Model Operations for Secure and Reliable AI
Artificial intelligence and Machine Learning are expressing incredible potential in various application fields; however, very few companies engaged in a 4.0 transition path can successfully implement these technologies in business processes. What needs to be done to make such applications profitable?
Why is AI Important?
Artificial Intelligence represents a set of studies and techniques, typical of information technology but with significant philosophical and social implications, which has as its purpose the realization of programs and technological systems capable of solving problems and carrying out tasks normally attributable to the mind and human capabilities. Given recent progress, it is possible to identify Artificial Intelligence as the discipline that deals with creating machines (hardware and software) capable of operating autonomously.
The growing attention created in this discipline is motivated by the results that can be achieved thanks to the technological maturity achieved, both in the computational calculation and in the ability to analyze in real-time and in a short time of huge amounts of data in any form [Big Data Analytics].
AI is a popular branch of computer science that concerns building “intelligent” smart machines capable of performing intelligent tasks. With rapid advancements in deep learning and machine learning, the tech industry is transforming radically.
Learning systems and algorithms have become more and more effective and efficient, with a great improvement of techniques related to neural architectures (even in multiple layers) with incremental and not necessarily supervised learning; learning automatic has been successfully applied, for example, in the classification and processing of documents, in the understanding of natural language, in bioinformatics and image processing. More and more effective methods for speech recognition and image classification have been developed, successfully applied to robotics and computer vision; many web search algorithms, translators, speech recognizers, image and photo classifiers that we use daily take advantage of these ever-evolving techniques.
AI can assist new technologies, which are rapidly evolving, both as regards the design of the most suitable tools, and in terms of methodological contribution (for example, sophisticated sensors that require the development of advanced systems capable of processing, intelligently and in real-time, the information they produce, to automatically understand the situations of interest and plan actions in dynamic contexts).
The use of AI techniques allows a wide range of applications such as integrated systems for surveillance, monitoring and diagnosis, remote assistance systems and logistic transport planning, self-driving vehicles etc. The availability of technological tools for home automation opens up the possibility of applications in problems relating to the ageing of the population.
Another very interesting application field is the so-called ‘Future Internet’, characterized as an open network composed of self-organized and intelligent entities, such as software (agents, web services, softbots, avatars), hardware (objects, sensors, robots). or human beings.
Weak and Strong AI
Taking as a starting point the functioning of the human brain, an AI can perform some functions:
- act humanly (i.e., indistinctly concerning a human being);
- think humanly (solving a problem with cognitive functions);
- think rationally (that is, using logic, as a human being does);
- act rationally (starting a process to obtain the best-expected result based on the information available).
These considerations classify AI into two main types, weak and strong:
- Weak AI: it identifies technological systems capable of simulating some cognitive functions of man without however reaching the real intellectual abilities typical of the latter (problem-solving mathematical programs through which functions for solving problems or for allowing machines to make decisions).
- Strong AI: in this case, we speak of ‘wise systems’, that is, able to develop their intelligence without emulating thought processes or cognitive abilities like humans.
Machine Learning e Deep Learning
The distinction between weak and strong Artificial Intelligence forms the basis for the definition of ‘Machine Learning’ and ‘Deep Learning’, two fields of study that fall within the broader AI discipline.
What characterizes AI from a technological and methodological point of view is the learning method with which intelligence becomes skilled in a task or action. These learning models characterize and distinguish Machine Learning from Deep Learning. The following diagram illustrates the differences between Artificial Intelligence, Machine Learning, and Deep Learning
Machine Learning is a form of applied statistics, aimed at using computers to statistically estimate a complex function. It is a set of techniques (such as computational statistics, pattern recognition, artificial neural networks, adaptive filtering, the theory of dynamic systems, the processing of images, data mining, adaptive algorithms, etc.) that allow machines to ‘learn’ from data and, subsequently, make decisions or make a prediction about them.
