Deep Learning with CNTK and F#
summary
Shows how to train real models with CNTK .NET API and F#: a toy classifier on 8x8 binary digits (32 ones, 32 twos) learns weights after ~1,000 passes and yields interpretable positive/negative weight maps, and a small CNN on MNIST-style data (5,000 training, 500 validation) built with conv 3x3 + ReLU + pooling + dense achieves ~9 errors on 500 images (~98% acc) while running on CPU. Builds a tiny higher‑level sequential layer library (NuGet v0.1.3), saves interoperable CNTK model files, and notes the default API is low‑level but functional.
AI-written summary · about
chapters (24)
- 0:00Start
- 0:48Context
- 2:15Plan
- 4:45Activation Gate
- 5:04Displaying a T
- 6:45Weights to control signal strength
- 8:10Extend to 1, 2, 3...
- 10:35Learning by example
- 11:33How should the weights look like?
- 13:24Robustness
- 15:12Learning a Circuit / Board
- 15:49Composing more complex circuits
- 17:27Model: Composition of Blocks
- 17:53Model: Composition of Layers
- 19:19What is CNTK?
- 21:39What does CNTK do?
- 22:30Core components
- 23:47What the hell is a Tensor?
- 25:18Functions
- 40:10Sequential Models
- 41:44Demo: MNIST dataset
- 43:25MNIST input
- 52:14Conclusion
- 53:55Thank you :
related talks
55mDeep Learning with CNTK and F#
57mDeep Learning with CNTK and F#
01h 05mDeep Learning with PyTorch
55mA Tour of Deep Learning With C++
54mNeural networks by example
26mIntroduction to the Microsoft Cognitive Toolkit (Deep Learning Framework)
01h 57mUsing CNTKs Python Interface for Deep Learning
01h 06mDeep Learning with TensorFlow