Advanced Deep Learning

 EECS E6691 - Topics and Data Driven Analysis and Computation 

Advanced Deep Learning

Columbia University Course

Zoran Kostic, Ph.D., Dipl. Ing.,  Professor of Professional Practice, zk2172(at)columbia.edu

Electrical Engineering Department, Data Sciences Institute, Columbia University in the City of New York

Course in a nutshell: 

 Advanced theory and practice of Deep Learning.  Applications and projects.

Description: Advanced (Second) Course on Deep Learning

Bulletin Description: Regularized autoencoders, sparse coding and predictive sparse decomposition, denoising autoencoders, representation learning, manifold perspective on representation learning, structured probabilistic models for deep learning, Monte Carlo methods, training and evaluating models with intractable partition functions, restricted Boltzmann machines, approximate inference, deep belief networks, deep learning in speech and object recognition.

Detailed Description for Spring 2024

EECS E6691 Advanced Deep Learning (TOPICS DATA-DRIVEN ANAL & COMP)

Spring 2024, 3 credits

Professor Zoran Kostic              zk2172 (at) columbia.edu     

A second-level seminar-style course in which the students study advanced topics in deep learning. Prior to this course, students must have previously taken a first course in deep learning. The course consists of: (i) studying state-of-the art architectural and modeling concepts, (ii) systematic review of recent literature and reproduction of the results, (iii) pursuing novel research ideas, (iv) participating in local and potentially in public contests on Kaggle or elsewhere, (v) class presentation(s) of paper studies during the semester, (vi) final project, (vii) quizzes during the lecture time. The course will address topics beyond the material covered in the first Deep Learning course (such as the Columbia course ECBM E4040), with applications of interest to students. Example topics are object detection and tracking, smart city and medical applications, use of spectral-domain processing, applications of transformers, and capsule networks.

Students entering the course must have prior experience with deep learning and neural network architectures, including Convolutional Neural Nets (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memories (LSTMs), and autoencoders. They need to have a working knowledge of coding in Python, Python libraries, Jupyter notebook, Tensorflow, both on local machines and on a cloud platform (such as Google Cloud Platform, GCP), and of GitHub or similar. The framework and associated tools, which are the focus of this course, are PyTorch and Google Cloud. The course will leverage the infrastructure and Python coding templates from the ECBM E4040 assignments (the first deep learning course by Prof. Kostic). Students must be self-sufficient learners and take an active role during the classroom activities.

Semester class assignments (paper reviews) will consist of reading, coding, and presentations. Every week, several student groups will give presentations reviewing selected papers from recent conferences such as NIPS and ICLR, including students’ results in reproducing the papers, followed by open discussion. Quizzes are a part of the class time. 

The final project is a group project (up to 3 students). The topic will be selected by students or by the instructor. It needs to be documented in a conference-style report and the code deposited in a GitHub repository. The code needs to be documented and instrumented so the instructor can run it after downloading from the repository. A Google Slides presentation of the project, suitable for a poster version, is required. Students will present the project at the end of the semester using the slides. 

Prerequisites

(i) Machine Learning (taken previously, or in parallel with this course). 

(ii) ECBM E4040 Neural Networks and Deep Learning, or an equivalent neural network/DL university course taken for academic credit. Although the quality of online ML and DL courses (Coursera, Udacity, edX) is outstanding, many takers of these courses complete the hands-on coding assignments superficially and therefore do not gain practical coding skills, which are essential to participate in this advanced course. Therefore, online courses are not accepted as prerequisites.

(iii) The course requires an excellent theoretical background in probability and statistics and linear algebra. 

Students are strongly advised to drop the class if they do not have an adequate theoretical background and/or previous experience with programming deep learning models. It is strongly advised (the instructor’s requirement) that students take no more than 12 credits of coursework (including this course) during the semester in which this course is taken.

Registration

The enrollment is limited to several dozen students. Instructor’s permission is required to register.  Students interested in the course need to add their name to the SSOL waitlist and may be asked to complete a questionnaire. The instructor will move the students off the SSOL waitlist.