Machine Learning is a type of AI that enables machines to learn from data and deliver predictive models
A Machine Learning system can be applied to a knowledge base coming from multiple sources to solve different tasks: facial classification, speech recognition, object recognition, etc. Unlike heuristic algorithms, which are those algorithms that follow a specific set of instructions for solving a given problem, Machine Learning enables a computer to learn how to recognize ‘perceptual configurations’ on its own and make predictions about them.
Machine Learning can be adapted to three different types of tasks:
What characterizes Machine Learning is the learning model; based on these models it is possible to make a classification of the algorithms:
- supervised learning: learning through examples of inputs and outputs to allow the AI to understand how to behave;
- unsupervised learning: learning through results analysis; in this case, the software understands how to act and the learning model adapts based on outputs that allow mapping the results of certain actions and tasks that the software will be required to perform;
- reinforcement learning: AI is rewarded when it reaches objectives, results, performs an action, etc… In this way, it can recognize the correct actions from the wrong ones.
Deep Learning is represented by learning models inspired by the structure and functioning of the biological brain and, therefore, of the human mind. If Machine Learning can be defined as the method that ‘trains’ AI, Deep Learning is what allows you to emulate the human mind.
Deep Learning is a sub-area of Machine Learning that makes use of “Deep Neural Network”, that is, equipped with many layers and new algorithms for the preprocessing of data for the regularization of the model. Deep Learning is inspired by Neuroscience (neural networks are a model of the neuronal activity of the brain). Unlike the biological brain, where any neuron can connect to any other neuron under some physical constraints, Artificial Neural Networks (ANN) have a finite number of layers and connections and a predetermined direction of information propagation.
The first requirement for training a Deep Learning model is to have very large training sets available; this makes Deep Learning extremely suitable for facing the era of Big Data. The reasons behind the popularity of Deep Learning are linked to the advent of Big Data and GPUs. Referring to a massive amount of data, the network ‘learns’, through the training algorithm, how to achieve objectives. Deep Learning is very sensitive to so-called ‘bias’: in a supervised model, if the labels are created incorrectly, the model will learn from the wrong data.
Deep Learning and Industrial Applications
Deep Learning has greatly influenced industrial applications: it can process a huge amount of data and recognize some discriminatory characteristics. Text-based searches, fraud or spam detection, recognition of writing, image search, speech recognition, NLP (Natural Language Processing) systems, Recommendation Systems, Street View Change Detection and language translation are just some of the tasks that Deep Learning is able to tackle; at Google, the Deep Networks have already replaced dozens of ‘rules systems’.
Deep Learning is a subfield of machine learning that is concerned with algorithms inspired by the brain’s Structure & functions known as artificial neural networks. A computer model can be taught using Deep Learning to run classification actions using pictures, texts or sounds as input
Today, Deep Learning for Computer Vision shows that it possesses skills that are already superior to those of humans, ranging from the recognition of common figures to the identification of cancerous nodules in lung tomographic images.
Artificial Intelligence in various Sectors
All the large multinational companies are competing not only to bring innovative startups in the field of AI into their own but also to initiate and feed research projects already in place (such as the recognition of images, faces, voice applications, translations linguistics, etc.). In the business world, the maturity (and availability) of technological solutions has brought the potential of AI to many segments; here are some of the most significant:
Marketing and Artificial Intelligence
Diffusion of virtual assistants (chatbot, Apple’s Siri, Microsoft’s Cortana, Amazon’s Alexa) that use AI both for natural language recognition and for learning and analyzing user habits and behaviours; real-time analysis of large amounts of data for the understanding of ‘sentiment’ and needs of people, to improve customer care, user experience, assistance and support services, but also to create and refine sophisticated engagement mechanisms with activities that go as far as forecasting purchasing behaviour, from which to derive communication strategies and proposal of new services. The application of AI in Marketing has achieved significant results. the main area of use is that of managing the relationship with users. A real discipline has recently spread, the Marketing of Artificial Intelligence, a branch of Marketing, which uses the most modern technologies that fall within the field of AI, such as Machine Learning and Natural Language Processing, integrated with mathematical, statistical and behavioural techniques. It is the use of AI and Machine Learning to persuade people to perform a certain action, such as buying a product or accessing a service; aggregation and analysis of data (even unstructured and based on natural language) in a continuous learning and improvement process to identify, from time to time, the actions, strategies and communication and sales techniques (those that have the greatest potential effectiveness/success for individual target users).
