Throughout your career as a project manager, you have had the opportunity to work on an extensive range of IT projects, ranging from complex monolithic structures to Software as a Service (SaaS) web applications. However, with the ongoing advancement of artificial intelligence (AI) and machine learning, an increasing number of projects with unique requirements and challenges are emerging.
As the development of Artificial Intelligence (AI) technologies continues to progress, it has become increasingly essential for technical project managers to maintain a positive relationship with these concepts. According to Gartner, by 2020, AI will create 2.3 million jobs, exceed the 1.8 million jobs it will displace, and generate an estimated $2.9 trillion of value to the economy by 2021. Sundar Pichai, CEO of Google, has even gone so far as to declare that, “AI is one of the most important undertakings that humanity is currently working on; it is deeper than fire or electricity.” With the potential of AI so far-reaching and its impact on the economy so great, it is essential for project managers to form a strong connection with these new technologies.
As a Technical Project Manager (PM), it is essential to understand the distinctions between traditional project management and AI project management in order to take advantage of the opportunities presented by the current business environment. Artificial Intelligence (AI) applications are having a profound effect on industries across the board, from finance to healthcare, and Technical PMs must ensure they are prepared for this shifting landscape.
Theory: AI vs. ML What It All Means
It is imperative to have an in-depth understanding of what Artificial Intelligence (AI) is before progressing further. To begin, it is beneficial to become acquainted with the definitions of the terms most commonly used in this context.
ARTIFICIAL INTELLIGENCE (AI)
Artificial Intelligence (AI), a branch of computer science, can be used to address a variety of issues related to pattern recognition, learning, and generalisation. AI is a rapidly growing field of computer science, exploring ways to make computers smarter, faster, and more efficient at solving complex problems. AI techniques can be used to automate processes, identify patterns, and provide valuable insights into data. These techniques can also be used to develop predictive models and create intelligent systems that can use data to make decisions and solve real-world problems.
Artificial general intelligence (AGI), a concept that has been overutilized in recent years, remains a distant reality for the foreseeable future. Instead, the majority of current artificial intelligence (AI) systems are classified as “narrow AI”, meaning they are designed to carry out one specific cognitive task to an outstanding degree, rather than demonstrating independent thought.
MACHINE LEARNING (ML)
Artificial Intelligence (AI) has given rise to a branch known as Machine Learning. This area of AI utilises statistical techniques to allow computers to learn from data without the need for explicit programming.
Due to the advancements in the field of artificial intelligence (AI) and the success of various machine learning (ML) techniques, there has been an increasing tendency for businesses to use the terms AI and ML interchangeably. It is important to note, however, that AI includes learning as well as other capabilities, whereas ML is solely focused on a program’s capacity for learning.
The Difference Between AI and Standard Algorithms
Artificial Intelligence (AI) has taught us an important lesson: its algorithms are highly dependent on data to make adjustments to their internal structures so that new data can be classified according to the data that has already been provided. Instead of merely following the rigid instructions that are coded into the algorithms, these adjustments are referred to as “learning” from the data. This concept is critical for AI to evolve and become more accurate and efficient.
It can be challenging to program software that can accurately differentiate between trucks and automobiles using the conventional approach. This involves manually coding for certain characteristics, such as larger wheels or a longer body, which could be used to distinguish between the two. However, this approach is highly prone to producing false positives and negatives, and can be extremely difficult to develop, potentially leading to an unusable outcome.
In scenarios such as these, Artificial Intelligence (AI) algorithms can be of great assistance. Once an AI algorithm has been trained, it can be supplied with a substantial amount of data, allowing it to modify its internal structure to start searching for characteristics that are essential in accurately categorising the images, rather than simply depending on static, pre-defined feature definitions.
Project Management for AI in Action
Data Is The Most Important
Due to our limited capacity to process large quantities of data, the sheer volume of data which is available to us can often prevent us from making use of it. Artificial Intelligence systems can be of great assistance in this scenario and enable us to benefit from this data in a productive manner.