Grading

• Assignments: (2-4 assignments per semester)

• Paper review presentations: 2-3 student presentations + code + discussions

• Project (proposal slide presentation + final slide presentation + final report + code repository)

• Exam and Quizzes

Assignment submission policy:

• The total number of days late across all assignments is 4.

• Late days do not apply to the final project (report, slides, presentations, or code)

2025 Spring Projects

• Automated Brain Aneurysm Detection in Time-of-Flight Magnetic Resonance Angiography (TOF-MRA)

• YOLO-based model for baggage security screening

• Self-Supervised Contrastive Pre-Training for Time Series via Time-Frequency Consistency

• Ordering Message Passing to Deal with Heterophily and Over-Smoothing

• SFT dataset for self debugging in code gen LLMs

• Optimizing MoE Routers: Design, Implementation, and Evaluation in Transformer Models

• VeriClaim: Interpretable Fact Verification for Claims

• Enhancing Generalizable Reasoning in LLMs via GRPO on Diverse Reasoning Tasks

• Predicting Human Actions from Pose Estimation Using Deep Learning

• LLMs with an Opinion

• Autoencoder prediction of NYC real estate value

• Reduced-order Neural Surrogate Modeling of Fluid Physics.

• Comparing Audio Tagging Models with fMRI Data

• Automated Quantum Circuit Builder

• Multimodal Skin Diseases Analysis

• Visual Reasoning Model with Chain of Thought for Robust Vision-Language Understanding

2024 Spring Projects

• AudioMamba: Mamba Architecture for Audio Classification

• Music Generation with Music transformer and LSTM 

• COVID-19 Forecasting using Spatio-Temporal Graph Attention Networks

• ConMamba: Convolution-augmented Mamba for Speech Recognition 

• Diminished Reality for Emerging Applications in Medicine through Inpainting

• Efficent Deep Learning Investigation

• Real-time Automatic Face Anonymization in Video Streams

• Development of Discrete Prompt Learning via Evolutionary Search 

• Heart Disease Detection Using Transformer Models

• 3D Tumor Segmentation with U-Net: Analyzing MRI Scans for Medical Insights 

• Low-Light Raw Image Enhancement 

• Direct Preference Optimization and Proximal Policy Optimization on Small Language Models 

• 3D Scene Reconstruction using Neural Radiance Fields and Structure from Motion 

• Dynamic Video Generation from Static Comic Panels 

• Small Object Detection using improved YOLO 

• Predictive Modeling of Tennis Player Poses and Ball Trajectory 

• VMamba: Visual State Space Model 

2023 Spring Projects

• Object Recognition and Seq2Seq Models for Handwritten Equation Recognition and LaTeX Translation

• Deep Learning for Soccer Pass Receiver Prediction in Broadcast Images

• Automatic Person Removal Pipeline

• AI Photographic Assistant: Implementation of Deep Learning Photographic Tools

• The Development of Athena: Leveraging GPT-3.5 for Adaptive and Interactive Intelligent Tutoring for Personalized Learning

• What’s Happened, Happened: Leveraging Past Decisions for Improved Interactive Image Segmentation

• ASL Detection Correction and Completion

• A hierarchical attention based model for biopsy classification

• Enhancing Indoor Bouldering Experience for Color-Blind Climbers: A Deep Learning Approach for Route Identification on Climbing Walls

• SwagGAN (StyleGAN for Fashion)

• Learning Multi-scale Visual Representation via Language Description for Segmentation

• Deep Learning For Financial Time Series

• Classifying neuron cell types within Drosophila melanogaster using graph convolutional networks

• SAM Based Cell Blood Classification Model

• Instance-Level Image Retrieval While Navigating Interactive 3D Environments

• Gaze and head redirection model based on StyleGAN3

2022 Spring Projects

• Pix2Pix Image-to-Image Translation with Conditional Adversarial Networks

• Jump rope counter

• Black Box Adversarial Attack with Style Information

• Learning Signed Distance Function for 3D Shape Representation (DeepSDF)

