Study notes, AI updates and paper summaries by Sahaj Saliya.https://blog.sahaj.si/https://blog.sahaj.si/2025/09/23/cv-08-augmentation/https://blog.sahaj.si/2025/09/23/cv-08-augmentation/Label-preserving random transforms (flips, crops, rotation, colour jitter) with torchvision.transforms.v2, applied consistently to images and boxes.Tue, 23 Sep 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/09/16/cv-07-segmentation-unet/https://blog.sahaj.si/2025/09/16/cv-07-segmentation-unet/Semantic segmentation as per-pixel classification, the U-Net encoder–decoder with skip connections, Dice/IoU, and a tiny U-Net trained on synthetic shapes.Tue, 16 Sep 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/09/08/cv-06-detection-yolo/https://blog.sahaj.si/2025/09/08/cv-06-detection-yolo/Detection as boxes plus classes, IoU, non-maximum suppression, YOLO's single-pass grid idea, and running a pretrained torchvision detector on a NASA photo.Mon, 08 Sep 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/09/01/cv-05-transfer-learning/https://blog.sahaj.si/2025/09/01/cv-05-transfer-learning/Reusing an ImageNet-pretrained ResNet-18: frozen features plus a new linear head versus training from scratch, on the 8×8 digits upsampled.Mon, 01 Sep 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/08/25/cv-04-vgg-resnet/https://blog.sahaj.si/2025/08/25/cv-04-vgg-resnet/VGG's stacks of 3×3 convolutions, why very deep plain networks got worse, and ResNet's skip connections, with a residual block checked in code.Mon, 25 Aug 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/08/15/cv-03-lenet-alexnet/https://blog.sahaj.si/2025/08/15/cv-03-lenet-alexnet/The two early milestones, LeNet-5 for digits and AlexNet for ImageNet: their layers, parameter counts, and what changed between them.Fri, 15 Aug 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/08/08/cv-02-opencv/https://blog.sahaj.si/2025/08/08/cv-02-opencv/The OpenCV operations I use most: colour conversion, resizing, Gaussian blur, Canny edges, thresholding and contours.Fri, 08 Aug 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/08/01/cv-01-image-basics/https://blog.sahaj.si/2025/08/01/cv-01-image-basics/Images as arrays: height × width × channels, uint8 vs float, RGB vs grayscale, cropping, and the channel-order and range traps between libraries.Fri, 01 Aug 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/07/23/nlp-10-text-classification/https://blog.sahaj.si/2025/07/23/nlp-10-text-classification/Multi-class topic classification on 20 Newsgroups, comparing Naive Bayes, logistic regression and a linear SVM, with a per-class report and the most confused pairs.Wed, 23 Jul 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/07/14/nlp-09-sentiment/https://blog.sahaj.si/2025/07/14/nlp-09-sentiment/Classifying movie-review sentiment on the Rotten Tomatoes dataset: a TF-IDF plus logistic regression baseline, its most telling words, and where it fails.Mon, 14 Jul 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/07/04/nlp-08-ner/https://blog.sahaj.si/2025/07/04/nlp-08-ner/NER as token-level sequence labelling with BIO tags, running spaCy's pretrained pipeline, and why entity spans make evaluation tricky.Fri, 04 Jul 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/06/25/nlp-07-attention/https://blog.sahaj.si/2025/06/25/nlp-07-attention/Additive attention lets the decoder look at every encoder state instead of one summary vector; the same reversal task with attention, and its alignment weights.Wed, 25 Jun 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/06/17/nlp-06-seq2seq/https://blog.sahaj.si/2025/06/17/nlp-06-seq2seq/The encoder–decoder architecture for sequence-to-sequence tasks, teacher forcing, and a small GRU model learning to reverse digit strings.Tue, 17 Jun 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/06/06/nlp-05-glove/https://blog.sahaj.si/2025/06/06/nlp-05-glove/GloVe vectors from global co-occurrence statistics, loading pretrained 50-dimensional vectors, analogies, and the bias these vectors inherit.Fri, 06 Jun 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/05/28/nlp-04-word2vec/https://blog.sahaj.si/2025/05/28/nlp-04-word2vec/Dense