TensorFlow Implementation of Neural Networks

• TensorFlow is a leading framework for deep learning

• “When I’m building projects, TensorFlow is actually a tool that I use the most often”

• Alternative: PyTorch (not covered in this specialization)

• Neural networks can be applied to many different applications

• Example: Coffee bean roasting optimization

  • Parameters: temperature and duration
  • Dataset: different temperatures/durations labeled as good/bad coffee
  • Good coffee (y=1) only in a specific region
    • Too low temperature → undercooked
    • Too short duration → not nicely roasted
    • Too high temperature/duration → burnt beans

Implementing Inference with TensorFlow

• Coffee roasting example implementation:

• Input: x = [200, 17] (temperature in Celsius, duration in minutes)

• Layer 1 (first hidden layer):

  • Layer 1 = dense(units=3, activation="sigmoid")
  • Computes: a1 = Layer1(x) → produces 3 values (e.g., [0.2, 0.7, 0.3])

• Layer 2:

  • Layer 2 = dense(units=1, activation="sigmoid")
  • Computes: a2 = Layer2(a1) → produces 1 value (e.g., 0.8)

• Final prediction:

  • Optional thresholding: y_hat = 1 if a2 >= 0.5 else 0

• Handwritten digit classification example:

• Input: x = numpy array of pixel intensity values

• Layer 1:

  • Layer 1 = dense(units=25, activation="sigmoid")
  • Computes: a1 = Layer1(x)

• Layer 2:

  • Layer 2 = dense(units=15, activation="sigmoid")
  • Computes: a2 = Layer2(a1)

• Layer 3:

  • Layer 3 = dense(units=1, activation="sigmoid")
  • Computes: a3 = Layer3(a2)

• Final prediction:

  • Optional thresholding for binary prediction

Note: “Dense” is another name for the standard neural network layers covered previously.

• Implementation details:
• Additional details in lab (loading TensorFlow library, loading parameters)
• TensorFlow treats data in specific ways (covered in next video)

TensorFlow provides a clean syntax for implementing neural network inference, allowing the application of the same fundamental algorithms to vastly different problems. The framework handles the complex computations while maintaining a relatively simple interface for model creation and prediction.