About
Articles by Balázs
Activity
2K followers
-
Balázs Fehér shared thisI initially prepared this analysis for my own reference, but it shaped up so nicely I decided to share more widely. It could be a useful read for anyone interested in this space—whether you're an AI/ML Engineer, Data Scientist, or someone in a leadership position. I developed the framework and outlined the core components for analysis, then iterated with Gemini for the deep-dive research and final composition. (Honestly, I have neither the patience nor the talent to write it up this well! 😄) While you should expect some errors and omissions, I hope it provides a good overview of the key considerations for modern RAG systems. Enjoy the read!RAG: An Architectural Review and Strategic Outlook for 2025RAG: An Architectural Review and Strategic Outlook for 2025
-
Balázs Fehér shared thisExcited to kick off Nvidia GTC 2024 with the first session: "Deploying, Optimizing, and Benchmarking Large Language Models with Triton Inference Server." In a typical setup where multiple users are served simultaneously, requests are processed in batches. This process traditionally forces us to wait for the next available slot to batch new requests, even if parts of the current batch are already completed. The introduction of the iterative batch scheduler improves this approach by allowing for the immediate utilization of freed-up slots, enhancing GPU efficiency and significantly boosting the user experience with quicker response times. #GTC2024 #llms #generativeai
-
Balázs Fehér shared thisThis article discusses common myths about the brain. Overcoming myth number two would lead to important new research directions towards AGI. "Myth number two is that your brain reacts to events in the world. Supposedly, you go through your day with parts of your brain in the off position. Then something happens around you, and those parts switch on and “light up” with activity. Brains, however, don’t work by stimulus and response. All your neurons are firing at various rates all the time. What are they doing? Busily making predictions. In every moment, your brain uses all its available information (your memory, your situation, the state of your body) to take guesses about what will happen in the next moment. If a guess turns out to be correct, your brain has a head start: It’s already launching your body’s next actions and creating what you see, hear, and feel. If a guess is wrong, the brain can correct itself and hopefully learn to predict better next time. Science, like your brain, works by prediction and correction. Scientists use their knowledge to fashion hypotheses about how the world works. Then they observe the world, and their observations become evidence they use to test the hypotheses. If a hypothesis did not predict the evidence, then they update it as needed."
-
Balázs Fehér posted this“Whenever a theory appears to you as the only possible one, take this as a sign that you have neither understood the theory nor the problem which it was intended to solve.” ― Karl R. Popper
-
Balázs Fehér shared thisUnfortunately, most of what we need to know about the virus isn't known. Yet.A Comic Strip Tour Of The Wild World Of Pandemic ModelingA Comic Strip Tour Of The Wild World Of Pandemic Modeling
-
Balázs Fehér shared thisWe thought beating top #StarCraft players is still a few years away. It is now at #DeepMind.
-
Balázs Fehér shared thisA mistaken view of AI is that it is about predicting sense-impressions. The correct view is that it is about understanding reality, of which we only ever experience an infinitesimal portion. Finding the rules for predicting the outcome of every possible experiment depends on ever more accurate explanatory theories. And so, accurately rendering an environment depends on understanding its physics. This brilliant work presents a neural net architecture capable of rendering environments from arbitrary viewpoints.