word vectors learned by predicting context (skip-gram), trained with gensim on the 20 Newsgroups corpus, with nearest neighbours.Wed, 28 May 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/05/20/nlp-03-bow-tfidf/https://blog.sahaj.si/2025/05/20/nlp-03-bow-tfidf/Turning documents into vectors with CountVectorizer and TfidfVectorizer, n-grams, and finding similar documents with cosine similarity.Tue, 20 May 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/05/09/nlp-02-tokenisation/https://blog.sahaj.si/2025/05/09/nlp-02-tokenisation/Word, character and subword tokenisation compared, out-of-vocabulary words, and what a real BERT WordPiece tokenizer does to unusual words.Fri, 09 May 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/05/01/nlp-01-preprocessing/https://blog.sahaj.si/2025/05/01/nlp-01-preprocessing/Cleaning raw text with regex and spaCy: normalisation, lowercasing, stop words, lemmatisation, and why modern models need much less of this.Thu, 01 May 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/04/22/dl-13-lstm-gru/https://blog.sahaj.si/2025/04/22/dl-13-lstm-gru/Gated recurrent units: the LSTM's cell state and three gates, the simpler GRU, and an experiment measuring how far gradients reach back through 100 time steps.Tue, 22 Apr 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/04/11/dl-12-rnn/https://blog.sahaj.si/2025/04/11/dl-12-rnn/Recurrent networks carry a hidden state across time steps; the unrolled view, nn.RNN shapes, and why gradients vanish over long sequences.Fri, 11 Apr 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/04/03/dl-11-cnn-intro/https://blog.sahaj.si/2025/04/03/dl-11-cnn-intro/Convolutions as sliding filters with shared weights, padding, stride, pooling, output-size arithmetic, and a small CNN versus an MLP on digits.Thu, 03 Apr 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/03/25/dl-10-training-loop/https://blog.sahaj.si/2025/03/25/dl-10-training-loop/The standard PyTorch loop with Dataset, DataLoader, nn.Module, a loss, an optimiser and an evaluation pass, applied to scikit-learn's handwritten digits.Tue, 25 Mar 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/03/14/dl-09-pytorch-autograd/https://blog.sahaj.si/2025/03/14/dl-09-pytorch-autograd/PyTorch tensors (NumPy-like, with devices and dtypes), the dynamic computation graph, requires_grad, backward(), and no_grad.Fri, 14 Mar 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/03/06/dl-08-dropout/https://blog.sahaj.si/2025/03/06/dl-08-dropout/Dropout randomly zeroes activations during training as a regulariser; inverted scaling, train vs eval behaviour, and its effect on an overfitting MLP.Thu, 06 Mar 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/02/25/dl-07-batchnorm/https://blog.sahaj.si/2025/02/25/dl-07-batchnorm/Normalising each feature over the mini-batch, the learnable scale and shift, train vs eval mode, and its effect on a deep MLP.Tue, 25 Feb 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/02/17/dl-06-initialisation/https://blog.sahaj.si/2025/02/17/dl-06-initialisation/Why initial weight scale matters, Xavier/Glorot for tanh and He/Kaiming for ReLU, and watching activations explode or collapse through 50 layers.Mon, 17 Feb 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/02/06/dl-05-optimisers/https://blog.sahaj.si/2025/02/06/dl-05-optimisers/Mini-batch SGD, momentum, RMSProp-style scaling and Adam compared on an ill-conditioned valley, using torch.optim.Thu, 06 Feb 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/01/28/dl-04-activations/https://blog.sahaj.si/2025/01/28/dl-04-activations/Sigmoid, tanh, ReLU, Leaky ReLU and GELU compared by shape and gradient, and why saturating activations make deep networks hard to train.Tue, 28 Jan 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/01/20/dl-03-backprop/https://blog.sahaj.si/2025/01/20/dl-03-backprop/Backpropagation as the chain rule applied layer by layer; a two-layer network trained on XOR in pure NumPy, plus a numerical gradient check.Mon, 20 Jan 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/01/09/dl-02-mlp/https://blog.sahaj.si/2025/01/09/dl-02-mlp/Adding