Excalibur Data Systems’ Post

A new field promises to usher in a new era of using machine learning and computer vision to... | Click below to read the full article at The Digital Insider.

Deep learning and computer vision to classify land use and cover. The project utilizes a neural network model developed with PyTorch and trained to classify satellite images into 10 different land cover categories with high accuracy This work, carried out as part of the Data Science Academy's Artificial Intelligence Engineer program, used the EUROSAT dataset for satellite image analysis. Thanks to Hugging Face for access to the dataset, the EUROSAT team for their resources, and Data Science Academy The codes are available on github: https://lnkd.in/dpTim-VW #DeepLearning #ComputerVision #DataScience #ArtificialIntelligence

A new field promises to usher in a new era of using machine learning and computer vision to... | Click below to read the full article at The Digital Insider.

My latest project, where I personally collected a comprehensive dataset from LRH Hospital. This project not only includes detailed steps of data analysis and feature engineering but also explores the application of machine learning models like ANN (Artificial Neural Network Architecture) and CNN (Convolutional Neural Network) plus Embedded CNN and ANN. Additionally, I implemented and compared 8 baseline classification algorithms to establish robust benchmarks. This project serves as a fundamental guide for tackling real-world challenges and provides valuable insights for data science enthusiasts and professionals alike. I hope this work offers practical knowledge and inspires others to approach complex problems with data-driven strategies. #DataScience #MachineLearning #ArtificialIntelligence #Healthcare #DataAnalysis #FeatureEngineering #ANN #CNN #Classification #Algorithms

❓ Did you know there is a new-old deep learning architecture called Structured State Space Models? 👉 In part 2 of my series “Towards Mamba State Space Models for Images, Videos and Time Series” together with Towards Data Science we will explore state space models, how they build the foundation for Mamba, and how we can leverage some tricks in linear algebra making them tremendously fast. 💠 State Space Models (SSMs) are sequence models and can be described as recurrent model or as convolution model. 💠 They are inferred fast in their recurrent representation that scales linearly with the sequence length because they only depend on the current input and the previous state. 💠 They are trained fast because of their convolutional representation that can be parallelized. 💠 Structured state space sequence models aka. S4 use a certain structure on their state-matrices to make SSMs tremendously fast.

"Diving into Image Classification with CNNs: Exploring the CIFAR10 Dataset!" Excited to share my project on Image Classification with CNNs using the CIFAR10 dataset! 🚀 From splitting data to experimenting with network architectures and optimization techniques, this project dives deep into the world of computer vision. 🌟We define 10 tasks in this assignment: 1. Split data and build convolutional networks 2. Train convolutional networks 3. Add dropout layer 4. Add batch normalization 5. Try different initialization strategies 6. Try different nonlinearities 7. Add L2 regularization 8. Add data augmentation 9. Try a different architecture 10. Monitor the training procedure Check out the code on my GitHub and let's continue exploring the power of convolutional neural networks together! 🖥️🔍https://lnkd.in/daArBmVV #ImageClassification #CNNs #ComputerVision #GitHub #MachineLearning

Hey Connections ! I Completed Handwritten Digit Classification Project with ANN using TensorFlow!🚀 Thrilled to share the success of my recent project on digit classification. I developed a Convolutional Neural Network (CNN) from scratch for the MNIST handwritten digit dataset. 🖋️ Project Highlights: 1)Dataset: MNIST consists of 60,000 small grayscale images of handwritten digits (0 to 9). 2)Model Development: Built a robust CNN model to classify digits. 3)Evaluation: Developed a test harness to evaluate model performance. 4)Improvements: Explored enhancements to boost learning and capacity. 5)Final Model: Achieved impressive accuracy and saved the model for predictions. Excited to continue my journey in AIML and Data Science ! 🤖📊 #MachineLearning #DeepLearning #DataScience #ArtificialIntelligence

ML for ML <<TpuGraphs: A Performance Prediction Dataset on Large Tensor Computational Graphs” (presented at NeurIPS 2023)>>

Reset Gate in GRUs: A Deep Dive into Its Mathematics, role and functionality What is the Reset Gate? In GRUs, the reset gate is a crucial component that determines how much of the past information should be forgotten. This gate allows the model to decide which parts of previous hidden states are irrelevant to the current computation, thereby optimizing memory usage and improving model performance. Mathematical Intuition The reset gate, denoted as r_t at time step t, is calculated using the sigmoid function: r_t = sigmoid(W_r * x_t + U_r * h_{t-1} + b_r) - sigmoid: Sigmoid activation function, ensuring output values between 0 and 1. - W_r and U_r: Weight matrices for the input x_t and previous hidden state h_{t-1}. - b_r: Bias vector. The reset gate modulates the influence of previous hidden states on the candidate hidden state: h_tilde_t = tanh(W * (r_t ⊙ h_{t-1}) + U * x_t) Here, ⊙ represents element-wise multiplication. The reset gate thus controls how much past information is considered when computing the candidate hidden state. Role and Importance The reset gate's ability to filter out unnecessary historical data is vital for capturing relevant dependencies in sequences. This mechanism helps GRUs effectively manage long-term dependencies without suffering from the vanishing gradient problem that affects traditional RNNs. Key Takeaways - Selective Memory: By deciding what past information to discard, the reset gate enhances model efficiency. - Computational Efficiency: GRUs, with fewer parameters than LSTMs, offer faster training times while maintaining performance. - Versatility: Ideal for applications like language modeling, speech recognition, and time-series forecasting. Understanding these components enhances our capability to leverage GRUs in various deep learning applications. 🌐 Follow & Connect with me: GitHub: https://lnkd.in/dvtE-xRR Instagram: https://lnkd.in/gGhbJSCg X: https://x.com/__kanhaiya__ #DeepLearning #MachineLearning #AI #NeuralNetworks #GRU #ResetGate

📢🚀 🖥 Exciting News! 🖥🚀📢 I am thrilled to announce the publication of our latest research paper in the Journal of Systems Architecture: "Parallel GEMM-based convolutions for deep learning on multicore ARM and RISC-V architectures". 🏅 📄Read the full paper here: https://lnkd.in/drH9dwYP 🔍 In this work, we introduce high-efficiency 📈, multi-threaded implementations of three GEMM-based convolution algorithms optimized for multicore processors using ARM and RISC-V architectures 🔋. These implementations are integrated into a library named CONVLIB, which features several distinct characteristics: 1️⃣ Scripts that automatically generate a crucial component of GEMM, known as the micro-kernel, which is typically coded in assembly language. 2️⃣ An enhanced analytical model that automatically adjusts the algorithms to match the cache architecture. 3️⃣ The capability to dynamically choose among four hyper-parameters: micro-kernel, cache parameters, parallel loop, and GEMM algorithm without needing to recompile the library. 4️⃣ A driver designed to identify the optimal hyper-parameters. Furthermore, we provide an in-depth performance analysis of the convolution algorithms on five different ARM and RISC-V processors 📊. 📚 The code is available on GitHub: https://lnkd.in/d4NDKWZq A big thank you to my co-authors Héctor Martínez Pérez, Adrián Castelló Gimeno, and Enrique S. Quintana Ortí for their work. Looking forward to the impact this research will have on the development of high-performance deep learning applications! #Research #DeepLearning #GEMM #MulticoreProcessing #ARM #RISCV #CONVLIB #HighPerformanceComputing #AI #MachineLearning #ScientificPublication #JournalOfSystemsArchitecture