There are many recent trends in machine learning, and they offer numerous opportunities for Machine Learning professionals. These are:
Machine Learning Operations
Deloitte predicts that by 2025, the Machine Learning Operations (MLOps) market will grow by nearly $4 billion. While launching machine learning pilots is deceptively easy, deploying them into production is a major challenge. "Although the potential for success is enormous, Artificial Intelligence initiatives take much longer to implement than anticipated," says Chirag Dekate, senior director analyst at Gartner.
It is well known that many Artificial Intelligence pilot projects have yet to be shipped into production due to various constraints such as cost, latency and quality requirements. As a data scientist, I have experienced the same difficulties in the projects I have worked on. To successfully transition models to production, you need technical skills (or you need to work closely with the technical team) and machine learning skills.
Most of the active research in machine learning is on modeling techniques, while companies need help to use the models in the real production environment. Understanding how to design the models for optimization and use in real production environments is important. MLOps is the machine learning function that encompasses data preparation to deployment and monitoring model performance in production.
It is often a collaborative activity between data scientists and AI Engineers. Depending on the size and structure of the company, the responsibilities of the data scientist, data engineer and MLOps engineer differ. In smaller companies, a data scientist might take on all roles!
When pursuing a career in MLOps, you must familiarize yourself with different database architectures, data platforms, and workflows. One must have a deep knowledge of machine learning or deep learning deployments. Understanding batch and online deployments, model size optimization, and distributed training are also skills required in this field. Adding alerts, monitoring, and responding to data shifts are other essential skills to master.
Generative AI
Generative AI technology focuses on generating text, images, music, or even videos. With the recent ‘ChatGPT’ revolution, it is clear how powerful Generative AI will be. When an image generated by MidJourney, an Artificial Intelligence program that turns lines of text into hyper-realistic graphics, won the first prize in the Artificial Intelligence competition, it also triggered a debate about human art and Artificial Intelligence art. AI-generated music albums are already flooding YouTube.
The AI-generated short film "The Crow" won the Cannes Film Festival award. This fuels the debate about the copyright of AI-generated art and the possible side effects of misusing this technology. But this technology has definitely created a storm, and companies are considering how to use it to cut costs in their respective businesses.
In addition to art generation, Generative AIs like GPT-4 and ChatGPT can generate code themselves. These systems are likely to create non-existent models for the fashion industry, generate images for content creation, help companies with marketing, and assist startups with website creation - all in a very short time. This has birthed a new discipline called "prompt engineering.
Generative models like GPT-3, GPT-4, and Bard have recently gained much popularity. At the same time, image generation tools like DALL-E, Stable Diffusion, and Midjourney have already started a revolution in image generation engineering. Besides prompt engineering, companies can train or tune the GPT models for a specific domain. For example, developing a chatbot for mental health.
They could also build generative models from scratch for certain complex applications. For example, Amazon has used Generative AI techniques to show virtual fittings for makeup products. The old generative techniques, such as Generative Adversarial Networks (GAN)s, Neural Radiance Field (NeRF), etc., are still relevant and actively used in the field of Generative AI. In addition to pure generation tasks, these models help data scientists generate more synthetic training data and build better models.
If you want to make a career in this field, understanding Generative AI technologies is a basic requirement. One should also study the latest GPT technologies like GPT-4, Bard, etc. and understand how to create, train, or fine-tune them to cater to specific business needs.
Multimodal Systems
In isolation, NLP and computer vision have solved many text or speech understanding image and video understanding tasks. The idea of multimodal transformers is to use multiple data inputs for decision-making. This can combine text, integer inputs, speech, and images. This process is similar to human decision-making using multiple sensory inputs such as vision, speech, etc.
In 2013, I worked on a shopping experience using the Bing search engine (called Bing Shopping). Based on the product description, we had to categorize products into clothing, electronics, computers, etc.. The technology to understand images could have been more advanced at that time. Now, we can solve the same problem effectively with multimodal systems. We can use both product descriptions and product images and build multimodal systems.
Another example of a multimodal system is the multimodal sentiment analysis system I recently developed at MOST Research. This system takes users' video reviews as input and extracts sentiment from the user's body language, speech, and the actual content of speech. This is a much more effective way to understand user ratings.
The development of Transformers, a deep learning model, has supported multimodal development. Transformers were originally developed for NLP tasks. Later, Vision Transformers (ViT) were developed to solve complex computer vision tasks. Transformers also performed well with audio input. Finally, it made sense to use transformers for multimodal inputs. CLIP vATT, GATO, and BERTWithTabular are some of the most popular multimodal transformers.
