DCT, DFT, Transformation, Noise Reduction: The patterns in an image and how they affect frequency spectrum and its direction which leads to designing noise reduction methods using DCT and DFT methods. This phenomenon has been shown on both natural real-world images and synthetic images for better comprehension. Also, the usage of filtering high-frequency for compression has been demonstrated.
Morphological Operation, Hit-or-Miss, Textural Feature, Soft LBP, HOG, SVM and kNN: Discussion around how morphological operations act on images and different operators such as hit-or-miss. This post ends with implementation of a machine learning model for Persian handwritten digit classification using textural and geometrical feature engineering and models such as Random Forest, kNN, SVM, etc. In the end, Confusion matrix is being used to evaluate methods.
Movie Review Sentiment Analysis: We start of by preparing the dataset, from cleaning to lemmatization; then we extract features using Bag of Words, BeRT embeddings, TF-IDF and Word2Vec methods. We train machine learning models such as SVM and Naive Bayes to learn this data and in the end, we report metrics such as F1, precision, recall and ROC_AUC curves.
Imbalance Learning and Evaluation using AdaBoost, RUSBoost, SMOTEBoost, RBBoost and ANOVA Measure: In this post, we dive deep into Imbalance Learning pipeline from data preparation to model fitting. We discuss normalization and K-fold validation for data preparation, then we define and discuss Ensemble Learning methods, bagging, boosting and implement them in numpy from scratch. These methods include, AdaBoost, AdaBoostM2, SMOTEBoost, RBBoost. Then we compare accuracies and precision-recall bars and AUC_ROC curves. In the end, to show statistical differences between models, we implement and report ANOVA test.
Perceptron With Hebb Rule For Linear Data: In this post, we first explain then implement a basic Perceptron model using numpy and update its weight using Hebb rule for a linearly separable data. In the end we compare the effect of choosing Tanh and Sigmoid as the activation function.
MLP With Backpropagation For Non-Linear Data (XOR): Here, we extend previous model that only worked for linear data, to a Multi-Layer Perceptron (MLP) to fit non-linear data and we update its weights using backpropagation. This model has been explained and implemented in numpy.
Kohonen Self Organizing Maps For Learning MNIST: In this post, we discuss the Kohonen Self-Organizing Maps as a different type of neural networks that can fit non-linear data. We test this by implementing the model in numpy and test it on MNIST dataset. In the end, we argue the method for computing accuracy and the best model size and its symmetric structure.
Training Tensorflow 1.x Model for MNIST: This post just contains a common implementation of TF1.x model for MNIST dataset, but this time we explain for a bit why the particular architecture/structure has been used in comparison to older methods.
4. Computer Vision (CV)
Phase Amplitude Combination, Hybrid Images and Cutoff Effect On It: Have you ever seen those images that look different to different people? Or more closer to this post, when you see them in different distances, you see different image? Well, this has to do with the Phase and Amplitude of combination of two images. By combining low-frequency information of one image and high-frequency of the other image, we can generate Hybrid Images.
5. Data Mining (DM)
Statistical Evaluation Metrics: This post compares many of statistical metrics for evaluating models with same result but different scenarios. We try to bring real-world examples here too. Some of these metrics are: ROC-AUC vs F1 and PR-Curve; P@1, P@10, P@K, MAP, MRR, NDCG, r-Prec, bPref; F-Macro and F-Micro; Miss and Fallout; Specificity, Sensitivity, PPV and NPV in Medical Predictor Assessment.