Moritz Wildenhain

AnchorGT: A Novel Attention Architecture for Graph Transformers as a Flexible Building Block to Improve the Scalability of a Wide Range of Graph Transformer Models Transformers have taken the machine learning world by storm with their powerful self-attention mechanism, achieving state-of-the-art results in areas like natural language processing and computer vision. However, when it came to graph data, which is... https://lnkd.in/eedeAnTa #AI #ML #Automation

2

Retrieval Augmented Generation (RAG) is a straightforward concept that any company can implement with little effort. You can think of RAG as an NN-based document summarization, like merging the results of a simple elastic search/apache solr query with a text summary. The original concept involved retraining the Llm to have a similar latent space, however in applied concepts the retraining has been dropped and replaced with a vector database and prompt engineering that can be paired with any Llm that exists. Based on Meta's LLAMA3, LlamaIndex, Qdrant, Ollama, Angular, and FastAPI, I built the RAG demo about funding lines for universities of applied sciences in hesse. The demo is available at (not well optimized for mobile use): https://lnkd.in/eBYQwKh4 This demo is a weekend project when the little ones were asleep, so it took me less than 3 days. Imagine your company creating a bot like this with more manpower and effort. Best of all, it's completely free and costs you nothing. Your data stays in your company without any security or GDPR violations. P.S.: On my own behalf: I also work as a speaker and consultant on AI topics. Through my research in context-aware fault diagnosis in industrial production lines, I know the state of the art in AI and have implemented various AI use cases myself using different neural networks under industry constraints. If you need some advice, just get in touch with me. I'm looking forward to talking with you about self-built solutions without buying expensive cloud services.

11

2 Kommentare

For those who are interested in attending the Coding.Waterkant event but are unable or unwilling to commit to the full week, I have good news: the Monday evening session is open to everyone! Join us for a fascinating keynote speech "LAION and its Role in Open Source AI" by Christoph Schuhmann and take the opportunity to learn more about us and the unique Coding.Waterkant format. I encourage you all to mark your calendars and spread the word. I hope to see a strong turnout and many familiar faces there! #MachineLearning #AI #CodingWaterkant24 #opensource

14

1 Kommentar

🚀 Control Systems in AI 🚀 Control systems often face challenges with high-dimensional, noisy data. The ACDF algorithm, improves the learning of action-dependent latent states, and ensures more accurate representation and control in complex environments. Paper 🔗 - https://lnkd.in/gi82aUqn 💡 Keen on AI breakthroughs? DM us to explore opportunities!

9

Till Adam (KDAB) and Florian Gilcher (Ferrous Systems GmbH) discussed features, pitfalls and challenges of Rust at Oxidize 2024. On top of this, check out this jointly written whitepaper about hybrid #RustLang and #Cpp applications: https://lnkd.in/ey4CuCme #CPlusPlus

31

I like Agnieszka Mikołajczyk-Bareła's paper reviews because she really has an eye for relevancy in the field of LLMs: 💡 As someone with a PhD she is familiar with many scholars and academics so when a paper comes out, like xLSTM, she can recognise relevant academics. Like Sepp Hochreiter in this case who was one of the original authors of the original LSTM paper back in the 90's which had a profound impact on the field of ML. 💡She explains papers in a concise and comprehensive way, so that you don't need to be spending hours reading the actual paper, you can spend 5 minutes reading her post in order to understand if it makes sense for you to read the entire thing. My thoughts on xLSTM: 📌 xLSTM definitely appears as a cheaper alternative to the Transformer based models (which is most LLMs at this point in time) 📌 My intuition is that Transformer-based models will outperform xLSTM on most benchmarks, however, I'm looking forward to the first industrial applications of xLSTM, as well as review/benchmark papers comparing it to the most popular LLMs today. 📌 Overall impression: So far, it looks like the advantage of xLSTM technology is price, so I believe xLSTM might have some industrial applications. 📌 However, I believe that for the forseeable future, transformer-based models will continue to be the base for most chatbot-like applications. Let me know your thoughts ✨