It is a generally accepted principle in the field of Artificial Intelligence (AI) that the accuracy of a system’s predictions is highly dependent on the amount and quality of data used in its training. For example, an algorithm which has been fed a million data points is likely to be more accurate than one based on just 10,000 data points. In addition, the Boston Consulting Group (BCG) has emphasised the importance of data and training to the success of AI systems, pointing out that in many cases, the quality of data is more important than the sophistication of the algorithm itself. This is analogous to the argument that nurture has a greater influence on human development than nature.
Given the high likelihood of organisations lacking the necessary data in the right forms, the development of an AI system will often require a considerable amount of preparation and data cleansing. This is often the most labour-intensive and time-consuming stage of the project process, meaning that data analysts may need to dedicate a significant amount of time to ensure the successful completion of this critical step.
Furthermore, compared to conventional software development, the building of the data infrastructure and the data cleansing processes are more methodical, requiring the application of a specific project management approach.
In conclusion, it is important for project managers to bear in mind that the process of developing a suitable data infrastructure and preparing the data for usage could take significantly more time than constructing the machine learning model that will utilise the data. This is a factor to be considered when managing teams, establishing the scope of the AI project, and estimating the associated costs.
It is essential to maintain the quality of the dataset by regularly adding fresh data. This can be a decisive factor in choosing the most suitable machine learning solution. To ensure the best results for your machine learning project, even after its launch, it is important to stay up to date with the latest trends in the field.
The Lifecycle of AI Development
It is widely understood that the standard systems development lifecycle (SDLC) is a cornerstone of the IT industry. However, it is also necessary to recognise that the advances in Artificial Intelligence (AI) are revolutionising the SDLC by introducing new challenges and opportunities. The AI development process can be broken down into three distinct stages: ideation and data discovery, prioritising of Minimum Viable Products (MVPs), and the transformation of these MVPs into finished products. It is essential to remain informed and up-to-date on the impact of AI on the SDLC in order to remain competitive and successful.
IDEAGENESS AND DATA RESEARCH
At this early stage, the primary focus should be on the end-user of the Machine Learning product, as well as the data pools that are accessible to them. Ensuring that they have the resources and necessary information to be successful should be the highest priority.
As a project manager, there are two distinct approaches that can be taken to quickly reduce the machine learning product options available within a company. By leveraging their knowledge of the machine learning field, top project managers can use this phase to gain a deeper understanding of the intricacy of the given task. With rapid developments in machine learning research, certain difficult issues can be drastically simplified.
It is essential that the data be thoroughly cleansed and processed once it has been discovered. This task cannot be easily incorporated into project management methods such as Agile or Waterfall, as it usually takes place in distinct phases that do not lend themselves to sprints. To ensure accuracy and completeness, data cleaning is typically conducted iteratively, whereby datasets are gradually increased and prepared in conjunction with other development activities.
SETTING THE MINIMUM VIABLE PRODUCT AS A PRIORITY (MVP)
When it comes to developing machine learning products, it is still beneficial to have a functional prototype of a smaller product rather than a large, incomplete one. In order to maximise the value to the business, it is important to prioritise new machine learning Minimum Viable Products (MVPs) based on how quickly they can be delivered. A rapid victory for the whole team can be achieved if smaller, yet valuable products are delivered quickly; thus, these products should be given top priority.
It is beneficial to use the traditional Agile methodology when creating Minimum Viable Products (MVPs), while the development team should focus on developing Machine Learning (ML) models based on independently created and continuously evolving datasets. It is important to note that the data team does not need to be subjected to the same Sprint organisation as the team creating the MVP.
From MVP TO FULFILLED PRODUCT
At this point, it is essential to have a strong data infrastructure in place. Organisations should consider how to scale their infrastructure to meet the demands of their ML product if it requires frequent, global API access.