-
Balázs Fehér shared thisRoyal Mail has one of the most mature data science practice in the UK. A great place to develop or start your career. #datascience #hackathonBalázs Fehér shared thisWant to hear a secret? We’re into Data Science too! We're hosting an exciting Charity Hackathon on Thursday 24th May. Be quick, applications close Monday 21st May! Sign up here: https://goo.gl/Ftjkxq #datascience
-
Balázs Fehér liked thisBalázs Fehér liked thisToday we launch Olea Venture Partners, with a first closing at €60 million. We are proud to be backed by the European Investment Fund (EIF), alongside a select group of private LPs from leading organizations across pharmaceuticals, healthcare, financial services, and industry. My partner Lampros Kourtis and I have built and scaled companies in healthcare across international markets. At Olea, we invest and work alongside founders from the earliest stages through scale to deliver real-world outcomes that last. Healthcare and life sciences are at an inflection point. AI, the pace of scientific discovery, aging demographics, and increasing pressure on health systems are reshaping the landscape. This is the moment to back companies that will define the next decade of healthcare and biotech innovation. On a personal note, to those I have worked with over the years, I look forward to reconnecting and driving healthcare forward together. Learn more: https://oleavp.com/Olea Venture Partners | Healthcare & Life Sciences VCOlea Venture Partners | Healthcare & Life Sciences VC
-
Balázs Fehér liked thisThank you for the insightful discussions and engaging conversations 😊 Zsombor Nánai-Nagy Evangelia (Lia) Valavani Oguz Ergun Juan Carlos Maltez Romillo Hülya Peker Sandra Stanimirov Chiara Baldelli Alexis Vrachimis Boyan NedyalkovBalázs Fehér liked thisThe DANTE Fusion team attended the Balkan Congress on Nuclear Medicine 2026 in Sofia last week. The conversations confirmed what we are already seeing in the market. Demand for nuclear medicine is growing across the Balkans, but isotope supply is not keeping pace. Lu-177 and Ac-225 are already constrained. Interest in next-generation isotopes like Tb-161 and Cu-64/67 is building fast. Clinicians, regional radiopharmaceutical companies, and larger players such as Siemens Healthineers and Curium Pharma all identified the same structural problem: production and distribution of isotopes remains fragmented, unreliable, and insufficiently localised. Without reliable regional supply, building treatment programmes that clinicians and patients can depend on remains very difficult. We also spoke with clinicians from North Macedonia on the practical challenges of expanding nuclear medicine in smaller healthcare systems. These conversations matter to us. Without a stable and accessible isotope supply, the clinical and infrastructural investment required to build these capabilities cannot be sustained. This is one of the core challenges we are working to address. The congress reinforced the direction we are taking. We have a lot of work ahead. #NuclearMedicine #Theranostics #Radioisotopes #HealthcareInnovation #Balkans #DeepTech #Fusion #Innovation
-
Balázs Fehér reacted on thisBalázs Fehér reacted on thisHappy Birthday Aleksei Zolotarev 💥 It was such a pleasure reconnecting with the Next Step Fusion team at FusionX Global 😊 We are excited to support your journey and look forward to continuing our long standing collaboration as the European roadmap unfolds 🚀 And we have to say, the fusion stickers were a highlight 🔥 Anyone who wants one should definitely reach out to Alexei Zhurba 😄
-
Balázs Fehér liked thisBalázs Fehér liked thisIntroducing WorldGen, new research that can generate walkable 3D worlds from simple text prompts like “medieval village town square.” Led by director of AI research Rakesh R and product manager Mahima Gupta, this Reality Labs team is hard at work cracking 3D generation challenges. Learn more: ➡️ https://lnkd.in/g2bmuDJ9 #TechInnovation #AI #MetaCareers #LifeatMetaWorldGen — Text to Immersive 3D Worlds | Reality LabsWorldGen — Text to Immersive 3D Worlds | Reality Labs
-
Balázs Fehér reacted on thisBalázs Fehér reacted on this𝗜’𝘃𝗲 𝗯𝗲𝗲𝗻 𝗮 𝗹𝗶𝘁𝘁𝗹𝗲 𝗾𝘂𝗶𝗲𝘁 𝗮𝗯𝗼𝘂𝘁 𝘀𝗼𝗺𝗲 𝘃𝗲𝗿𝘆 𝗯𝗶𝗴 𝗻𝗲𝘄𝘀. There’s been 𝙖 𝙣𝙚𝙬 𝙖𝙧𝙧𝙞𝙫𝙖𝙡 𝙞𝙣 𝙩𝙝𝙚 𝙛𝙖𝙢𝙞𝙡𝙮! I held off sharing because, in truth, the important thing was making sure mum, dad, and the little one were safely home, well supported, and getting settled in ... before sleep becomes a mythical thing of the past! So ... I’m now a grandparent! 😮 Please welcome 𝙊𝙡𝙞𝙫𝙚𝙧 𝘼𝙙𝙖𝙢 𝙇𝙤𝙣𝙜, born on 𝟮𝗻𝗱 𝗡𝗼𝘃𝗲𝗺𝗯𝗲𝗿 𝟮𝟬𝟮𝟱, weighing in at 𝟵𝗹𝗯 𝟲𝗼𝘇 (a solid unit!) I had the privilege of holding him for the first time last weekend ... there’s something about that moment, tiny little fingers and toes, short-quiet breaths, and a scrunched up windy face! He is perfect. His parents are incredible Toby Long. We're all very proud of them all. ❤️ To new beginnings ... tiny socks .. lots of love that appears out of nowhere!
