QuantUp

CTO & Co-Founder

PrediqTank

Member

suchwalko.pl

Improving companies' data-driven decisions with #machinelearning, #AI, #datascience and optimization:  modeling, consulting, short courses and workshops (with R or without).

Owner, Data Scientist, Instructor

Bioavlee S.A.

Expert in Machine Learning algorithms

Apperise Group Sp. z o. o.

Apperise offers commercial Business Intelligence solutions.

Principal specialist: data analysis, data mining

MedicWave AB (publ)

Everything: from software development to marketing and sales.

Vice Chief Executive Officer

Chief Science Officer

Chief Business Development Officer

Wroclaw University of Science and Technology

Bioengineering, statistics, data mining, computer science.
Lectures, computer labs, supervision of M.Sc. and B.Sc. theses, popularization of science.

Lecturer

Institute of Biomedical Engineering and Instrumentation, Faculty of Fundamental Problems of Technology

Teaching and Research Assistant

Lukas Bank

Building of statistical models: scoring systems, forecasting, provisioning, portfolio modeling, Basel II modeling. Software development: data analysis, building predicitive models, forecasting.

Credit Risk and Modeling Statistician / Expert

● 25 years of professional experience in applying mathematics (Data Science, Machine Learning etc) in many organizations (corporations, universities, startups) and many roles (from an employee to an owner). 

● Over 100 of real-world Data Science projects delivered. Hundreds of hours of commercial trainings conducted. 

● With QuantUp works for companies to improve their data-driven decisions.

● A co-author of a book on forecasting in R.

#datascience #machinelearning #artificialintelligence #AI

Dear data scientist, 
you are not alone! 😉

🥊 It’s not easy to beat simple models.
😟 Most new attempts during an analysis fail.
😢 Usually miraculous ideas do not work.
😁 Fortunately, there are also approaches that end with (at least) moderate success.

👊 Some of the most important qualities of a data scientist are patience and resistance to failures.

#datascience #machinelearning #ai

Artur Suchwalko, PhD

This is mine:

Difference between machine learning and AI:
➡️ If it is written in Python, it's probably machine learning
➡️ If it is written in PowerPoint, it's probably AI

(Mat Velloso, Twitter)

❓What's yours❓

#datascience #machinelearning #ai 

I see many ChatGPT prompts that can generate code doing data analysis or building models.
They are useless, IMO. Why?

Even more, some claim that these codes “solve real-world problems.” It’s really tempting…

❓This is a question of risk and reward:
Is the risk of making mistakes using ChatGPT lower or higher compared to the standard process? If it is higher, then what is the reward for this risk?

The risk is much higher and there is no substantial reward, IMO.

❓ Questions related to the ChatGPT generated code – as food for thought:

🔴 Is the code correct?
🔴 What’s the code quality?
🔴 Is the applied method correctly implemented?
🔴 Is the method right for the data?
🔴 Is the method good enough?

🔵 What part of the process will be simplified thanks to ChatGPT?
🔵 How important is this part?
🔵 What are the hardest parts of a data science project, and how does this approach help?

🟡 How does using ChatGPT influence risks in the project?
🟡 Does it increase or decrease cost of the project?
🟡 What about results of the project?

🟣 What about the intellectual property of the code?

❗DS projects are extremely sensitive to errors. I would even say that they are fragile.

And these errors can be very painful, even if they are seemingly small ones. For example:

➡️ poor understanding of the data,
➡️ minor errors in the data,
➡️ differences in the way training data is collected and for production use of the model,
➡️ minor programming bugs like changing the order of arguments or misplacing parentheses,
➡️ incorrect assumptions that cannot be verified, 
➡️ errors in thinking, 
➡️ being not precise enough
➡️ carelessness in stating and verifying hypotheses
➡️ …and many more…

❗ One of the most painful mistakes I've seen in data science projects is a misunderstanding of the business problem to be solved or a mistranslation into the mathematical language.

Making a seemingly small mistake can change everything. It can make it appear that the expected project result is achievable, but in reality it isn't. 

I have listed above errors that, in most cases, cannot be detected with the help of (software) tests or simply by using the built solution. You need to have a process of building DS / ML / AI solutions that reduces the risk of their occurrence. And then "sense" that there are errors there and figure out where they are 😉

❓ The final question:
What's your own conclusion?

Last but not least, in some cases, these AI-generated codes can be helpful, e.g., for someone experienced and skilled to remind them how to write some particular code snippets.

#machinelearning #ai #datascience #chatgpt

Do not overcomplicate. Don’t build complex predictive models first. 
Start with simple models to have benchmarks. Why?

From my experience, if one has a simple solution like logistic regression or a tree for classification it is not that easy to build something much better. Surprisingly, the situation is similar for time series forecasting, where it is hard to beat a seasonal naive method.

That's the reason to use simple benchmarking methods.
If one takes, e.g., xgboost, applies it properly, and tunes hyperparameters correctly, it's usually enough for classification / regression.

Despite this, many people start with deep learning, relying on raw events instead of a simple method based on aggregates.

When working on projects at QuantUp we always use simple methods first:
➡️ to deliver the first version of the solution as soon as possible,
➡️ to know where we can be using simple means and not much of the budget,
➡️ because we want to solve the problem with appropriate tools, not to build complex models instead,
➡️ a small improvement in accuracy is not always worth the effort.

For the sake of simplicity, I silently assumed above that we are dealing with (potentially) tabular data and a classification / regression problem.

Of course, using deep learning or other approaches in many cases is much better for some problems and data.

#machinelearning #datascience #ai

ChatGPT can only pretend to be a virtual machine. Moreover, you never know where the accurate pretending of a virtual machine ends.

ChatGPT is only a language generation model. It has been designed to spew out human-like responses that are based on the patterns it learns from large amounts of data from the internet. So it can neither imitate nor simulate a real computer. It can only pretend to be one. 

So if you ask ChatGPT to act like a rabbit 🐇, you will get this answer:
Thump, thump, munch, munch, hop, hop! (proven!)

Is it really a rabbit? I don’t think so.

Try asking ChatGPT if you're not convinced. Like me, you should receive a similar response.

Of course, there is a philosophical question: can we say that something is a virtual machine when it behaves exactly like one? 😉 

#AI #machinelearning #chatGPT 

Taplio

Artur Suchwalko, PhD's Linkedin Analytics

Artur Suchwalko, PhD

Artur Suchwalko, PhD's Best Posts (last 30 days)

Want to drive more opportunities from LinkedIn?

Taplio

.css-1jgreyo{background-image:linear-gradient(284deg, rgba(39,43,72,1) 0%, rgba(33,47,183,1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Artur Suchwalko, PhD's Linkedin Analytics

Artur Suchwalko, PhD

Artur Suchwalko, PhD's Best Posts .css-qqfgvy{font-size:var(--chakra-fontSizes-md);}(last 30 days)

Want to drive more opportunities from LinkedIn?

Artur Suchwalko, PhD's Linkedin Analytics

Artur Suchwalko, PhD's Best Posts (last 30 days)