This illustrated guide walks from raw tabular data to production-ready neural-network insight, using the classic Iris flowers as a didactic sandbox.  We begin with data-quality triage—missing-value scans, outlier flags, and correlation checks—then motivate feature-wise normalization and stratified splitting.  Architecture design is treated as an intuition exercise: depth as sequential transformations, width as parallel pattern capacity, and activation choice as the valve that keeps gradients healthy.  A compact 4 → 50 → 3 ReLU network is trained and its parameters regularised with weight-decay and early stopping.

Interpretability is attacked on three fronts: permutation-feature drops quantify global importance; SHAP values assign per-sample attributions; t-SNE of hidden activations reveals class clusters that echo the confusion matrix.  Robustness drills inject Gaussian noise, occlude key features, and apply distribution shifts to expose accuracy and calibration drift.  Throughout, key metrics—accuracy, class-wise precision/recall, F1—are logged in the same loop you would use in production.

The result is a self-contained blueprint for why each layer, activation, chart, and metric exists, and how to reason about a neural network’s behavior long before it meets real-world data.

Full Post: Link 

Humans use natural language for communication. It plays a vital role in conveying thoughts, ideas, and emotions. Humans have an innate capacity to learn and use language, acquiring it effortlessly during childhood through exposure and interaction with their environment. Language acquisition involves grasping grammar, vocabulary, and the rules that govern language structure.

Robots operate using a structured set of computer code, which is different from the natural language humans use. A robotics expert compiles specific instructions to enable robots to perceive and interact with their environment. This code explicitly encodes information about the robot's self, the surroundings, and objects. Higher-intelligence robots need to be programmed with abilities like logical reasoning, mathematical understanding, and spatial awareness. However, the need of expertise in robotics leaves a language gap between a layman and a robot. Bridging the gap becomes necessary since robots are becoming a part of our world to assist us with everyday tasks.

Full Post: Link 

Selection of a motion planner for an application can be a daunting task. This article is the first of a series of articles that will clarify the first principles of path planning, types of planners and their respective strengths, and criteria to select the right motion planner.

Full Post: Link

Large Language Models have gained wide adaptability and promise an AI powered future. LLMs transcend the conventional AI that is trained on a finite amount of data with an objective of performing analytics or making predictions. The underlying architecture of LLMs develops a so-called “world sense”. A conventional AI agent in the simplest way will learn how to interpolate or recognize patterns based on information it has been trained on. LLMs on the contrary develop cognitive abilities in addition to inferring from trained data. Their architecture captures human behavior, common reasoning, and creative abilities as part of their weights and architecture. LLMs have displayed behaviors that tend to be more human-like and not always logical or rational. Consider the worldly color link: warm is yellow and cold is green. LLMs possess the capability to understand this link and direct a conventional AI agent or robot towards either of the colors inferring from dialogue. The world sense that the models possess can be leveraged to accelerate the introduction of collaborative machines in today’s workforce.

Full Post: Link

Embark on an intellectual journey, exploring two fascinating realms - the Free Energy Principle (FEP) and Large Language Models (LLMs) like GPT-4. While seemingly disparate, understanding their differences enriches our perspective on artificial intelligence and cognitive science.

Full Post: Link