-
Balázs Fehér liked thisBalázs Fehér liked thisAI is moving at eye-watering speed, and standing still isn’t an option. ⏳ The pressure on business leaders is real. Knowing where to start and how to succeed in today’s market begins with understanding where you are on your AI journey. That’s why our founder, Steve Cracknell, created a short survey to help organisations quickly assess their position on the AI curve and what that means for their next steps. ⚡ Take the AI Readiness Survey → https://lnkd.in/ek6Yq2uc In his latest article, Steve shares what inspired the survey and why clarity matters now more than ever. While some teams are still evaluating their options, others are already using AI to drive execution, unlock growth, and gain serious competitive advantage. 📖 Read the article → https://lnkd.in/ewzzApH5 #AIReadiness #DigitalTransformation #LeadershipTools #EnterpriseAI #StrategyExecution
Experience & Education
View Balázs’s full experience
See their title, tenure and more.
or
Licenses & Certifications
Publications
-
Auditory event-related potentials reflect dedicated change detection activity for higher-order acoustic transitions
Biological psychology
The processing of auditory changes at cortical level relies partly on dedicated change-detectors whose activity is reflected in the elicitation of the N1 and P2 event-related potentials (ERPs). In previous studies, N1 and P2 have been found only for first-order frequency transitions (i.e. constant-to-glide) but not for higher-order transitions (i.e. glide-to-constant). We tested whether this asymmetry is due to the complete lack, or the smaller number of dedicated higher-order change detectors…
The processing of auditory changes at cortical level relies partly on dedicated change-detectors whose activity is reflected in the elicitation of the N1 and P2 event-related potentials (ERPs). In previous studies, N1 and P2 have been found only for first-order frequency transitions (i.e. constant-to-glide) but not for higher-order transitions (i.e. glide-to-constant). We tested whether this asymmetry is due to the complete lack, or the smaller number of dedicated higher-order change detectors compared to first-order change detectors by recording ERPs to constant-to-glide and glide-to-constant frequency transitions within pure and complex tones. For constant-to-glide transitions ERP amplitudes increased with the rate of frequency change and spectral complexity. Importantly, for glide-to-constant transitions, N1 was elicited, even though only for spectrally rich tones when the frequency-change rate was fastest. Thus, the asymmetry in auditory change-related N1 elicitation is attributable not to the lack of higher-order change detectors, but to their relatively low number.
Other authors -
-
Stimulus-focused attention speeds up auditory processing
International Journal of Psychophysiology
Stimulus-focused attention enhances the processing of auditory stimuli, which is indicated by enhanced neural activity. In situations where fast responses are required, attention may not only serve as a means to gain more information about the relevant stimulus, but it may provide a processing speed gain as well. In two experiments we investigated whether attentional focusing decreased the latency of the auditory N1 event related potential. In Experiment 1 slowly emerging, soft (20dB sensation…
Stimulus-focused attention enhances the processing of auditory stimuli, which is indicated by enhanced neural activity. In situations where fast responses are required, attention may not only serve as a means to gain more information about the relevant stimulus, but it may provide a processing speed gain as well. In two experiments we investigated whether attentional focusing decreased the latency of the auditory N1 event related potential. In Experiment 1 slowly emerging, soft (20dB sensation level) sounds were presented in two conditions, in which participants performed a sound-detection task or watched a silent movie and ignored the sounds. N1 latency was shorter in the sound-detection task in comparison to the ignore condition. In Experiment 2 we investigated whether the attentional N1 latency-decrease was caused by a frequency-specific attentional preparation or not. To this end, tone sequences were presented with a single tone frequency or with four different frequencies. N1 latency was shorter in the sound-detection task in comparison to the ignore condition regardless the number of frequencies. These results suggest that stimulus-focused attention increases stimulus processing speed by generally increasing sensory gain.