Artificial Intelligence in HealthCare
AI has had the advantage of improving many technological systems already in use by people with disabilities (for example, voice systems have improved to the point of allowing completely natural communication even to those who are unable to speak) but it is on the front of the diagnosis and treatment of tumours and rare diseases that the potential of AI will be seen. Currently, cognitive systems are available on the market capable of drawing, analyzing and learning from an infinite pool of data (scientific publications, research, medical records, drug data, etc.), at an unimaginable speed for humans, accelerating processes of often very critical diagnoses for rare diseases or suggesting optimal treatment pathways in the case of tumours or particular diseases. Not only that, AI-based virtual assistants are starting to see each other more frequently in operating theatres, in support of reception staff or those who offer first aid services.
Cybercrime and risk management
Fraud prevention is one of the most mature applications where AI materializes with what is technically called “advanced analytics”, very sophisticated analyzes capable of correlating data, events, behaviours and habits to understand in advance any fraudulent activity (such as cloning a credit card or executing an unauthorized transaction); these systems can also find application within other business contexts, for example for the mitigation of risks, the protection of information and data, the fight against cybercrime.
Artificial Intelligence and Supply Chain Management
The optimization and management of the supply and distribution chain now require sophisticated analysis and, in this case, the AI represents an effective system for connecting and monitoring the entire supply chain and all the actors involved; a very significant case of application of the AI to the Supply Chain Management sector is related to the management of orders (in this case the technologies that exploit the AI not only aim at the simplification of processes but also the total integration of them, from purchases to inventory, from warehouse to sales, up to even integration with Marketing for the preventive management of supplies according to promotional activities or communication campaigns).
Artificial Intelligence for Public Safety
The ability to analyze huge amounts of data in real-time and to deduce, through correlations of events, habits, behaviours, attitudes, systems and data for geo-localization and monitoring of the movements of things and people, offers enormous potential for the improvement of efficiency and effectiveness of public safety, for example for safety and crime prevention at airports, railway stations and metropolitan cities or the prevention and management of the crisis in cases of natural disasters such as earthquakes and tsunamis.
Artificial intelligence and business for value creation
The acceleration given by AI and Deep Learning in data analysis can translate into value for the business, provided that multiple challenges are overcome. The most complex is the preparation of the data that feeds the AI, given their origin from multiple sources and the time required for their transformation.
The potential of AI surprises every day with new extraordinary applications. However, making AI in operation and introducing it in a persistent and widespread manner into business processes represents a challenge that few excellences have so far been able to successfully face. Indeed, nowadays, just one in ten organizations can promote more than 75% of the AI models developed into production. A forecast made by Gartner analysts assumes that between now and 2023, at least 50% of Data & Analytics department heads will have difficulty getting their AI projects through the Proof of Concept (PoC) phase.
The industry 4.0 paradigm encourages companies to modernize and optimize their processes and, in the context of AI. However, if on the one hand there are numerous pilot initiatives in every industrial sector, on the other hand, there are still few projects that have turned into digital products or services launched on the market or released in production to transform the customer experience on digital channels.
The reasons behind the complexity are different. Security and privacy, inadequate volume and quality of data, accessibility constraints, limited understanding of use cases and related benefits, inadequate internal skills, just to name a few.
Operationalization: the work practices to be developed
What to do to manage this complexity? The common element underlying the interdependencies is data. It is necessary to work not only on technologies but also to consider the aspect of work practices and operating models. That is all that is necessary to integrate data and AI in an industrialized way in operations and business processes.
There are some work processes and practices that need to be developed. They are now defined as XOps (primarily ModelOps and DataOps) and are united by the need to achieve economies of scale, reliability, reuse and repeatability, regardless of the underlying technological infrastructure.
