The document surveys 42 neural network architectures and training schemes from perceptron to diffusion LLMs, using diagrams that map networks from input to output. It explains how to read the diagrams, what symbols mean, and how repeating components reuse weights, with each panel noting the year and learning objective.
The course explains artificial neural networks through hands-on experimentation, focusing on internal machinery and intuition about suitable problems; students will be prepared to study algorithms or build a network themselves.
A learning path that guides building neural networks from a single neuron to forward propagation, loss, and backpropagation with step-by-step explainers.
BREM-NET is a nonlinear multimodal model that fuses behavior with multiple neural modalities to separate behaviorally relevant neural dynamics from residual activity, improving decoding and prediction, even with asynchronous data or dropouts in a primate 2D reaching task.
The article argues that neural networks may be easier to understand than commonly thought, offering explanations to demystify AI decision-making and improve transparency.
Researchers show that a self-organizing artificial neural network, guided by simple local rules, can learn and reorganize itself to perform information processing, offering insights into brain-like learning and potential neuroscience applications. The work compares self-organizing networks to backpropagation-based models and discusses implications for brain-inspired AI and neurotechnology.
Describes a compact convolutional neural network for sparse chess control, detailing input channels, architecture, loss, optimization, and replication steps using Lc0 data and MLflow for tracking.
A tutorial-style repository demonstrates neural networks built from scratch in NumPy, from a single perceptron to a multi-layer network, covering activation functions, optimizers, regularization, initialization, batch normalization, and hyperparameter tuning, with consistent synthetic datasets and a common baseline for comparing techniques.
The paper argues that deep nets’ expressiveness creates interpretable challenges, showing (i) semantic information lies in the space of high-level representations rather than individual units, and (ii) small, input perturbations can reliably cause misclassification across networks trained on different data subsets.
The SaluNet repository provides code and experiment results for CNNs and Vision Transformers that replace feature normalization with SALU and test gated activation variants. It includes CIFAR implementations and ImageNet-1K reference scripts, with reported CIFAR-100 ResNet-18 results reaching 83.42% EMA accuracy in a separate deterministic run.