a hidden layer with a non-linear activation turns the perceptron into an MLP that can solve XOR; a forward pass written in NumPy.Thu, 09 Jan 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2025/01/01/dl-01-perceptron/https://blog.sahaj.si/2025/01/01/dl-01-perceptron/Rosenblatt's perceptron from scratch: a weighted sum, a step function, and the update rule; it learns AND but can never learn XOR.Wed, 01 Jan 2025 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/12/23/ml-18-hyperparameter-search/https://blog.sahaj.si/2024/12/23/ml-18-hyperparameter-search/Tuning with GridSearchCV and RandomizedSearchCV, nested evaluation on a held-out test set, and why random search often wins.Mon, 23 Dec 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/12/13/ml-17-feature-engineering/https://blog.sahaj.si/2024/12/13/ml-17-feature-engineering/Creating better inputs: ratios, date parts, binning, interactions, log transforms for skewed values, and target-free encodings.Fri, 13 Dec 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/12/05/ml-16-pipelines/https://blog.sahaj.si/2024/12/05/ml-16-pipelines/Combining preprocessing and model with Pipeline and ColumnTransformer for mixed numeric and categorical data, so one object is fitted, tuned and saved.Thu, 05 Dec 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/11/27/ml-15-cross-validation/https://blog.sahaj.si/2024/11/27/ml-15-cross-validation/k-fold, stratified and grouped cross-validation, why one train/test split is noisy, and avoiding leakage by preprocessing inside the folds.Wed, 27 Nov 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/11/19/ml-14-metrics/https://blog.sahaj.si/2024/11/19/ml-14-metrics/Confusion matrix, accuracy and why it fails on imbalanced data, precision, recall, F1, ROC-AUC, and regression metrics.Tue, 19 Nov 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/11/11/ml-13-pca/https://blog.sahaj.si/2024/11/11/ml-13-pca/Principal component analysis: rotating data onto directions of maximum variance, explained variance ratio, and compressing the digits dataset.Mon, 11 Nov 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/11/01/ml-12-hierarchical-dbscan/https://blog.sahaj.si/2024/11/01/ml-12-hierarchical-dbscan/Two alternatives to k-means: agglomerative clustering, which builds a tree of merges, and DBSCAN, which finds dense regions of any shape and labels noise.Fri, 01 Nov 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/10/24/ml-11-kmeans/https://blog.sahaj.si/2024/10/24/ml-11-kmeans/Unsupervised clustering with k-means: the assign/update loop, inertia, choosing k with the elbow and silhouette score.Thu, 24 Oct 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/10/16/ml-10-svm/https://blog.sahaj.si/2024/10/16/ml-10-svm/SVMs find the widest-margin boundary; support vectors, the C parameter, and the kernel trick for non-linear data.Wed, 16 Oct 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/10/08/ml-09-boosting-xgboost/https://blog.sahaj.si/2024/10/08/ml-09-boosting-xgboost/Boosting builds shallow trees one after another, each fixing the previous errors; gradient boosting, learning rate, early stopping, and XGBoost.Tue, 08 Oct 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/09/30/ml-08-random-forest/https://blog.sahaj.si/2024/09/30/ml-08-random-forest/Bagging many decorrelated trees: bootstrap samples, random feature subsets, out-of-bag score, and feature importances.Mon, 30 Sep 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/09/20/ml-07-decision-trees/https://blog.sahaj.si/2024/09/20/ml-07-decision-trees/How a decision tree chooses splits using Gini impurity, why deep trees overfit, and reading the learned rules.Fri, 20 Sep 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/09/12/ml-06-knn/https://blog.sahaj.si/2024/09/12/ml-06-knn/k-NN classifies by majority vote of the closest training points; why scaling matters and how k controls the bias–variance trade-off.Thu, 12 Sep 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/09/04/ml-05-regularisation/https://blog.sahaj.si/2024/09/04/ml-05-regularisation/Overfitting with too many features, and how L2 (ridge) shrinks weights while