A featured contribution from Leadership Perspectives: a curated forum reserved for leaders nominated by our subscribers and vetted by the Construction Tech Review Advisory Board.
thyssenkrupp
AI For ROI - How To Choose The Right AI Solution?
.jpg "Abhinav Singhal")
Abhinav Singhal
Companies looking to adopt AI today are bombarded with technology companies and start-ups selling advanced machine learning-based solutions built on exciting use cases. However, before kickstarting newer pilots and investing in these advanced solutions it is useful to step back and reflect on the overall intent of using AI for the organization and the traditional suite of analytical techniques and resources available.
A recent study by McKinsey & Company compiling hundreds of AI use cases and applications across multiple industries found out that just in 16% of use cases advanced machine learning-based AI solutions were only applicable and traditional analytical methods were not effective. In 69% of the use, cases advanced AI methods such as deep neural networks helped in improving performance of already established methods and for remaining 15% could provide only limited benefit.
This is also consistent with our experience that the greatest potential for AI is to create value in use cases where already established analytical techniques (such as regression or classification) can be used, but where neural network techniques can generate additional insights or broaden the application base. It is a common pitfall for companies to get caught up in all the hype surrounding the advanced AI developments and miss the rationale of adopting AI in the first place. Oneway, CIOs can assess the suitability of an AI solution is it to break it down into simpler elements and ask five basic questions.
1. What is the core business problem to be solved?
Over two-third of the expected value from using AI is in either revenue generation use cases (e.g., product recommendation, customer service management, pricing & promotion) or operational improvement (e.g., predictive maintenance, yield optimization, supply chain). While consumer-led industries such as retail and high tech tend to see more potential from marketing and sales-related AI applications, manufacturing, and other heavy industries see more benefit in using AI for operational excellence. The remaining value pool is distributed across the support functions, for example, task automation, people analytics, risk assessment, etc. Given the wide range of applicability of AI techniques, it is important to pin down the source of value creation for the company and then determine the pain points or opportunities where it makes most sense to investin AI deployment.
2. Which analytical method or technique is best applicable?
Most of the business problems can be classified into few standard types and have a corresponding set of established techniques to solve them.
Some common examples (not exhaustive) include:
• Classification (e.g., categorizing products of acceptable quality) – Logistic regression, Discriminant analysis, Naïve Bayes, Support Vector Machines, CNNs
• Estimation (e.g., forecasting sales demand or predictive analytics) – Linear Regression, Feed Forward Neural Networks
• Clustering (e.g., segmenting customers or employee) – K means, Gaussian mixture, Affinity propagation
• Optimization (e.g., capacity or route optimization) – Genetic Algorithms, Differential Evolution, Markov decision process
• Recommendation (e.g., next product to buy) – Collaborative filtering, Content filtering, Hybrid
The most prevalent problem types are classification, estimation, and clustering, suggesting that developing the capabilities in associated techniques could be a good starting point. Also, in most cases, for a specific problem both the traditional techniques or advanced deep neural network-based techniques could be applicable and its important to assess the trade-off between them before selecting.
3. What types of data-sets are needed?
Data sets also play an important role while choosing between traditional and advanced techniques. They can vary from being structured or time series based to text, audio, image or video based. Typically, neural AI techniques excel at analysing image, video, and audio data types because of their complex, multi-dimensional nature compared to traditional techniques. However, sometimes even more value can be extracted from mining insights from traditional structured and time series data types rather than going for audio visual data sets. Building a deep understanding of use cases and how they are associated with particular problem types, analytical techniques, and data types can help guide companies regarding where to invest in the technical capabilities and the associated data that can provide the greatest impact.
4. What are the ‘training’ requirements?
Large labeled training data sets are required in most applications to make effective use of advanced techniques such as neural networks, and in some cases, millions of them to perform at human levels. As training data set increases, performance of traditional techniques tends to flatten and advanced AI techniques tends to increase. However, if a threshold of data volume is not reached, AI may not add value to traditional analytics techniques making it also an important consideration. While promising new ML techniques such as reinforcement learning, generative adversarial networks, one-shot learning are trying to overcome the requirements of supervised learning, which requires humans to label and categorize the underlying data. Companies still need to assess their ability to collect, integrate, govern and process data at scale before deciding on the right AI solution.
