@machinelearningtogo: ➡️ Part 8 of learning ML code from scratch: Every model you have ever heard of, up to ChatGPT, learned with this one idea: walk downhill. Two questions guide the video: what even is gradient descent, and why do we need it? Step 1, the error. Our line from part one guesses points from hours studied. Guess minus truth, squared, averaged: that one number is the loss. At the start it is 57.3. Step 2, the landscape. Every pair of weights gives one loss. Put the weights flat as the floor and the loss up as height, and the error becomes a landscape. One point on it is one complete setting of the model. Learning has one goal: the lowest point, where the error is smallest and the line fits best. Step 3, the direction. The model cannot see the valley, it only feels the ground under its feet. The derivative tells it which way is down: nudge one weight by a tiny bit, watch the loss, and step against the slope. For a whole network, backpropagation does this for every weight at once. Step 4, the walk. Measure, take the slopes, step downhill, again and again. In the video the loss goes 57.3, then 1.97, then 0.12, then settles at 0.022, where the line runs through the points. Real networks do the same in millions of directions. Last knob: the learning rate, the step size. 0.1 jumps across the valley and explodes to 4e19. 0.0002 crawls and is still at 3.4 after 300 steps. 0.05 settles at the bottom. Every number on screen comes from running the shown code. The landscape is the code's own loss function evaluated on a grid. #machinelearning #gradientdescent #neuralnetworks #python #coding

QuantMachineLearn

Open In TikTok:

Region: DE

Monday 21 September 2026 13:11:31 GMT