Other authors -
-
Attention decreases auditory N1 latency
Frontiers in Neuroscience Conference Abstract: 13th Conference of the Hungarian Neuroscience Society (MITT)
Attention related changes can be observed even approximately 100 ms after stimulus onset, in the time window of the N1 event-related potential. Previous studies on attentional N1 deflection mainly reported amplitude enhancement as a function of attention. It seems plausible that in cases where attentional demands are high, stimulus detection gets faster. According to this assumption attention effects are not only reflected in the enhancement of the N1 amplitude, but also in the decrease of the…
Attention related changes can be observed even approximately 100 ms after stimulus onset, in the time window of the N1 event-related potential. Previous studies on attentional N1 deflection mainly reported amplitude enhancement as a function of attention. It seems plausible that in cases where attentional demands are high, stimulus detection gets faster. According to this assumption attention effects are not only reflected in the enhancement of the N1 amplitude, but also in the decrease of the N1 latency. We tested this hypothesis using soft tones (20 dB SL) with different rise times in attended and unattended conditions. We found that the latency of the N1 decreases for all the tones in the attended condition compared to the unattended condition. As the stimuli did not differ in any physical attributes, we interpret the differences between the attended and unattended conditions as an effect of attention.
Other authors -
-
Attentional N1 latency modulation: a frequency general effect
Frontiers in Neuroscience Conference Abstract: 13th Conference of the Hungarian Neuroscience Society (MITT)
Attention dependent changes can be observed even in the time window of the N1 component. In a previous experiment we demonstrated that in a paradigm using low intensity tones attention not only enhances the amplitude of the N1 but also decreases its latency. In the present experiment we tested whether this early attentional effect is a general one or stimulus specific, using soft tones (20 dB SL) with four different frequencies. Our results suggest that in an attended condition the mechanism…
Attention dependent changes can be observed even in the time window of the N1 component. In a previous experiment we demonstrated that in a paradigm using low intensity tones attention not only enhances the amplitude of the N1 but also decreases its latency. In the present experiment we tested whether this early attentional effect is a general one or stimulus specific, using soft tones (20 dB SL) with four different frequencies. Our results suggest that in an attended condition the mechanism behind the decrease of the N1 latency is a general one, which means that attention increases the general sensitivity of the auditory system, thereby improving the processing of the attended stimuli.
Other authors -
Courses
-
Information Management Solution Architecture
-
-
Issues of Statistical Significance Testing
-
-
Statistik und Grafikerstellung mit R
-
Projects
-
Kaggle - Allstate Claims Severity
Allstate is currently developing automated methods of predicting the cost, and hence severity, of claims. In this challenge, I have created an algorithm which accurately predicts claims severity. With the solution it is possible demonstrate insights into better ways to predict claims severity to ensure a worry-free customer experience.
In this project I have:
• developed a solution with deep artificial neural nets to predict insurance claim severity
Key technologies used:
•…Allstate is currently developing automated methods of predicting the cost, and hence severity, of claims. In this challenge, I have created an algorithm which accurately predicts claims severity. With the solution it is possible demonstrate insights into better ways to predict claims severity to ensure a worry-free customer experience.
In this project I have:
• developed a solution with deep artificial neural nets to predict insurance claim severity
Key technologies used:
• Amazon AWS, Python, Keras, Deep Neural Nets
Data:
• Historical data (188,000+ claims) -
Deep Learning for Autonomous Vehicle Control
See projectIn this project I have:
• developed an image recognition system with deep convolutional neural nets to recognize traffic signs
Key technologies used: Python, Keras, Deep Learning, PCA, Convolutional Neural Nets, Raspberry Pi
-
Reinforcement Learning for Self Driving Cab
See projectIn this project I have:
• applied reinforcement learning techniques for a self-driving agent in a simplified world to aid it in effectively reaching its destinations within an allotted time.