5. What is the optimum system (or hardware) configuration?
Depending on the end-use case and its requirements, data can be generated from a variety of sources including sensors, IoT devices & machinery or social networks. This could result in inmassive amount of fast, structured, semi-structured or unstructured data, time series or real-time which can be stored in original granularity or aggregated or pre-analysed. The choices are immense. It’s important to maintain focus on the ultimate requirements of the end application and prioritise accordingly. Similarly, the AI application can reside in the cloud or at the edge to minimise latency. Cloud typically offers a flexible and scalable environment at relatively low-cost and without huge initial investments. Many providers also offer APIs for computer vision, speech recognition, and natural language processing (NLP) or other cognitive domains on their cloud platforms which are pre-trained and preconfigured for a certain task and serve as gateways to AI applications. All these configuration options need to be considered before finalising the optimum system infrastructure to enable the preferred AI solution.
Ultimately, for CIOs to create value from their AI initiatives, it is important to develop an understanding about which business use cases have the most value creation potential, as well as which AI or any other analytical techniques is most suited to capture that value and can be deployed at scale across the company. The decision of choosing the right AI solution is often driven less by the sophistication of the technique but more by the available skills, capabilities, IT infrastructure and data at the end for organization-wide implementation.
• Classification (e.g., categorizing products of acceptable quality) – Logistic regression, Discriminant analysis, Naïve Bayes, Support Vector Machines, CNNs
• Estimation (e.g., forecasting sales demand or predictive analytics) – Linear Regression, Feed Forward Neural Networks
• Clustering (e.g., segmenting customers or employee) – K means, Gaussian mixture, Affinity propagation
• Optimization (e.g., capacity or route optimization) – Genetic Algorithms, Differential Evolution, Markov decision process
• Recommendation (e.g., next product to buy) – Collaborative filtering, Content filtering, Hybrid
The most prevalent problem types are classification, estimation, and clustering, suggesting that developing the capabilities in associated techniques could be a good starting point. Also, in most cases, for a specific problem both the traditional techniques or advanced deep neural network-based techniques could be applicable and its important to assess the trade-off between them before selecting.
3. What types of data-sets are needed?
Data sets also play an important role while choosing between traditional and advanced techniques. They can vary from being structured or time series based to text, audio, image or video based. Typically, neural AI techniques excel at analysing image, video, and audio data types because of their complex, multi-dimensional nature compared to traditional techniques. However, sometimes even more value can be extracted from mining insights from traditional structured and time series data types rather than going for audio visual data sets. Building a deep understanding of use cases and how they are associated with particular problem types, analytical techniques, and data types can help guide companies regarding where to invest in the technical capabilities and the associated data that can provide the greatest impact.
4. What are the ‘training’ requirements?
Large labeled training data sets are required in most applications to make effective use of advanced techniques such as neural networks, and in some cases, millions of them to perform at human levels. As training data set increases, performance of traditional techniques tends to flatten and advanced AI techniques tends to increase. However, if a threshold of data volume is not reached, AI may not add value to traditional analytics techniques making it also an important consideration. While promising new ML techniques such as reinforcement learning, generative adversarial networks, one-shot learning are trying to overcome the requirements of supervised learning, which requires humans to label and categorize the underlying data. Companies still need to assess their ability to collect, integrate, govern and process data at scale before deciding on the right AI solution.
5. What is the optimum system (or hardware) configuration?
Depending on the end-use case and its requirements, data can be generated from a variety of sources including sensors, IoT devices & machinery or social networks. This could result in inmassive amount of fast, structured, semi-structured or unstructured data, time series or real-time which can be stored in original granularity or aggregated or pre-analysed. The choices are immense. It’s important to maintain focus on the ultimate requirements of the end application and prioritise accordingly. Similarly, the AI application can reside in the cloud or at the edge to minimise latency. Cloud typically offers a flexible and scalable environment at relatively low-cost and without huge initial investments. Many providers also offer APIs for computer vision, speech recognition, and natural language processing (NLP) or other cognitive domains on their cloud platforms which are pre-trained and preconfigured for a certain task and serve as gateways to AI applications. All these configuration options need to be considered before finalising the optimum system infrastructure to enable the preferred AI solution.
Ultimately, for CIOs to create value from their AI initiatives, it is important to develop an understanding about which business use cases have the most value creation potential, as well as which AI or any other analytical techniques is most suited to capture that value and can be deployed at scale across the company. The decision of choosing the right AI solution is often driven less by the sophistication of the technique but more by the available skills, capabilities, IT infrastructure and data at the end for organization-wide implementation.
The articles from these contributors are based on their personal expertise and viewpoints, and do not necessarily reflect the opinions of their employers or affiliated organizations.