ModelOps takes its name from the concept of AI Model Operationalization and is undoubtedly at the heart of an AI strategy. These practices bring government and a life-cycle management approach to the world of AI, that is, the management of the entire life cycle of an AI decision-making model.
ModelOps is a combination of culture, policies, practices and tools that is key to an organization’s ability to operationalize AI, ML and analytic models at scale, evolving and improving model governance and performance at a faster pace than existing/homegrown methods. This speed enables organizations to employ AI decisioning to better serve their customers, secure their businesses, and compete more effectively in their market.
In addition to ModelOps, we often hear about MLOps as well. These two terms are used interchangeably. However, there are key distinctions between them that depend on the functionality and features provided by each technology, and of course, the value and scalability of AI depend on them.
Recently, ModelOp, Inc. published a comprehensive guide to understanding the differences between MLOps and ModelOps “Top 10 Questions about MLOps and ModelOps”
As stated by Stu Bailey, Co-founder and Chief Enterprise AI Architect at ModelOp “Understanding and valuing the distinction between ModelOps and MLOps is important because while both are needed, only one fully addresses the operational and governance process issues that are holding back nearly two-thirds of enterprise AI programs (the 2021 State of ModelOps Report).”
Also read Unlocking the Value of AI in Business Applications with ModelOps
Basically, ModelOps is an enterprise-wide capability that ensures validation of and accountability for all models in production. This includes managing all IT, risk, compliance, and business requirements throughout the model’s life cycle, as well as the health and operational efficiency of models once they are in production. ModelOps encompasses all models (not just ML) regardless of type of model, how developed, or where they run (on-premises, cloud, edge).
It’s based on the concept of DevOps but adapted to ensure good quality of AI/ML models. In general, ModelOps includes:
- model management, repository and model store;
- model versioning (to allow the evolution of models without defects generally caused by regressions);
- model training (i.e. management of learning process of models);
- Continuous Integration & Continuous Delivery (CI/CD);
- development environments;
- roll-out / roll-back;
- testing, etc.
Automating and orchestrating all aspects of the model life cycle ensures reliable model operations and governance at scale. Each model in the enterprise can take a wide variety of paths to production, have different patterns for monitoring and various requirements for continuous improvement or retirement.
DataOps must ensure an orderly flow of data from their origin to the points of use and consumption. The ultimate goal is to generate value from data as simply and quickly as possible. DataOps operate on pipeline dates for which they define aspects of data architecture, data modelling, data configuration, data quality and data integration.
DataOps combines Lean, Agile, and DevOps. It started with a series of best practices and has grown into a methodology. There are more and more data management, integration, warehousing, business intelligence, cloud vendors who are embracing DataOps.
ModelOps and DataOps have a common element represented by governance. It must allow for an adequate balance between instances of control, accessibility, accountability and traceability in the use of data.
We are in an accelerating phase of Digital Transformation. Organizations that will be able to grasp the cross-functional and multi-disciplinary nature of AI, carrying out operationalization, combining traditional professions (eg, data architect, data modeler, data steward) and new professions (eg, data engineer, AI engineer, data scientist, ML validators) will have the greatest opportunities to extract value by seizing the transformative opportunities of these technologies.
Follow me for daily updates on Technology and Innovation
ModelOps-CIOs | MLOps-Data Scientists | ModelOp
ModelOps is a process to manage AI initiatives across the enterprise. MLOps is a process designed for the Data…
ModelOp, “State of ModelOps 2021” report. April 15, 2021
Gartner,” Innovation Insight for ModelOps”, Farhan Choudhary, Shubhangi Vashisth, Arun Chandrasekaran, Eric Brethenoux, 6 August 2020
Forrester,” Introducing ModelOps to Operationalize AI”, Kjell Carlsson, Ph.d., and Mike Gualtieri, August 13, 2020.
What Is ModelOps and How Is It Different From MLOps? - neptune.ai
In the past few years, we've seen an increase in Artificial Intelligence and Machine Learning solutions in real-life…