Jasper Busschers on LinkedIn: I start to see a pattern from my experience as data scientist. Whatever… (2024)

Last 2 days did a workshop on discovering IBM Watsonx. After seeing only genAI and LLM everywhere I found it humble to see them target any ML model, not jumping on the hype. Similar to what I aspired to do with previous product design, they try to take away the responsibility of choosing what model can go to production away from the data scientist. I am happy to see AI model development go in this direction and am sure a big step towards more ethical AI is to involve more stakeholders in model selection.

14

🎼 Music production with GPT ~ Code release Since I had a week between jobs, I finally had time to experiment with music production and AI.But as I wanted to try ableton their official integration, I noticed they wanted me to upgrade my version for 429 euro extra.As programmer I started working on my own integration to link python to ableton.Like with any solution I design, I tried to least distupt the workflow by keeping most of it in the original software (ableton) and enhanced it only with a minimal ui.I happily share my project for others in my network who are into music production.GITHUB: https://lnkd.in/e5NskpRpUpdates on the planning- Keep a database of generated melodies to quickly search matching melodies- Also generate a timeline when to activate and deactivate melodies- Fine tuned models for better melody generation#ai#gpt#music#production#integration

17

Hello Network,After 3 years working at Cnext to bring our vision of a decision intelligence platform to live. I have decided to make my next carreer step and join Cognizant as senior data scientist.This was not an easy choice as anyone working in a start up would understand. But making an AI product in the current market is incredible challanging.1) Regulation such as gdpr, AI act, medical device regulation create significant barrier to entree towards many use cases.2) As a product vendor that does not own data but only technology, companies that do own the data have to allow you to enter the market.3) Scaling is hard, customers might have the same data sources, but that does not guarantee the solution that works for 1 customer will work for another.The best tip I can give to new founders is to not start from technologie or value. Instead start from a MOAT, "If I was able to have this dataset, I would have such competative advantage"As such I am excited to get started as senior data science consultant at Cognizant designing tailor made solutions for different clients.

This content isn’t available here

Access this content and more in the LinkedIn app

32

8 Comments

Not just in RL, but in any optimization, doing single-objective optimization almost always leads to unwanted side effects.The paper mentions :"MORL methods could model thereputational harm as an additional objective to be minimised. But this may be very difficult to define quantitatively. "This goes for any optimization whether it is:𝐌𝐚𝐜𝐡𝐢𝐧𝐞 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠-Create a recommendation system that both increases revenue, customer satisfaction with reduced risk of unfairness.-Find ML models that solve the problem best, while using as little computation as required.𝒄𝒉𝒂𝒏𝒈𝒆𝒔 𝒊𝒏 𝒚𝒐𝒖𝒓 𝒃𝒖𝒔𝒊𝒏𝒆𝒔𝒔 𝒑𝒓𝒐𝒄𝒆𝒔𝒔-Change your pricing strategy to minimize churn, maximize revenue and conversion.𝑲𝑷𝑰'𝒔 𝒕𝒐 𝒋𝒖𝒅𝒈𝒆 𝒑𝒆𝒓𝒇𝒐𝒓𝒎𝒂𝒏𝒄𝒆-Make best strategic decisions that both maximize growth, while minimizing employee turnover and without reducing the quality.For that reason I am working on a framework that unifies any optimization problem under the same multi-objective definition.

13

💡 Announcing the Development of a new Optimization Libraryhttps://lnkd.in/gF73c5zPAs an AI professional, I'm excited to share that I'm developing a new optimization library, planned for release later this year. This library aims to address key challenges in the field:❓ Why another optimization library?- Current tools like Scikit-learn, PyTorch, and TensorFlow offer limited support for multi-objective optimization. This library simplifies finding trade-off models for any combination of objectives, aimed to ease the use of multi-objective optimization in model development.- Users will have the ability to choose whether to train models end-to-end or as separate modules. This feature is particularly useful for optimizing decisions composed of multiple sub-decisions.- Built on a low-code language (DMN) for both the input and output layers, the library ensures interpretability, interoperability, and interchangeability of model chains.- The library will support single-objective reinforcement learning, multi-objective reinforcement learning, supervised learning, and multi-objective supervised learning, all through a single, user-friendly interface.❓ How does it work?It is build on Evotorch(https://lnkd.in/eSxqUjJV), an evolutionary optimization library that efficiently solves multi-objective optimization problems.❓ How do you use it?The library is being developed to have a simple interface to manipulate the requirements of any function you wish to model.#decisionmaking #decisionintelligence #optimization #ai #python

25

6 Comments

I wanted to give this paper attention since I believe it is close to my master research in multi-objective reinforcement learning which I believe is becoming very important in reducing negative side effects of AI.Paper : https://lnkd.in/eSKeUp3Wcode : https://lnkd.in/eEXECTuPContext : Assistant AI's are trained first to predict the most likely next word, then fine tuned to predict words such that people like the answer the most. As I mentioned in my thesis, having an AI optimize for a complicated reward such as producing the most liked output can cause the agent to learn undesired behavior.Contribution : The paper changes the objective of the second stage to not just optimize for the most liked output, but also include an objective to reduce hallucinations in their example.After that it is possible to choose between all trade-off models that are both liked and have little hallucinations. As illustrated below.What can you do?- Identify all negative side effects for your use case- Collect a dataset used to classify whether the negative side effect occurred- Use these methods to find good models that suffer less from these side effects.Or contact me to set up a session on how to apply this in your business.#ai #research #RL #learning #rlhf

8

16,809

1,580 Comments