Prior to introducing modifications to the existing machine learning (ML) modules, it is essential to consider their potential impact on the functionality of the current product. To ensure that the product quality is not compromised, rigorous testing should be conducted before deploying it live. It is important to note that retraining the ML modules with new methods or datasets does not always lead to linear performance improvements. Although GAN attacks and testing of edge cases in ML modules are still in the early stages, project managers should bear this in mind when utilising a live ML product.
Significant Positions in the AI Development Lifecycle
The successful implementation of Machine Learning (ML) applications necessitates a wealth of data, thus introducing new responsibilities into the software development lifecycle of AI products. To be an effective project manager in this field, one needs to have an in-depth understanding of the three key roles of Data Scientists, Data Engineers and Infrastructure Engineers. These roles may sometimes be referred to by other titles such as Machine Learning Scientists, Engineers or Infrastructure Specialists, however it is paramount to comprehend their impact on the development of ML systems.
DATA SPECIALIST
Data scientists are the professionals who develop machine learning models. Drawing on their expertise in applied statistics, machine learning, and analytics, they integrate concepts to find solutions for real-world business problems.
At times, data scientists may be compared to more experienced data analysts. However, data scientists tend to possess strong programming skills, have comfort in working with large amounts of data from multiple sources, and are knowledgeable in machine learning concepts.
Candidates for this position must possess the ability to work autonomously and carry out exploratory research and data analysis. They must be comfortable analysing data to uncover useful information and trends, as well as possess a strong understanding of data architecture and data mining fundamentals.
Examples of programming languages that are fundamental to the practice of data science and related disciplines include Python, R, Scala, Apache Spark, Hadoop, Machine Learning, Deep Learning, Statistics, Jupyter, and RStudio. These languages are essential for performing complex data analysis and manipulations, and are used in many organisations to power their data-driven initiatives.
DATA ENGINEER
Software engineers with a primary focus on developing the necessary infrastructure and software to enable the functioning of Machine Learning (ML) products are known as data engineers. While data engineers may not be experts in machine learning, analytics or big data, they must possess a comprehensive understanding of these disciplines to be able to test their software and infrastructure. Data engineers prefer to concentrate on the overall architecture of ML products to ensure that the data scientist’s machine learning models are correctly integrated and applied to the real world.
Examples of fundamental skills include programming languages such as Python and Hadoop, data streaming techniques like MapReduce and Hive, as well as databases such as NoSQL, SQL, DashDB, MySQL, MongoDB, and Cassandra.
ENVIRONMENTAL ENGINEER
Infrastructure engineers are responsible for the infrastructure layer, which serves as the foundation for Machine Learning (ML) products. Although some of this infrastructure is put together by data engineers, it is typically built on top of the layer that has been set up and validated by the infrastructure team.
In order to establish a scalable and efficient setting for Machine Learning (ML) projects to serve a large number of users, Infrastructure Engineers must collaborate with members of various ML teams. These engineers are responsible for providing the necessary software for the platform and working closely with data centre partners to ensure that all aspects of the system, from the physical location of stored data to the hardware, are running efficiently. As AI-driven companies continue to grow, Infrastructure Engineers are becoming more and more critical to the success of ML initiatives.
Professionally: The fundamental skills required for end-to-end computing, IT infrastructure, and service management include Kubernetes, Mesos, Amazon Elastic Kubernetes Service (EKS), Google Kubernetes Engine (GKE), Hadoop, Apache Spark, Hadoop Distributed File System (HDFS), CEPH, Amazon Web Services (AWS) cloud computing, and data centre operations.
Common Issues of Today
As the use of Artificial Intelligence (AI) and Machine Learning (ML)-based technologies become more and more commonplace, it is essential for project managers to be prepared for potential issues that may arise. Experienced project managers are conscious of the potential difficulties that can occur at any stage of the project, ranging from the initial scoping to the completion of the project. It is important for project managers to be aware of both common and unique problems that may occur throughout the project.
