What Data Science research topics should you follow?

1. Generative AI: to create the various content 2. Explainable AI: AI which can help to understand and interpret behind Machine learning and Deep learning models 3. Faster Deep learning model for real time usage:

We can start with existing DL model in GitHub. I start to search model that I like, for example TTS (text to speech). We can use it first, to understand how it works. After that, we can start re-train the model or Fine-Tune it. In the end we can know state-of-the-art of TTS.

Although Deep Learning has become a goto method for almost any task that you can think of today, it is important to also think of it's downsides. It lacks explainability It is more compute intensive than many of the traditional Machine Learning techniques (Although quantization, distillation etc. are bridging this gap now) It is prone to overfitting given a smaller quantity of data Despite these with appropriate care, Deep Learning continues to dominate the performance across classification,regression, generative models etc. but one should really ponder if one can also achieve comparable performance with traditional rule-based or machine learning algorithms which can give good explainability and save computational resources...

I aim to understand and integrate specific self-healing mechanisms in CI/CD and CD4ML, recognizing their role in enhancing deep learning systems. By identifying how these mechanisms directly reduce downtime, minimize interventions, and improve model quality, I plan to blend this knowledge with deep learning fundamentals and emerging trends. My approach involves identifying knowledge gaps, exploring case studies, and engaging with recent research to effectively incorporate these advancements into my AI skill set.

Deep learning stands as the pinnacle of machine learning, leveraging artificial neural networks for profound data analysis. Its applications includes; image recognition, natural language processing, speech generation, and beyond, have created groundbreaking advancements in data science. To master this domain, one must delve into statistics, foundational concepts, frameworks, and the latest research challenges, appreciating the transformative potential that deep learning holds for our technological landscape.

There are several great branches in data science but in my point of view, explainable AI is a crucial one. As the models that we usually utilize get more and more complex, explainability of them get decreased. Several industries around the world are heavily regulated by the governments like finance or insurance and the governing agencies demand explanations towards the results of the models. The outcome should be understood either by the customer or the business personnel. As the world embraces the use of large language models in their daily activities; bias and discrimination should be main dangers towards the AI use in building predictive blocks. We see that several companies like Google started building departments for XAI.

As we have reached to the phase of LLM where interpretation of model is getting tough day by day.Like when working with finance or medical industries it’s hard to rely on the outcome as the cost of failure would be higher and in medical it might affect someone’s life.So, explaining the interpretation of model is crucial to both data scientists and to the stakeholders by being very transparent about the limitations or effectiveness. I would suggest if someone is working on any problem start with basic models that are easy to explain and avoid using advanced models in very beginning like linear regression models can be explained very easily as compared to random forest or XGboost.

As AI models become more complex and approach AGI, understanding their decision-making is crucial. Explainable AI (XAI) research focuses on: - Methods to interpret model predictions: Explain why a model made a specific decision. - Developing fair and unbiased AI: Ensuring models don't discriminate based on sensitive data. - Building trust in AI systems: Making AI more secure and transparent to users and regulators. It also involves enforcing new/amending old regulations (importance evident with the increasing lawsuits) but that's a topic for another day.

In my line of work, the ultimate customer is usually a CXO. Whether you are in a regulated or unregulated business a 'black box' model built on 'secret sauce' doesn't help build trust with stakeholders. You need the ability and aptitude to deconstruct one feature at a time to show no causality is broken and spurious data has been used. e.g. using lower variability range of response variable to reduce overall modeling error. The top 5 research areas my customers are interested are as follows: 1. Causality in Explainable AI 2. Ethical Considerations in XAI 3. Quantitative Metrics for Explainability 4. Improving Interactive Explainability 5. Robustness of Explanations

Incorporating Explainable AI not only addresses ethical dilemmas but also empowers users to comprehend, critique, and enhance AI systems. This transparency fosters a partnership between humans and AI, ensuring that technological advancements align with societal values and human expectations.

🕵🏽♀️ Data Science is closely related to data organization and is susceptible to cause or impacted by data breaches. Many common practices, such as CSV data for exploration, can result in vulnerable data sets. 🧑🏽🏫 When a Data Science experiment involves PII, protected and IP-sensitive data, it is necessary to educate the users about data protection practices. 📕Establish mandatory data protection and privacy audits and annual training for the Data Science project/community in the enterprise. 🧐 Establish a responsible AI/read team to assess the risk of bias and security threats arising from Data Science Projects. 💡Establish MLSec operations standards and start educated enforcement of the practice.

