Md. Asif Uddin
    Figura IIOne forward pass, one backward pass — and it continuesSignal crosses; the units take their values. Gradient returns along the same wires, and the bright ones are where the loss is most sensitive — which is to say, where an update would move things. Then it happens again. The repetition is the training.

    A structured study of modern artificial intelligence

    Elementa

    Read. Learn. Understand. Research. A structured journey through the core mechanisms of modern AI. Read in order, like Euclid. Begin with the simple. Follow each idea to its consequences. Gradually build toward the difficult, from representation and computation to neural networks, vision, language, multimodal systems, causality, and biological intelligence. Each entry is a proposition, not a post. It makes a claim, builds upon what came before, and develops the idea through intuition, mathematics, examples, and figures. Every proposition should leave you with something you can explain, question, and build upon. Elementa is not meant to be read and forgotten. Read it. Learn it. Understand it. Research it. Question what you learn. Connect it to what came before. Test it. Extend it. Knowledge is not a collection of isolated facts. It is a structure of ideas, where each thing understood makes the next thing possible. Elementa is an attempt to build that structure carefully, one proposition at a time. If the figure cannot be drawn, the proposition is not yet understood.

    8 books · 74 chapters100 propositions written

    Md. Asif UddinDhaka · MMXXVI

    Contents

    1. Book INeural networks14 chapters · 51 written · 13 load-bearingthe perceptron, backpropagation, optimisation, convolution, recurrence, latent variables, graphsHow a neural network computes and how it learns. Fourteen chapters, the largest book here, and the one every other book borrows from.
    2. Book IIFoundations8 chapters · 22 written · 6 load-bearingtokenisation, attention, the transformer blockBuild the conceptual and mathematical footing the rest of the corpus stands on. By the end you should understand how a network represents information and how a transformer processes a sequence.
    3. Book IIIVision7 chapters · 26 written · 3 load-bearingViT, DINOv2, self-supervised visual representationExplain how a machine represents and interprets visual information, from pixels to self-supervised representations.
    4. Book IVLarge Language Model (LLM)9 chapters · 0 written · 5 load-bearingTransformers, attention mechanisms, prompt engineering, fine-tuning.Move from the transformer block to modern language models: how they are trained, scaled, aligned and made to fail.
    5. Book VVision-language (VLM)7 chapters · 0 written · 1 load-bearingcontrastive pretraining, fusion, VLM architecturesExplain how visual and linguistic representations are joined, and what becomes possible once they share a space.
    6. Book VICausal inference9 chapters · 0 written · 7 load-bearingDAGs, interventions, counterfactuals, perturbationMove from prediction to intervention: what a causal claim is, when data can support one, and when it cannot.
    7. Book VIIBioinformatics10 chapters · 0 written · 4 load-bearingsingle-cell data, gene regulatory networks, perturbation screensShow how machine learning meets biological data, and where a perturbation turns a correlation into a causal question.
    8. Book VIIIPractice10 chapters · 1 written · 2 load-bearinglosses, training dynamics, evaluationTurn knowledge into research: reading papers, forming a question, running an experiment that means something, and writing it down.
    Book 0 · ApparatusThe mathematics the other seven books assume47 entries across seven parts — linear algebra, matrix calculus, probability, information, optimisation, statistics, numerics. A reference volume, not a linear read, and never a gate: start Book I today and follow the links when a step stops making sense.

    The Work at a Glance

    Every chapter declares how much mathematical load it carries, and the declaration has consequences. A load-bearing chapter owes five worked problems across four distinct variants, ten solved exercises, eight numbered definitions, two named results each with the boundary of its guarantee, and two concrete failures — before it can be marked published. These minima are summed from what the chapters actually declare, so the table cannot drift from the corpus.

    M0NarrativeM1DefinitionalM2SubstantiveM3Load-bearing

    BookM0M1M2M3DefinitionsProblemsExercises
    I · Neural networks——113137 / 11085 / 68136 / 136
    II · Foundations——260 / 601 / 361 / 72
    III · Vision—1330 / 460 / 250 / 51
    IV · Large Language Model (LLM)—1350 / 620 / 350 / 71
    V · Vision-language (VLM)11410 / 390 / 180 / 37
    VI · Causal inference——270 / 680 / 410 / 82
    VII · Bioinformatics1—540 / 650 / 350 / 70
    VIII · Practice33220 / 490 / 190 / 41
    All eight562241137 / 49986 / 277137 / 560

    That total is not a burden to be reduced. It is the difference between a body of knowledge and a set of articles, and it is why the books are built one at a time rather than all at once.