In our last reading group, dida’s Machine Learning Scientists discussed the paper ( from deepseek about the DeepSeek-R1 model and their approach of using reinforcement learning directly without prior fine-tuning.

OpenAI’s new o1 model boosts accuracy with chain-of-thought reasoning, excelling in complex tasks like IMO challenges (83% vs. 13% for GPT-4o) and competitive programming. Here you can read our summary for the o1 model: #AI #OpenAI #ML #innovation

NVIDIA Shrinks LLama3.1 8B to 4B with Pruning and Distillation. NVIDIA's latest research reduces LLama3.1 8B to 4B parameters by pruning 50% of its layers. Retrained with 40x fewer tokens, the model sees a 16% MMLU score boost. Read more: #AI #research.

We would like to present a paper on Physics-Informed Neural Networks (PINNs) that use physical laws to improve neural network training. It features a recent PyTorch implementation demonstrating PINNs by calculating gravity from thrown ball data. #ML.

We recently examined MambaVision, a new backbone combining Vision Transformers (ViT) with Mamba. This integration boosts scores and performance for tasks like object detection and instance segmentation. More details here: #machinelearning #tech #ML #AI.

Choosing the Right GPT Model for Your Business. For processing 4,500 documents per month:. ◾ It could either spend 126 EUR for the overall task or up to 7.150 EUR .◾ GPT-3.5 is 50 times cheaper than GPT-4 .◾ GPT-4o is still 5 times cheaper than GPT-4. #AI #LLM #GPT #TECH4ALL

Visual Instruction Tuning. Learn about "Visual Instruction Tuning": GPT-4 generates multimodal instruction tracking data. ViT-L/14 integration with Vicuna enables image comprehension and conversational applications. Read more:

dida will be at #ICLR2024 with two papers in Vienna this year. Check them out in case you are there.

We are thrilled to announce the launch of our brand-new website design!. We aimed to create a user-friendly, informative, and visually appealing platform for our blog readers, customers, and partners. Take a look at our new website design at

follow us on X for more interesting reading group posts: @dida_ML.

paper:

It seems that just by comparing predictions of a network to those from a weighted average of its past parameters you can produce information from almost nothing.

This week in our reading group we covered #BYOL. A very unusual form of self-supervised learning that seems to learn reasonable representations almost from nothing!

𝐆𝐢𝐭𝐇𝐮𝐛 𝐒𝐨𝐮𝐫𝐜𝐞 𝐂𝐨𝐝𝐞: 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐏𝐚𝐩𝐞𝐫: Follow us on X for more interesting #ML papers that you can practically apply to your projects.

The network's code and trained weights are publicly accessible, signaling its readiness for integration and further exploration in various real-world applications. #AI.

This approach allows it to manage occlusion and non-repeatable key points adeptly. It is trained end-to-end, enabling it to learn significant priors for pose estimation from a comprehensive annotated dataset, furthering its ability to reason about the 3D scene and assignments.

The paper presents SuperGlue, a neural network designed for efficient feature matching. It determines correspondences and filters out non-matchable points, utilizing a graph neural network and attention mechanisms to solve an optimal transport problem.

𝐒𝐮𝐩𝐞𝐫𝐆𝐥𝐮𝐞: In our last reading group we discussed a paper called "#SuperGlue: Learning Feature Matching with Graph Neural Networks"

Celebrating 5 years with the dida-conference!🎂 . We had an amazing day filled with machine learning (#ML) talks, panels, and workshops featuring top organizations. We captured some moments of it and turned them into a short video. 𝐯𝐢𝐝𝐞𝐨:

Link to the paper: For more summaries of interesting research papers, make sure to follow us here on 𝕏 so you don't miss out on any updates. #AI #ML.