← AI Terminology

TensorFlow

TensorFlow is Google's open-source machine learning framework — initially built around static computation graphs (define-then-run), now supporting eager execution (2.0+), with a mature production ecosystem including TensorFlow Serving, TFLite (mobile), and TensorFlow.js (browser).

It was the dominant deep learning framework 2015–2019 before PyTorch overtook it in research.
Why It Matters in AI
TensorFlow was the first framework to combine a complete ML production stack: training (tf.keras), serving (TensorFlow Serving), mobile deployment (TFLite), and browser execution (TF.js). Google's massive infrastructure bet on TF means it remains dominant in Google-internal systems, cloud ML platforms (Vertex AI), and edge deployment on Android/embedded. Keras 3 (which runs on TF, PyTorch, or JAX) re-positions TensorFlow as a backend rather than a standalone framework.
Key Points
Aspect Description
XLA TensorFlow uses XLA compiler for TPU — same backend as JAX
TF 2.x Eager execution default + tf.keras API — made TF competitive with PyTorch's usability
TFLite Optimised inference runtime for Android, iOS, and embedded devices — INT8 quantised models
Keras 3 os.environ["KERAS_BACKEND"] = "tensorflow" — TF is now just one Keras backend
TF Serving Production model server — REST/gRPC API, model versioning, zero-downtime updates
tf.function @tf.function — traces and compiles eager code to a static graph for production speed
Simple Analogy
A Swiss factory with a consumer product line (Keras), an industrial division (TF Serving), a miniature product for watch-sized devices (TFLite), and a browser app version (TF.js) — the same factory (computational graph) adapted to every deployment context, optimised for Google's global infrastructure.
Common Usage Examples
  • import tensorflow as tf; model = tf.keras.Sequential([Dense(128, activation='relu'), Dense(10)])
  • model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
  • TFLite: converter = tf.lite.TFLiteConverter.from_saved_model(path); tflite_model = converter.convert()
  • TF Serving: tensorflow_model_server --model_name=my_model --model_base_path=/models/
  • @tf.function def train_step(x, y): with tf.GradientTape() as tape: loss = model(x, training=True)
Summary
In short: TensorFlow is Google's production-grade ML framework — the backbone of Google's AI infrastructure and Android ML deployment (TFLite), with a mature serving ecosystem, though PyTorch has overtaken it in research and now dominates new model development.