TRUTH CHECK
It is highly probable that the problem you are attempting to solve does not necessitate a sophisticated Artificial Intelligence solution, despite the widespread appeal and potential of AI.
Prior to commencing any project, it is essential for a Project Manager to conduct a preliminary analysis to ensure that a Machine Learning approach is the most appropriate solution for resolving any potential predictive issues. In certain scenarios, an alternative and potentially more precise option may be to use statistical regression models, which may be simpler to implement.
In certain cases, it may be beneficial to employ a machine learning-based solution in addition to a basic statistical model. For example, when constructing a recommendation engine, it is advisable to begin with a straightforward approach and a shorter development timeline. This will serve as a reliable benchmark that the ensuing machine learning model should surpass.
SCOPE CREEP OF AI
Frequent causes of scope creep for Machine Learning projects often arise from trying to take on too many tasks simultaneously and underestimating the amount of time needed for data preparation. Scope creep can be detrimental to a project’s success, as it can lead to decreased productivity, missed deadlines, and project delays. Therefore, it is important for project teams to be mindful of the potential for scope creep and to set realistic expectations for the completion of the project.
It is important to effectively manage stakeholders in order to ensure that the initial issue is adequately addressed. To this end, it is beneficial to begin by setting achievable goals and objectives, rather than attempting to tackle the issue with overly ambitious targets. As the project progresses, it is essential to maintain regular communication regarding this strategy in order to ensure its successful implementation.
Begin by establishing and verifying basic, minute, and elementary components. If you confront a difficult task, try to break it down into smaller objectives that act as suitable substitutes for the bigger ones. The purpose of these assignments should be easy to articulate.
For example, one could attempt to predict whether a person will completely disregard an advertisement prior to attempting to forecast when they will click on it. This approach simplifies the problem and makes it easier for current machine learning (ML) models to anticipate and address it. Facebook has recently developed a great series exploring this topic in more depth, focusing particularly on the ML pipeline from model creation to model deployment.
It is of utmost importance to ensure that adequate data is available to facilitate your machine learning activities in addressing the second cause of scope creep. One of the most common mistakes made by project managers when beginning ML initiatives is failing to appreciate that the data needed may not already be available and/or in the required format. Appropriately managing this stage is essential, as it is frequently the longest phase of the ML project. To ensure a successful ML product is developed, it is imperative to guarantee that the data scientist has access to the essential data and is able to assess the accuracy and legitimacy of the information.
It is important to be prepared to continually label and clean data throughout the project, rather than just in the beginning stages. This task is not always the most exciting, so it is recommended to break it down into sprints so that the data team can appreciate the progress they are making, rather than having an overwhelming amount of work in the backlog.
Companies may find it cost-effective to outsource the labelling of their data to external organisations. While this may save money and time in the short-term, it may also lead to data inaccuracies, which can impede the performance of Machine Learning models. To prevent this, companies should employ the multiple overlap strategy. This entails having multiple people examine each piece of data before it is used, only if their findings are in agreement.
When you are formulating your project plan, make sure to allocate sufficient time for the data team to make necessary adjustments in the event that your labelling requirements change during the project, and new labels need to be implemented.
It is important to consider whether the data available can be used with existing machine learning (ML) techniques, as attempting to develop new ones may substantially increase the project’s timeline and scope. It is necessary to acknowledge that there is a high probability of failure if one attempts to address an ML issue that has not been solved before. Despite the increasing recognition of ML and the abundance of research papers that have been published regarding it, tackling ML problems can be incredibly difficult. Rather than attempting to create something entirely new, it is generally simpler to begin with an area of ML that already has many successful examples and approaches and endeavouring to improve them.
