Theory & Code-along: Triplet loss with cosine distance for Siamese Networks on CIFAR-10 dataset — What if we could encode every object image (human faces, etc) into a template — a vector of numbers? Thereafter, we could objectively determine the similarity between objects by numerically comparing — finding the distances — between their templates. In deep learning, this is exactly what the Siamese Neural Networks…

Part 1 — Cognitive Revolution — of the complete summary of Sapiens: A Brief History of Humankind — Once in a few years, a revolutionary and eye-opening book is published that questions our assumptions and reshapes our understanding of the world around us. Sapiens: A Brief History of Humankind, written by Yuval Noah Harari in 2014, is an exemplar of such literature. …

A comprehensive guide to data generation strategy for data-centric Machine Learning projects — In the world of deep learning, complex models are data-hungry — in both quality and quantity — to perform well on inference at test time. However, in real-world Artificial Intelligence projects by commercial companies, Machine Learning Engineers seldom have the luxury of readily available training data. …

A complete discussion on the nature of happiness, and the strategies to achieve well-being — Hundreds of thousands of humans are born each day into a world already occupied by billions. As with everyone who came before them, they will go through the motions of life — growing up, forging their future, and retiring in old age. Yet, despite the multitude of beings, we can…

Achieving optimal throughput and latency with model inference on high client-server traffic — In real-time AI model deployment en masse, efficiency of model inference and hardware/GPU usage is paramount. The speed of a single client-server inference request depends on the Latency and Throughput of the server. This is because deep learning models are typically mounted on a server or a cluster of servers…

All about setting up, quantization and running inference with TensorRT in Linux environment — During real-time AI model deployment, speed of model prediction (or inference) is paramount. To ensure seamless user experience, the speech engine by Apple’s Siri cannot take forever to decipher a speaker’s intent and produce an appropriate response. Neither can its FaceID lag when unlocking the iPhone. Fortunately, it is possible…

A complete guide from CNN to Transfer Learning for Kaggle’s Cat versus Dog dataset — In the modern era, Neural Networks triumph over traditional Machine Learning algorithms when modelling complex data sets, especially unstructured data — spatial data (Computer Vision) or sequential data (NLP and Time Series). While leading Neural Network architectures for NLP are helmed by Transformers (since 2017, with the paper “Attention Is…

Developing deep neural networks from scratch with Mathematics and Python — Unlike traditional machine learning frameworks, deep neural networks are extremely powerful because it can capture highly non-linear relations between the features and target. And the degree of non-linearity is usually controlled by adjusting the depth of the neural network, or the number of Hidden Layers it contains. But how should…

Creating shallow neural networks from scratch with Mathematics and Python — In the previous article, we have described that the vanilla neural networks are basically Multi-Layer Perceptron (MLP), or the regular feedforward networks. Other neural network architectures, such as CNN and RNN, are more complex customization of the structure behind Perceptron layers, but forward propagation and backpropagation process similar to MLP…