• implemented a Q-Learning algorithm for the self-driving agent
• improved upon the Q-Learning algorithm to find the best configuration of learning and exploration factors to ensure the self-driving agent is reaching its destinations with consistently positive results
Key technologies used:…In this project I have:
• applied reinforcement learning techniques for a self-driving agent in a simplified world to aid it in effectively reaching its destinations within an allotted time.
• implemented a Q-Learning algorithm for the self-driving agent
• improved upon the Q-Learning algorithm to find the best configuration of learning and exploration factors to ensure the self-driving agent is reaching its destinations with consistently positive results
Key technologies used: Python, Reinforcement Learning, Q-learning -
Customer Segmentation for Wholesale Retailer
In this project I have created customer segmentation to help a wholesale grocery distributor determine which changes will benefit their business.
In this project I have:
• used unsupervised techniques to see what sort of patterns exist among existing customers, and what exactly makes them different
Key technologies used: Python Scikit-learn, ICA, PCA, K-means clustering
-
Building a Student Intervention System
A local school district had a goal to reach a 95% graduation rate by the end of the decade by identifying students who need intervention before they drop out of school.
In this project I have:
• modeled the factors that predict how likely a student is to pass their high school final exam, by constructing an intervention system that leverages supervised learning techniques
• identified the most effective model that uses the least amount of computation costs to save on the budget -
Tweeting Recurrent Neural Network
See projectIn this project I have:
• used Twitter's API to access tweets about the Star Wars Force Awakens movie
• trained a recurrent neural net to generate tweets about the movie
Key technologies used: Python, R, Recurrent Neural Nets -
Kaggle - Prudential Life Insurance Assessment - Making the Buying of Life Insurance Easier
-
In this project I have:
• developed a predictive model that accurately classified risk using an automated approach
• experimented with both different neural network designs as well as extreme gradient boosting, and blending the two approaches to achieve the most accurate results
The results of the challenge helped Prudential better understand the predictive power of the data points in the existing assessment, enabling them to significantly streamline the process.
Key…In this project I have:
• developed a predictive model that accurately classified risk using an automated approach
• experimented with both different neural network designs as well as extreme gradient boosting, and blending the two approaches to achieve the most accurate results
The results of the challenge helped Prudential better understand the predictive power of the data points in the existing assessment, enabling them to significantly streamline the process.
Key technologies used: R, Python, XGBoost, Neural Nets
-
Kaggle - Rossmann - Forecast Sales Using Store, Promotion, and Competitor Data
-
See projectRossmann operates over 3,000 drug stores in 7 European countries. Rossmann store managers were tasked with predicting their daily sales for up to six weeks in advance.
In this project I have:
• designed and implemented a complex neural network architecture to forecast sales figures for each individual store
• performed data exploration, data cleansing, and data transformation, as well as looking after the incorporation of additional external data sources to aid in better…Rossmann operates over 3,000 drug stores in 7 European countries. Rossmann store managers were tasked with predicting their daily sales for up to six weeks in advance.
In this project I have:
• designed and implemented a complex neural network architecture to forecast sales figures for each individual store
• performed data exploration, data cleansing, and data transformation, as well as looking after the incorporation of additional external data sources to aid in better forecasting capability of the network
• experimented with a large number of hyperparameter settings to find the optimal solution for this particular problem
• presented my results to a group of data scientists (30+)
Key technologies used: Python, IPython, Neural Networks, Scikit-learn, Keras, Theano, Azure
-
Deep Learning OCR
-
Developed an optical character recognition software to identify individual Tibetan characters. I used deep neural networks (i.e. deep learning).
Key technologies used: Octave, Python, Deep Learning
Honors & Awards
-
Research Grant
DAAD-MÖB
I was awarded a research grant to conduct research on auditory cognition at Lepizig University.
Languages
-
English
Full professional proficiency
-
German
Professional working proficiency
-
Hungarian
Native or bilingual proficiency
-
Tibetan
Professional working proficiency
-
Chinese
Elementary proficiency
Recommendations received
4 people have recommended Balázs
Join now to viewWebsites
Other similar profiles
-
Pavithra Rajendran, Ph.D.
Pavithra Rajendran, Ph.D.
Great Ormond Street Hospital for Children NHS Foundation Trust
Explore more posts
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More