On February 16, 17 & 20th, Goutham Gopalakrishna gave a mini lecture on Deep Learning and Macro-Finance Models. Gopalakrishna from École Polytechnique Fédérale de Lausanne and the Swiss Finance Institute and is currently a visiting research collaborator with the Bendheim Center for Finance.
PART 1: Introductions to Neural Networks
Download the presentation slides
Time Stamps
00:00:00 – Introduction to the mini-course
00:06:36 – Revival of neural networks
00:08:53 – Introduction to deep-learning
00:10:58 – Feed-forward neural networks
00:17:47 – A typical deep-learning problem
00:21:44 – Why deep learning works?
00:26:35 – Comparison to projection methods
00:31:30 – Learning by trial and error
00:36:44 – Neural networks as function approximators
00:43:17 – Sneak peek into Google Colaboratory
00:46:25 – Comparison to other methods
00:53:56 – Limitations
Part 2: Deep Learning Principles, High-Dimensional Optimization Techniques in Machine Learning
Download the presentation slides
Time Stamps
00:02:00 – Learning principles
00:03:50 – Introduction to gradient descent methods
00:13:40 – Momentum based gradient descent
00:15:12 – Implementation in Python Jupyter notebook
00:15:50 – Nesterov accelerated descent
00:26:15 – Variants of gradient descent method
00:28:27 – Computing gradients using backpropagation
00:32:06 – Basic activation functions
00:36:54 – Parameter initialization
00:41:55 – Object-oriented programming 101
00:47:25 – Decorators in Python
00:52:47 – Getting started with Tensorflow
Part 3: Application to Solve Macro-Finance Models with Aggregate Shocks
Download the presentation slides
Time Stamps
00:01:05 – Introduction to ALIENs
00:02:47 – General setup
00:05:15 – HJB equations in continuous-time
00:11:00 – Neural network solution method
00:16:14 – A mesh-free approach to solving the HJB equation
00:23:53 – Active deep-learning example
00:31:22 – Putting together the ALIEN architecture
00:35:02 – Comparison to simulation-based methods
00:37:27 – Automatic differentiation in practice
00:40:58 – Hands-on with the benchmark model
00:51:22 – Analyzing the computational loss
01:03:11 – Comparison with the finite difference solution
01:05:49 – A high dimension example
01:10:15 – Data parallelism
01:16:23 – Using high performance computing (HPC)