• Representation learning without any labeled data

• Comparison of Self-Supervised Models for Music Classification

• Subcellular localization of proteins using deep learning

• Image Descriptions Generator

• Predicting remaining surgery duration

• Vision Transformer

• Adversarial Audio Synthesis

• PlaNet - Latent Dynamics from Pixels

• RecoNET: Understanding What happens in Videos

2021 Spring Projects

• Multi-Graph Graph Attention Network for Music Recommendation

• 3D Facial Reenactment from 2D Video

• Temporal Fusion Transformers for Time Series Forecasting

• Deep Reinforcement Learning for Environmental Policy

• Stochastic & Split Convolutions for High Resolution Image Classification (Integrated with EfficientNet) 

• Temporal Fusion Transformers for Time Series Forecasting

• Forecasting Corn Yields with Semiparametric CNNs

• Pose Estimation + Instance Segmentation

• Speaker Independent Speech Separation

• Generalized Autoencoder-based Transfer Learning for Structural Damage Assessment

• Deep Reinforcement Learning for Environmental Policy

• End-to-end object detection with Transformers

• Exploring latent space of InfoGAN

• Multi-Graph Graph Attention Network for Music Recommendation

• 3D Facial Reenactment from 2D Video

• Pose Estimation + Instance Segmentation

2018-2020 Projects

• See list of  projects under E6040 link 

Course sponsored by equipment and financial contributions of:

• NVIDIA GPU Education Center, Google Cloud, IBM Bluemix, AWS Educate, Atmel, Broadcom (Wiced platform); Intel (Edison IoT platform), Silicon Labs.

PREVIOUS SEMESTERS

Detailed Description for Spring 2023

Instructor: Dr. Mehmet Kerem Turkcan              mkt2126 (at) columbia.edu     

This is an advanced-level course in which the students study topics in deep learning. It is required that students had previously taken a first-course in deep learning. The course consists of: (i) lectures on state-of-the-art architectural and modeling concepts, (ii) assignments, (iii) exam, and a (iv) final project. The course will address topics beyond material covered in the first course on Deep Learning (such as ECBM E4040), with applications of interest to students. In 2023, the main subject of the lectures will be object detection.

Students entering the course have to have prior experience with deep learning and neural network architectures including Convolutional Neural Nets (CNNs), Recurrent Neural Networks (RNNs), Long Short Term Memories (LSTMs), and autoencoders. They need to have working knowledge of coding in Python, Python libraries, Jupyter notebook, TensorFlow both on local machines and on Google Cloud, and of GitHub or similar code hosting tools. The framework and associated tools which will be the focus of this course are PyTorch and Google Cloud. Students have to be self-sufficient learners and to take an active role during classroom activities.

There will be a few (3-4) assignments throughout the semester focusing on coding. In the second half of the course, there will be a midterm exam comprised of multiple-choice questions.

Final projects need to be documented in a conference-style report, with code deposited in a GitHub repository. The code needs to be documented and instrumented such that the instructor can run it after a download from the repository. A Google Slides presentation of the project suitable for a poster presentation is required.

Prerequisites

(i) Machine Learning (taken previously, or in parallel with this course). 

(ii) ECBM E4040 Neural Networks and Deep Learning, or an equivalent neural network/DL university course taken for academic credit.

(iii) The course requires an excellent theoretical background in probability and statistics,  and linear algebra. 

Students are strongly advised to drop the class if they do not have an adequate theoretical background and/or previous experience with programming of deep learning models. It is strongly advised (the instructor’s requirement) that students take no more than 12 credits of any coursework (including this course and project courses) during the semester while this course is being taken.

Registration

The enrollment is limited to several dozen students. The instructor’s permission is required to register.  Students interested in the course need to populate the SSOL waitlist, and MUST also populate the questionnaire.  The instructor will move the students off of the SSOL waitlist after reviewing the questionnaire.

(Tentative) Grading for the course (2023 Spring)

Assignments: 30%

Midterm Exam (Delivered at Week 11): 30%

Project (Final report & Code Repository): 40% 

(Potential) Class Contribution: x