EXPECTATIONS, UX, AND MACHINE LEARNING
As project managers, it is essential to be prepared to evaluate the most effective way to lead the team tasked with developing AI products and creating a positive user experience. Recently, Google released a remarkable article that centred around their perspective on UX and AI, emphasising the importance of human interaction.
It is essential to bear in mind that the implementation of a machine learning solution should be designed to cause minimal disruption for operators and users. For example, chatbot technology frequently incorporates machine learning, yet a human operator should still be able to take control at any point. This is especially important if the machine learning solution is intended to either operate in place of human operators or to operate alongside them.
Moreover, it is possible that users of machine learning solutions may have an overly optimistic outlook due to the enthusiasm generated by the media’s coverage of artificial intelligence technology. As a result, it is essential for the project manager to ensure that expectations are realistic, as this can be a common issue.
It is imperative for any project manager to be mindful in communicating the capabilities of their AI tool to their stakeholders, in order to prevent any unrealistic expectations prior to the stakeholders’ experience of the technology. It is essential to have an understanding of Artificial Intelligence and its true capabilities in order to guarantee a satisfactory user experience for all involved. Being conscious of this is integral for a successful project outcome.
QUALITY ASSURANCE (QA) AND TESTING PRACTICES IN ML AI
Deep learning has come a long way since its inception and is now being utilised by an ever-increasing number of applications in order to achieve their desired outcomes. Despite the advantages that deep learning brings, there are some unique challenges that have emerged with its increased popularity, particularly in the area of testing.
Testing machine learning models, especially those created with neural networks, is significantly more difficult than testing standard software that has a set of rules established by humans. At this time, the majority of machine learning models are being evaluated solely by data scientists, and there is not a widespread consensus for a process to test ML products with traditional Quality Assurance teams in order to ensure they do not malfunction in unexpected ways.
As the development of novel methods, such as Generative Adversarial Network (GAN) attacks, continues to advance, it is increasingly essential for comprehensive model testing to be conducted in machine learning projects. This kind of “integration” testing will become even more important in years to come, and should be kept in mind, particularly when developing mission-critical machine learning products, even if it is not currently a concern for simpler projects.
ML PLAGIARISM AND MODEL THE FRAUD
Since the initial research was presented at the 2016 USENIX Security conference, and subsequent media coverage by Wired, it has become evident that it is possible to successfully steal a functioning machine learning model.
It is undeniably difficult to carry out this task successfully, however, it is essential to understand the implications if your model is accessible via a public API. In theory, a party with the necessary access could train their own network based on yours and essentially replicate your predictive capabilities.
Although the possibility of this occurring is still relatively low, it is important to take steps to develop a plan of prevention in case your project is impacted. Ensure that you and your team are in communication to create a plan that will protect your project from any potential attacks.
SHORTAGES OF TALENT
The current demand for highly-skilled Artificial Intelligence (AI) professionals is resulting in a higher cost of recruitment for such individuals. In fact, according to a New York Times report, top-tier AI professionals may receive salaries of up to $1 million a year when employed by major Silicon Valley technology companies. As a Project Manager, it is important to take these market dynamics into account when searching for AI specialists for your team, as they may have an impact on your recruitment strategies, budget, or the quality of the end product.
The scarcity of highly qualified data scientists and engineers goes beyond the individuals who have conceived the most recent deep learning algorithms. It is indicative of a broader need for professionals who possess the skills and knowledge to develop and utilise sophisticated artificial intelligence and machine learning processes.
Competing in machine learning contests such as Kaggle, where participants can win prizes of up to $100,000 for successfully tackling difficult machine learning tasks, often attracts the most talented individuals. If it is not possible to find machine learning experts in your local area, it can be beneficial to look into alternative solutions, such as hiring specialist contractors remotely or organising your own Kaggle competition for the most demanding machine learning tasks.
ETHICAL AND JURIDICAL DIFFICULTIES
There are two distinct legal and ethical problems associated with the use of AI in project management.