L1 (lasso) drives some exactly to zero.Wed, 04 Sep 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/08/27/ml-04-logistic-regression/https://blog.sahaj.si/2024/08/27/ml-04-logistic-regression/Logistic regression for binary classification: the sigmoid, log-loss, decision thresholds, and predicted probabilities.Tue, 27 Aug 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/08/19/ml-03-gradient-descent/https://blog.sahaj.si/2024/08/19/ml-03-gradient-descent/Fitting linear regression by gradient descent from scratch: the gradient of MSE, the learning rate, and what happens when it is too small or too big.Mon, 19 Aug 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/08/09/ml-02-linear-regression/https://blog.sahaj.si/2024/08/09/ml-02-linear-regression/Linear regression with many features: the model, the squared-error cost, the normal equation, and scikit-learn's LinearRegression with RMSE and R².Fri, 09 Aug 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/08/01/ml-01-workflow/https://blog.sahaj.si/2024/08/01/ml-01-workflow/The end-to-end supervised learning loop: define the task, split data, pick a baseline, fit, evaluate on held-out data, iterate.Thu, 01 Aug 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/07/25/stats-14-regression/https://blog.sahaj.si/2024/07/25/stats-14-regression/Fitting y = a + bx by least squares, reading slope and intercept, R², residuals, and the bridge to machine learning.Thu, 25 Jul 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/07/18/stats-13-correlation/https://blog.sahaj.si/2024/07/18/stats-13-correlation/Covariance, Pearson vs Spearman correlation, what r does and doesn't capture, and why correlation is not causation.Thu, 18 Jul 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/07/11/stats-12-t-chi-square/https://blog.sahaj.si/2024/07/11/stats-12-t-chi-square/One-sample, two-sample (Welch) and paired t-tests, and the chi-square test of independence for categorical data.Thu, 11 Jul 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/07/05/stats-11-hypothesis-tests/https://blog.sahaj.si/2024/07/05/stats-11-hypothesis-tests/The logic of a hypothesis test: null and alternative, test statistic, p-value, significance level, and the two types of error.Fri, 05 Jul 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/06/28/stats-10-estimation/https://blog.sahaj.si/2024/06/28/stats-10-estimation/Estimators and their bias, standard error, and building and interpreting 95% confidence intervals with the t distribution.Fri, 28 Jun 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/06/21/stats-09-sampling/https://blog.sahaj.si/2024/06/21/stats-09-sampling/Population vs sample, simple random vs stratified sampling, sampling bias, and the bootstrap for estimating uncertainty.Fri, 21 Jun 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/06/17/stats-08-clt/https://blog.sahaj.si/2024/06/17/stats-08-clt/Why sample means look normal even when the data doesn't, the σ/√n standard error, and a simulation from a skewed distribution.Mon, 17 Jun 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/06/10/stats-07-continuous-distributions/https://blog.sahaj.si/2024/06/10/stats-07-continuous-distributions/Uniform, normal and exponential distributions, z-scores, the 68–95–99.7 rule, and computing areas with cdf and ppf.Mon, 10 Jun 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/06/03/stats-06-discrete-distributions/https://blog.sahaj.si/2024/06/03/stats-06-discrete-distributions/Bernoulli, binomial, geometric and Poisson distributions: what each one counts, their means, and computing them with scipy.stats.Mon, 03 Jun 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/05/27/stats-05-random-variables/https://blog.sahaj.si/2024/05/27/stats-05-random-variables/Random variables as numbers attached to outcomes, PMF vs PDF vs CDF, expectation and variance, and linearity of expectation.Mon, 27 May 2024 00:00:00 GMTstudyhttps://blog.sahaj.si/2024/05/21/stats-04-bayes/https://blog.sahaj.si/2024/05/21/stats-04-bayes/Bayes' theorem as updating a prior with evidence, the base-rate trap in medical tests, and sequential updating.Tue, 21 May 2024 00:00:00 GMTstudy