. and Canada based IT and data firms, it was noticed that organizations who successfully deployed AI-powered applications and platforms have faced some common issues of poor data quality and management issues due to the handling of massive amounts of data.

A) Handling Bias

AI-based learning models rely on the data feed. The human biases like sub-conscious profiling and pre-judgment routinely cause the deep learning to take on the biases as a result of being trained on datasets being generated from imperfect humans. An AI chatbot from a major technology company had to be taken off twitter because it learned some of the worst aspects of humanity and started to behave like one in the recent past.

Advertisment

B) Dealing with larger volumes

Largely rule-based systems have operated by delegating edge cases to human staff and a lot of oversight by management teams. AI-based systems are designed to process a lot more volume of data because of their ability to handle edge cases compared to rule-based systems. So AI-based systems would require a measure of accuracy and appropriate staffing to handle edge cases and bursts of bad rules causing backlogs (requiring rollbacks on models generated)

C) Having a heart

Rule-based systems were criticized for not having a heart. Banking, Telecom, Healthcare, Insurance, Government and many other sectors deal with life, death and scenarios involving the pursuit of happiness, with newer AI trained models based on machine learning presents an opportunity to have a heart. An ad re-targeting company was roundly criticized for continuing to show ads for baby products to a mother who lost her unborn child. AI specialists must take precautions to not repeat the mistakes of older rule-based systems by using the new paradigms.

ai e2e-networks