Training on GPU with Colab — Deep Learning & Neural Networks

Training deep learning models on CPU is feasible only for toy experiments — a ResNet-50 that trains in 2 hours on a GPU would take 4 to 6 days on a modern CPU. GPUs accelerate neural network training because their thousands of parallel processing cores handle the matrix multiplications that dominate both forward and backward passes simultaneously. Google Colab provides free access to NVIDIA GPUs (typically T4, A100, or L4 depending on tier) through a browser-based Jupyter notebook interface, making GPU training accessible without any hardware investment. Understanding how to use Colab effectively — enabling the GPU runtime, managing memory, loading data efficiently, mounting Google Drive for persistence, and profiling GPU utilisation — is a practical prerequisite for training real models. Without GPU access, the lessons in this course that train CNNs, LSTMs, and transformers would take days rather than minutes.

Analogy🏏Cricket

🏏 Think of it like cricket: The Duckworth-Lewis-Stern (DLS) method, the ICC bowling economy metric, and the net run rate each measure team performance but optimise for completely different things — DLS cares about wickets-in-hand, economy cares about runs-per-over, NRR cares about tournament-wide margins. Using the wrong metric to evaluate a bowler is like using MSE for classification: technically computable but optimising the wrong thing entirely. Just as you would never rank a spinner by batting average, you should never use MSE when your model predicts probabilities. Just as focal DLS adjustments down-weight easy chases and amplify close finishes, focal loss down-weights easy examples and amplifies hard minority-class examples. The choice of scoring metric defines what excellence means — and so does the choice of loss function.