Learning deep architectures for AI - the architecture of transformers and the different architectures of state-of-the-art (SOTA) multi-modal AI - is important to make a career in this field. Learn application-based development with multi-modal AI.
Artificial Intelligence Governance
The debate on building ethical and responsible Artificial Intelligence is becoming more important every year. Ethics Artificial Intelligence is part of Artificial Intelligence governance, which includes ethics, moral values, and legal values in developing Artificial Intelligence systems. Every company needs to ensure that their Artificial Intelligence products are ethical, unbiased, and provide privacy to build user trust.
Any developer who trains Artificial Intelligence systems should properly understand responsible Artificial Intelligence. When the pre-trained models learn from the data universe, it is important to ensure that the data fairly represents races, genders, and social structures. For example, the Google pre-trained model Bidirectional Encoder Representations from Transformers (BERT) answered the question "man works as" with "caregiver, waiter, hairdresser." When asked what job a woman would do in high probability, the answers were "nurse, waitress, maid...". This reveals a bias in the product.
According to the Stanford AI Index Report 2022, most generative models are only truthful 25% of the time. Initially, ChatGPT's violent responses generated concern. Since then, GPT-4 has been improved and is generating safe responses. For example, it does not answer unacceptable questions like suicidal thoughts or violent behavior.
The European Union (EU) Commission has proposed the EU AI Act to review ethics in high-risk areas such as healthcare and recruitment. The US and China are also introducing relevant legislation on ethical and responsible Artificial Intelligence. The main problem is that many companies need to realize the potential negative impact their systems could have. Developers must consider important facts such as effectiveness, robustness or reliability, bias, explainability, and privacy.
Any Artificial Intelligence developer needs to understand Artificial Intelligence governance issues. Also, different companies may have different data governance rules depending on the requirements of the organization/country. Solutions are being developed to identify governance and ethical issues in Artificial Intelligence systems.
Ethical and Explainable Machine Learning
In the landscape of machine learning trends for 2023, a prominent and imperative focus lies on the ethics of machine learning. This trend sheds light on the escalating concern over the ethical dimensions and comprehensibility of machine learning processes as they weave deeper into the fabric of our society.
The ethical use of machine learning is gaining increased attention. Ethical machine learning involves ensuring that models are used responsibly, fair, and unbiased and respecting users' privacy. It also involves considering these models' potential implications and consequences, including how they could be misused.
The concept of explainable machine learning, alternatively termed explainable Artificial Intelligence (XAI), rests upon crafting models that yield transparent predictions. Traditional machine learning models, particularly intricate ones like deep neural networks, are termed 'black boxes' as it is hard to understand their internal mechanisms. XAI endeavors to unveil the decision-making rationale behind these models, making their logic accessible to human comprehension.
Artificial Intelligence and Machine Learning models are continuously used to make decisions that directly affect people's lives, such as loan approvals, medical diagnoses, or job applications; we must understand how they're making those decisions and that we can trust their accuracy and fairness.
No-code or Low-code AI
It is estimated that 70% of applications will use low-code or no-code AI by 2025. Building no-code or AI code solutions will significantly reduce development costs and time. Gartner called this one of the most important trends in machine learning or trends in technology in general, for which high-quality solutions still need to match the high demand.
In the early days of machine learning, applications took even longer to develop. The problems were new then, and much time was invested in data cleaning, algorithm selection, and metrics selection. As several industries solved similar problems, solution patterns emerged, and solutions became more or less streamlined. This has led to the development of AutoML solutions by several startups. Cognitive services are also developed by large companies such as Meta and Azure. Amazon, etc.
These are low-code solutions where the user can call the Application Programming Interface (API) with the data set and complete the task. In most cases, the AutoML solutions or the APIs are tied to cloud services, which makes them a complete package for enterprise solutions.
Transformers have also enabled zero-shot learning, meaning we can use the pre-trained model as is, provide the inputs to the model, and extract the outputs. Platforms like Hugging Face provide these models and are popular with people because they are low-code and low-cost solutions. They are well-suited for research and academic purposes.
When looking for career opportunities in this field, it is important to understand the market's existing players and how to develop end-to-end solutions for several well-known problems. Many good solutions are on the market for automating traditional machine learning algorithms. Going forward, the focus will be on scaling solutions in different domains or markets and automating solutions with deep learning algorithms.