Data privacy in data science involves cool stuff like making sure individual data doesn't mess up analysis (differential privacy), secret machine learning on encrypted data, and doing math on encrypted stuff. People need a say in how their data is used, and we're figuring out how to measure and defend privacy. It's a dynamic scene, so keeping up with tech and rules is key.

Data Privacy is almost ignored in any data science life cycle. It is key to note that, handling and processing Personal Identifiable Information should be accompanied by strong data protection and management guidelines. This is a wide and interesting area that has not been fully researched so far.

Data privacy is a critical priority in data science, particularly in light of stringent regulations like Europe's GDPR and California's CCPA. These laws, along with initiatives like the EU's AI Act, are shaping the handling and protection of personal and sensitive data. It's essential for data scientists to understand and comply with these regulations and stay updated with continual changes in the legal landscape. By doing so, they can ensure their data practices are not only technically advanced but also ethically responsible and legally robust, maintaining trust and security in an increasingly regulated environment.

Data privacy concerns are critical and likely to continue to have increasing importance in the future. Depending on the industry, data may be subject to different laws and regulations. It’s important to develop the appropriate data infrastructure from day one. All industries need to protect the data of their customers or clients. Data breaches in any industry can decrease trust, lead to lawsuits, and hurt the bottom line. Engaging a data security expert or building a data security team can be an important step in ensuring data security. Data collection is another important consideration. In the healthcare context, where data can be extremely sensitive and confidential, it is critical to obtain consent for data use and data collection.

No matter the type of data scientist you are and the type of work you engage in, data visualisation is the bread and butter of a good data scientist and the foundation for sharing any data-driven insight. Knowing which type of visual is the most effective at capturing the information you want for different scenarios is critical at capturing your audience and getting your point across.

De todos os tópicos de ciência de dados a serem seguidos, a visualização de dados com certeza é o meu preferido, não me entendam mal, Machine Learning, Deep Learning e demais tópicos de pesquisa são exatamente importantes de serem seguidos e acompanhados de perto, mas a Visualização de dados tem um lugar especial no meu coração. Os conteúdos a respeito são sempre muito ricos, intuitivos e agradáveis de ler, afinal de contas, eles são produzidos com os mesmos pilares que estão sendo ensinados. A satisfação de ler estes tópicos se torna a satisfação de produzir algo valorosamente semelhante.

Last year I had the privilege of working closely with a designer to create visualization standards. I learned a lot about accessibility, particularly the different types of color blindness and what palettes were differentiable to people with those challenges. In the end, we were able to create an optimal standard that was rigorous and accessible.

Very few recommendations in data science are as straight-forward as this one: Every research of data visualization should start with Edward Tufte’s “The Visual Display of Quantitative Information”.

Data visualization is the art and science of presenting complex information in a visually engaging and comprehensible way. By transforming raw data into charts, graphs, and interactive dashboards, it enables individuals to quickly grasp patterns, trends, and insights. Effective data visualization is not just about aesthetics; it's about storytelling. The choice of visual elements, color schemes, and interactivity should align with the narrative, simplifying the communication of data-driven messages. In a world inundated with information, mastering data visualization is a powerful skill for conveying compelling stories and aiding decision-makers in extracting actionable intelligence from data.

Large language models(LLMs) have taken the world by storm and made AI reach even the non technical users across the globe. These LLMs are trained on vast quantities of text data available on the internet and thus can be seen as having a vector map of the entire human knowledge inside of them. The areas in which we can effectively apply data science and AI using LLMs are only going to expand further in the coming years. That’s why I will recommend to start here. From LLMs, we are now moving towards LMMs(Large multimodal models) - models which can take text, image, video etc any kind of input and output in any format. This will turn out to be the melting pot where other areas also combine with NLP and we may move towards AGI.

Adding up to the context, mentioning about some of the crucial steps to follow in the field of NLP & LLMs. 1. Contextual Embedding Model Selection: Choose a state-of-the-art contextual embedding model, such as BERT (Bidirectional Encoder Representations from Transformers) or GPT (Generative Pre-trained Transformer). These models excel in capturing contextual information, vital for understanding the nuances of language. 2. Input Representation: Preprocess the input text by tokenizing it into subword or word-level tokens. 3.Pooling or Attention Mechanisms: Depending on the specific task, employ pooling or attention mechanisms to aggregate the contextual embeddings.

It is not just the main driver of the emerging Large Language Models that have taken over the spotlight recently, but it also encompasses everything to do with textual input. From basic text processing, to multilingualism inputs, improved information retrieval, data structuring, all the way to end-to-end conversational agents. Natural language processing is a must-have skill.