13

2 Kommentare

The minimum cut problem is a structural question about the connectivity of a graph and one of the key problems in algorithmic graph theory, with many practical applications. Read about the first deterministic, nearly-linear algorithm that works for general graphs →https://goo.gle/3xFQBY3

933

7 Kommentare

Exciting Spring Update from OpenAI! 🚀 The launch of the new GPT-4o model marks a groundbreaking moment in AI technology. With twice the speed and a 50% cost reduction compared to its predecessor, GPT-4o also offers a great five times higher rate limits. It’s truly an all-in-one model capable of handling text, video, and audio, supporting an impressive 50 languages. The fascinating part: The range of emotions and expressions GPT-4o can achieve seems almost unbelievable! It feels like we're stepping into the world of "Her," with AI becoming more natural and collaborative than ever before. OpenAI continues to astound me! #GPT4o #OpenAI

7

Debug like a Human: A Large Language Model Debugger via Verifying Runtime Execution Step-by-step Li Zhong, Zilong Wang, Jingbo Shang : https://lnkd.in/eM2WwSAm #ArtificialIntelligence #DeepLearning #MachineLearning

1

#3: Choose Your Model Architecture Selecting the appropriate model architecture is crucial. Different architectures excel at different tasks. Convolutional Neural Networks (CNNs) #CNNs #DeepLearning Use Case: Best for image-related tasks such as recognition and classification. Example Architectures: VGG, ResNet, EfficientNet. Example: Use ResNet for recognizing different species in wildlife images due to its ability to handle deep networks. Recurrent Neural Networks (RNNs) #RNNs #SequenceModeling Use Case: Suitable for sequential data like text or time series data. Example Architectures: LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit). Example: Use LSTM for sentiment analysis in text data as it can capture long-term dependencies. Transformers #Transformers #NLP Use Case: Excellent for natural language processing tasks. Example Architectures: BERT, GPT-3, T5. Example: Use GPT-3 for text generation tasks due to its ability to generate human-like text. Other Architectures #GraphNeuralNetworks #Autoencoders Graph Neural Networks (GNNs): Used for tasks involving graph-structured data like social networks. Example: Use GNNs to predict user interactions on social media platforms. Autoencoders: Good for unsupervised tasks like anomaly detection or data compression. Example: Use autoencoders for detecting fraudulent transactions in financial data. #EncephAI #ArtificialNeurons #ArtificialIntelligence #AI #MachineLearning #AIModels

1

Hugging Face has unveiled LeRobot, an open-source ML model designed to enhance robotic capabilities and usability. Built on the PyTorch framework, LeRobot combines cutting-edge techniques with real-world applications, focusing on reinforcement and imitation learning. With pretrained models, example datasets, and simulated scenarios, LeRobot simplifies the development process for developers and researchers alike. Discover more about how LeRobot is transforming robotics and explore the future of innovative robotic solutions. 🌟 🔗 Read the full blog post - https://lnkd.in/gYK2TUFX #Robotics #MachineLearning #OpenSource #AI #Innovation #HuggingFace

2

Mitko Vasilev Muzaffer Kal - instead of measuring performance against flops, I would like to see it measured (and optimized) again memory transfers or ideally total energy consumption. Here's my reasoning.... The starting assumption is that compute has grown much faster than memory or memory bandwidth. This is described as the "memory wall". So if the memory bandwidth is the limitation, you can often do more computation for 'free' without slowing down the overall AI task. An example of this is convolution networks where the memory bandwidth is trivial as the convolution weights are reused to a very large degree across the whole image. Compare that to the fully connected layers which do an expensive operation to move the weight from DRAM to the die, do one multiply add then throw it away. The convolution is far better from a cost / benefit analysis that the fully connected layer. Measuring the FLOPS without considering the reuse of weights doesn't reflect that. The memory transfers or ideally energy is a better proxy. For a small chip, the energy is very directly how much you can run the model before you exhaust your energy budget. For the datacenter, the energy is a limiting factor on how much you can push the performance before the cooling can't keep up.

8

2 Kommentare

In his introduction to Neural Speed, Benjamin Marie reviews the library's main optimizations, shows how to use it and benchmark its inference throughput, and compares its performance to llama.cpp's.

33