It is essential to be aware of the data source used to train Machine Learning (ML) models, as well as to determine if the necessary licencing and legal authorization has been acquired in order to utilise the data. Failing to have this knowledge can lead to a variety of difficulties.
In order to ensure compliance with applicable laws and regulations, it is imperative for Project Managers (PMs) to ensure that their teams are only accessing datasets that they have the proper licence to use. This is especially important when deploying models trained on data for which a licence may not be held. As the field of data licencing is still in its infancy, many of the solutions to these questions are uncertain. Therefore, it is essential to consult with a legal professional to address any potential issues.
This selection of datasets is available for use in training machine learning algorithms and can be accessed without charge. Each dataset provides an invaluable resource for developers, researchers, and data scientists to use in their machine learning projects.
The second set of challenges relates to managing the development of systemic prejudice within one’s system. Recently, there have been a number of cases which illustrate how this can occur. For example, one camera manufacturer was forced to admit that its smile recognition system only identified faces from a single racial group, due to the fact that it had been trained on data comprising solely of images from that race. Additionally, a high-profile software company had to take down its self-teaching Twitter bot after a few days of learning, as a collective of online users had collaborated to make it generate racist slurs and propagate irrational theories.
When creating systems that are essential to their success, Project Managers should ensure that they anticipate any potential problems and take steps to mitigate or avoid them as soon as possible. This is important to do, as the severity of these issues may vary from minor to catastrophic, with the latter having the potential to completely derail a project.
Solid Structures are Built on Solid Foundations
In conclusion, the coming age of AI will bring with it a wide assortment of exciting and dynamic projects that will require new approaches to development, a restructuring of the usual team typology, and the introduction of previously unknown challenges. This is certain to be a period of great innovation and progress, as the potential of AI is tapped into in ever more innovative ways.
Top-tier technical project managers possess a keen understanding of the intricacies of each project phase and the potential that their team can realistically achieve, while also having a sound grasp of the fundamental principles of Artificial Intelligence (AI). As AI is not a Commercial-Off-The-Shelf (COTS) solution, organisations that pursue the purchase of specific Machine Learning (ML) solutions must still invest in the research and testing of new concepts, as well as the proper management of their data and infrastructure.
It is evident that the advancement of Artificial Intelligence (AI) is transforming the types of software products being created and the ways in which they are being produced. Project managers, who can comprehend and implement these novel concepts, will be pivotal in the realisation of future machine learning products.
Additional Readings by the Author
DLs and NNs are two theories.
Project managers should be knowledgeable about the nuances between deep learning (DL), neural networks (NN), artificial intelligence (AI), and machine learning (ML). These distinctions can be beneficial in order to understand the capabilities of each technology, and help to determine the best approach for a given project.
DEEP LEARNING (DL)
In comparison to conventional algorithms which are designed for specific tasks, deep learning is part of a broader range of machine learning techniques which are centred on the acquisition of data representations.
Despite the utilisation of numerous other methods, most current deep learning models are based on an artificial neural network. Artificial neural networks are a type of machine learning algorithm modelled after the human brain, which is composed of interconnected “neurons” that are capable of learning and recognising patterns in data.
PHYSICAL NETWORKS (NN)
Neural networks are mathematical structures inspired by biology, which enable artificial intelligence (AI) systems to acquire knowledge from the data that is presented to them. By connecting neurons in a network-like structure, these systems are able to identify patterns, make predictions and make decisions based on the data they receive. This makes neural networks a powerful tool for AI systems, enabling them to continually improve and evolve their capabilities.
Neural networks may be conceptualised as a vast number of tiny gates that have the ability to open or close in response to input data. In recent years, the advancement of GPU processing power has enabled us to quickly modify a greater number of these small gates within neural networks, making these approaches more effective.
There exist a multitude of neural network varieties, each with its own unique applications and complexity. Different forms of neural network architecture can be referred to by names such as Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN).