The next big thing in NLP would be small and custom LLMs and LLMs in local languages. Most enterprises would want to train their own LLMs for their custom applications. Focusing on them would help

To Start in NLP I recommend starting from the traditional approaches like tf-idf and traditional models which work well with tf-idf - like approaches. A good next step would be to learn about embeddings starting from Word2Vec to BERT/GPT. Embeddings are extremely useful across a wide range of applications like topic modelling, Question Answering systems, etc.

🔧 Data Engineering: The Backbone of Data Science! 📊🔍 Data Engineering is the unsung hero 🦸♂️ in the journey of a Data Scientist. It's like the sturdy foundation beneath a towering skyscraper 🏢. Data Engineers build, maintain, and optimize data pipelines, ensuring Data Scientists have access to clean, well-organized data. 🚀🔑 Without solid data engineering, Data Scientists would be lost in a maze of messy data 🧩, unable to unlock valuable insights. 🤯🔓 So, to succeed in the world of data science, tip your hat to the diligent Data Engineers. They make the magic possible! 🎩🌟 #DataScience #DataEngineering #KeyToSuccess

Although data engineering is not always at the forefront of AI/ML news and research, it is absolutely vital to ensure a robust ML pipeline. Because large amounts of data are often ingested and transformed in Python, these workflows can be bogged down by Python's inherent speed limitations. However, there have been a slew of new tools being developed to combat this, of which Polars and cuDF are at the forefront. For example, cuDF is a 1:1 replacement for Pandas in Python and leverages Nvidia's CUDA GPU framework that allows dataframe operations to be executed orders of magnitude faster than base Pandas. Keeping up with these developments can save you time, money, and headaches!

Data engineering encompasses the elements of data science concerned with acquiring, preparing, storing, and delivering data. It empowers data scientists to handle diverse and complex data sources like databases, APIs, web scraping, streaming, and cloud platforms. Data engineering also facilitates the creation of scalable, dependable, and efficient data pipelines and architectures that support data analysis and machine learning. Proficiency in data engineering requires acquiring skills, tools, and best practices to effectively manage data challenges and optimize data performance.

Data engineering is the backbone of robust data infrastructure. It involves designing, constructing, and maintaining the architecture that enables efficient data processing and analysis. Data engineers leverage programming skills, database expertise, and knowledge of big data tools to build scalable data pipelines and ensure data quality.

Data engineering is crucial for structuring the vast modern data ecosystem, enabling advanced analysis and machine learning. Mastering it involves understanding how to collect, store, and process data from diverse sources, ensuring that pipelines are scalable and efficient. With robust data engineering skills, one can build systems that not only handle large volumes of data but also make it accessible and useful for meaningful discoveries and data-driven decision-making.

Diving into causal inference, especially with panel models, I believe it's a game-changer in data science. Business problems aren't always just about predicting outcomes, as most Machine Learning courses seem to suggest; sometimes, it's about understanding the why behind changes, offering insights for targeted actions. In the real-world of business, knowing if a new training program truly improves performance or if a marketing campaign increases sales depends on causal analysis. Since A/B testing isn't always possible, or it may have high costs, knowing how to use observational data is a smart approach. RDD, IV, propensity scores, diff-in-diff, etc., are great tools that not all data scientists know.

Newer doesn’t necessarily correlate with better. There are techniques that are decades old that have withstood the test of time because of their elegance. When exploring ways to add tools to your data science toolbox, don’t exclude the “historical data”, so to speak. As a bonus, these techniques typically require less compute and are therefore cheaper to implement and maintain.

Protecting your ML models. I have observed from my interactions that most data scientist don't focus on how to protect their ML model once they are done training and operationalizing it. There is a need for data scientists to also focus on the following: 1. ML model encryption 2. Working with the admin team to ensure access control is implemented to ensure unauthorized users can not access the model to replace it with malicious model. 3. Add a signature or watermark to your ML model. This will ensure that your flask application will only execute your model once it's able to validate the signature or the watermark. Model security is key.

Research in data science ethics is critically important as the field continues to expand its impact on society. Ethics in data science encompasses a range of issues that arise from the acquisition, analysis, and use of data, particularly when dealing with human subjects and sensitive information. One crucial area of ethics research is algorithmic fairness and bias. Data science algorithms can inadvertently perpetuate or even exacerbate societal biases if they're trained on unrepresentative datasets or if the algorithmic design does not account for fairness. Research is needed to develop techniques for bias detection and mitigation in machine learning models to ensure equitable outcomes across different demographic groups.