Coffee Roasting Lab

• Optional Lab: Build a small neural network using TensorFlow

• Coffee Roasting Dataset

• Features: Temperature (Celsius) and Duration (minutes)

• “Duration is best kept between 12 and 15 minutes while the temp should be between 175 and 260 degrees Celsius”

• As temperature rises, duration should shrink

• Data Normalization

• Features normalized using Keras normalization layer:

  • Create normalization layer (not part of model)
  • Adapt the data (learns mean and variance)
  • Normalize the data

• Results:

  • Temperature: Max/Min pre-norm: ~200-300 → post-norm: ~1.5-2.0
  • Duration: Max/Min pre-norm: ~10-20 → post-norm: ~-1.5-1.5

• Model Architecture: “Coffee Roasting Network”

• 2 input features → 3 hidden units → 1 output unit

• All activations use sigmoid function

• Input shape = (2,)

• Model definition:

  model = Sequential([
      tf.keras.Input(shape=(2,)),
      Dense(3, activation='sigmoid', name='layer1'),
      Dense(1, activation='sigmoid', name='layer2')
  ])

• Model Parameters

• Layer 1: 9 parameters

  • W1: shape(2,3) = 6 parameters
  • b1: shape(3,) = 3 parameters

• Layer 2: 4 parameters

  • W2: shape(3,1) = 3 parameters
  • b2: shape(1,) = 1 parameter

• Training Process

• Compile with:

  • Loss: BinaryCrossentropy
  • Optimizer: Adam(learning_rate=0.01)

• Fit for 10 epochs

• Output shows progress: “Epoch 1/10, 6250/6250 […] - loss: 0.1782”

• Model weights updated during training

• Making Predictions

• Input data must be normalized same as training data

• Output is probability (of good roast)

• Apply threshold (≥ 0.5) to make decision:

  yhat = (predictions >= 0.5).astype(int)

• Layer Functions Visualization

• First layer units specialize in different regions:

  • Unit 0: Detects when temperature is too low
  • Unit 1: Detects when duration is too short
  • Unit 2: Detects bad combinations of time/temperature

• “The network learned these functions on its own through gradient descent”

• “They are very much the same sort of functions a person might choose”

• Network Output Visualization

• Left graph: Raw output (probability) with training data overlay

• Right graph: Decisions after threshold with results shown

Note: The network demonstrated ability to handle complex decisions by dividing the decision task between multiple units.

This lab shows how to build, train and analyze a neural network in TensorFlow for coffee roasting classification. The network learns interpretable features in the hidden layer that correspond to meaningful aspects of the roasting process.