There are many recent trends in machine learning, and they offer numerous opportunities for Machine Learning professionals. These are:
Machine Learning Operations
Deloitte predicts that by 2025, the Machine Learning Operations (MLOps) market will grow by nearly $4 billion. While launching machine learning pilots is deceptively easy, deploying them into production is a major challenge. "Although the potential for success is enormous, Artificial Intelligence initiatives take much longer to implement than anticipated," says Chirag Dekate, senior director analyst at Gartner.
It is well known that many Artificial Intelligence pilot projects have yet to be shipped into production due to various constraints such as cost, latency and quality requirements. As a data scientist, I have experienced the same difficulties in the projects I have worked on. To successfully transition models to production, you need technical skills (or you need to work closely with the technical team) and machine learning skills.
Most of the active research in machine learning is on modeling techniques, while companies need help to use the models in the real production environment. Understanding how to design the models for optimization and use in real production environments is important. MLOps is the machine learning function that encompasses data preparation to deployment and monitoring model performance in production.
It is often a collaborative activity between data scientists and AI Engineers. Depending on the size and structure of the company, the responsibilities of the data scientist, data engineer and MLOps engineer differ. In smaller companies, a data scientist might take on all roles!
When pursuing a career in MLOps, you must familiarize yourself with different database architectures, data platforms, and workflows. One must have a deep knowledge of machine learning or deep learning deployments. Understanding batch and online deployments, model size optimization, and distributed training are also skills required in this field. Adding alerts, monitoring, and responding to data shifts are other essential skills to master.
Generative AI
Generative AI technology focuses on generating text, images, music, or even videos. With the recent ‘ChatGPT’ revolution, it is clear how powerful Generative AI will be. When an image generated by MidJourney, an Artificial Intelligence program that turns lines of text into hyper-realistic graphics, won the first prize in the Artificial Intelligence competition, it also triggered a debate about human art and Artificial Intelligence art. AI-generated music albums are already flooding YouTube.
The AI-generated short film "The Crow" won the Cannes Film Festival award. This fuels the debate about the copyright of AI-generated art and the possible side effects of misusing this technology. But this technology has definitely created a storm, and companies are considering how to use it to cut costs in their respective businesses.
In addition to art generation, Generative AIs like GPT-4 and ChatGPT can generate code themselves. These systems are likely to create non-existent models for the fashion industry, generate images for content creation, help companies with marketing, and assist startups with website creation - all in a very short time. This has birthed a new discipline called "prompt engineering.
Generative models like GPT-3, GPT-4, and Bard have recently gained much popularity. At the same time, image generation tools like DALL-E, Stable Diffusion, and Midjourney have already started a revolution in image generation engineering. Besides prompt engineering, companies can train or tune the GPT models for a specific domain. For example, developing a chatbot for mental health.
They could also build generative models from scratch for certain complex applications. For example, Amazon has used Generative AI techniques to show virtual fittings for makeup products. The old generative techniques, such as Generative Adversarial Networks (GAN)s, Neural Radiance Field (NeRF), etc., are still relevant and actively used in the field of Generative AI. In addition to pure generation tasks, these models help data scientists generate more synthetic training data and build better models.
If you want to make a career in this field, understanding Generative AI technologies is a basic requirement. One should also study the latest GPT technologies like GPT-4, Bard, etc. and understand how to create, train, or fine-tune them to cater to specific business needs.
Multimodal Systems
In isolation, NLP and computer vision have solved many text or speech understanding image and video understanding tasks. The idea of multimodal transformers is to use multiple data inputs for decision-making. This can combine text, integer inputs, speech, and images. This process is similar to human decision-making using multiple sensory inputs such as vision, speech, etc.
In 2013, I worked on a shopping experience using the Bing search engine (called Bing Shopping). Based on the product description, we had to categorize products into clothing, electronics, computers, etc.. The technology to understand images could have been more advanced at that time. Now, we can solve the same problem effectively with multimodal systems. We can use both product descriptions and product images and build multimodal systems.
Another example of a multimodal system is the multimodal sentiment analysis system I recently developed at MOST Research. This system takes users' video reviews as input and extracts sentiment from the user's body language, speech, and the actual content of speech. This is a much more effective way to understand user ratings.
The development of Transformers, a deep learning model, has supported multimodal development. Transformers were originally developed for NLP tasks. Later, Vision Transformers (ViT) were developed to solve complex computer vision tasks. Transformers also performed well with audio input. Finally, it made sense to use transformers for multimodal inputs. CLIP vATT, GATO, and BERTWithTabular are some of the most popular multimodal transformers.
