Wednesday, Jan. 22- DAILY AGENDA

TRANSFER LEARNING WORKSHOPS

Workshops will be on Colab using TensorFlow 2.0 - Keras

10:45 - 11:30 Morning Workshop: How to use existing models for transfer learning 

In the first Lab of day two we will introduce the concept of Transfer Learning through a simple image classification example. We will employ MobileNet architecture, which is a well known deep network used for classification tasks. This network performs well when classifying images; however, when MobileNet is used to classify images that have never been used in the training process, the performance dramatically decreases. One way to improve the performance is to train MobileNet in the new dataset of images, but this can be time consuming. With Transfer Learning, we will demonstrate how using the pre-trained MobileNet to replace the last layer, a fully connected classification layer, is a much faster way to improve the performance.

Workshop Link

KEYNOTE ADDRESS: "Jupyter meets the Earth: an open, collaborative approach for Earth data science"

10:00 - 10:30 Morning Lecture
Pavlos Protopapas, PhD
Scientific Program Director, IACS

Transfer learning is a machine learning method where a model developed for a task is reused as the starting point for a model on a second task. It is a popular approach in deep learning where pre-trained models are used as the starting point on computer vision and natural language processing tasks. This can be very important, given the vast computational and time resources required to develop neural network models on these problems and given the huge jumps in skill that these models can provide to related problems. In this part of the program we will examine various pre-existing models and techniques in transfer learning.

9:00 - 10:00

Professor Pérez will be speaking jointly to participants at Harvard IACS ComputeFest and the Harvard IQSS DataFest.

Fernando Pérez
Associate Professor of Statistics, UC Berkeley, Co-founder Project Jupyter
10:30 - 10:45 Coffee Break
8:30 - 9:00 Registration
Marios Mattheakis, PhD
Research Associate, IACS
11:30 - 12:45 Morning Workshop 2: Transfer Learning Across Tasks: Imagenet helps semantic segmentation
Robbert Struyven, MD
S.M. Data Science '20
12:45 - 1:45 Lunch Break
1:45 - 2:45 Afternoon Workshop 1: Network Distillation 
Robbert Struyven, MD
S.M. Data Science '20
Camilo Fosco
PhD Student, CSAIL, MIT
Vincent Casser
Research Scientist, Waymo
2:45 - 3:00 Coffee Break

3:00 - 4:30 Afternoon workshop 2 hosted by Weights and Biases (WandB)

 

This workshop will demonstrate the Weights & Biases experiment tracking platform.  You'll learn how to debug performance and dataset issues as well as problems with model convergence.  You will also perform hyperparameter sweeps and discuss current popular approaches for optimizing model performance.

Prerequisites for this workshop

  • Proficient in Python

  • Familiarity with Keras 

  • You will need a Weights & Biases account to access class material. We will allot time at the beginning of the workshop to register for an account. You're welcome to sign up for a free account before our workshop here

Useful Links

logo - Weights & Biases - black and gold
Chris Van Pelt
Co-Founder, Weights and Biases
Carey Phelps
Head of Product, Weights and Biases

PREREQUISITES:  Workshops assume fluency in Python and basic machine learning to the level of  Harvard's

CS 109a or a beginner data science/ML course. 

IACS.SEAS.HARVARD.EDU 

HARVARD INSTITUTE FOR APPLIED COMPUTATIONAL SCIENCE 

33 OXFORD ST. CAMBRIDGE, MA 02138

LRAY@SEAS.HARVARD.EDU 

IACSLogo_RGB_UseOnLightBackground (4).jp