Learning deep architectures for AI - the architecture of transformers and the different architectures of state-of-the-art (SOTA) multi-modal AI - is important to make a career in this field. Learn application-based development with multi-modal AI.
Artificial Intelligence Governance
The debate on building ethical and responsible Artificial Intelligence is becoming more important every year. Ethics Artificial Intelligence is part of Artificial Intelligence governance, which includes ethics, moral values, and legal values in developing Artificial Intelligence systems. Every company needs to ensure that their Artificial Intelligence products are ethical, unbiased, and provide privacy to build user trust.
Any developer who trains Artificial Intelligence systems should properly understand responsible Artificial Intelligence. When the pre-trained models learn from the data universe, it is important to ensure that the data fairly represents races, genders, and social structures. For example, the Google pre-trained model Bidirectional Encoder Representations from Transformers (BERT) answered the question "man works as" with "caregiver, waiter, hairdresser." When asked what job a woman would do in high probability, the answers were "nurse, waitress, maid...". This reveals a bias in the product.
According to the Stanford AI Index Report 2022, most generative models are only truthful 25% of the time. Initially, ChatGPT's violent responses generated concern. Since then, GPT-4 has been improved and is generating safe responses. For example, it does not answer unacceptable questions like suicidal thoughts or violent behavior.
The European Union (EU) Commission has proposed the EU AI Act to review ethics in high-risk areas such as healthcare and recruitment. The US and China are also introducing relevant legislation on ethical and responsible Artificial Intelligence. The main problem is that many companies need to realize the potential negative impact their systems could have. Developers must consider important facts such as effectiveness, robustness or reliability, bias, explainability, and privacy.
Any Artificial Intelligence developer needs to understand Artificial Intelligence governance issues. Also, different companies may have different data governance rules depending on the requirements of the organization/country. Solutions are being developed to identify governance and ethical issues in Artificial Intelligence systems.
Ethical and Explainable Machine Learning
In the landscape of machine learning trends for 2023, a prominent and imperative focus lies on the ethics of machine learning. This trend sheds light on the escalating concern over the ethical dimensions and comprehensibility of machine learning processes as they weave deeper into the fabric of our society.
The ethical use of machine learning is gaining increased attention. Ethical machine learning involves ensuring that models are used responsibly, fair, and unbiased and respecting users' privacy. It also involves considering these models' potential implications and consequences, including how they could be misused.
The concept of explainable machine learning, alternatively termed explainable Artificial Intelligence (XAI), rests upon crafting models that yield transparent predictions. Traditional machine learning models, particularly intricate ones like deep neural networks, are termed 'black boxes' as it is hard to understand their internal mechanisms. XAI endeavors to unveil the decision-making rationale behind these models, making their logic accessible to human comprehension.
Artificial Intelligence and Machine Learning models are continuously used to make decisions that directly affect people's lives, such as loan approvals, medical diagnoses, or job applications; we must understand how they're making those decisions and that we can trust their accuracy and fairness.
No-code or Low-code AI
It is estimated that 70% of applications will use low-code or no-code AI by 2025. Building no-code or AI code solutions will significantly reduce development costs and time. Gartner called this one of the most important trends in machine learning or trends in technology in general, for which high-quality solutions still need to match the high demand.
In the early days of machine learning, applications took even longer to develop. The problems were new then, and much time was invested in data cleaning, algorithm selection, and metrics selection. As several industries solved similar problems, solution patterns emerged, and solutions became more or less streamlined. This has led to the development of AutoML solutions by several startups. Cognitive services are also developed by large companies such as Meta and Azure. Amazon, etc.
These are low-code solutions where the user can call the Application Programming Interface (API) with the data set and complete the task. In most cases, the AutoML solutions or the APIs are tied to cloud services, which makes them a complete package for enterprise solutions.
Transformers have also enabled zero-shot learning, meaning we can use the pre-trained model as is, provide the inputs to the model, and extract the outputs. Platforms like Hugging Face provide these models and are popular with people because they are low-code and low-cost solutions. They are well-suited for research and academic purposes.
When looking for career opportunities in this field, it is important to understand the market's existing players and how to develop end-to-end solutions for several well-known problems. Many good solutions are on the market for automating traditional machine learning algorithms. Going forward, the focus will be on scaling solutions in different domains or markets and automating solutions with deep learning algorithms.
