Research

Jialin Lu

So here I am going to write about what I’ve done and what I am working on. Each comes with a description, or several slides.

Overview

Click to expand

Undergraduate


Click to expand

When I was taking my undergraudate in Zhejiang University, I was working on disentanglement1 Disentanglement of Hetergeneous Components: A Slightly Obsolete Attempt ) and also briefly on generative modelling2 A Review of Failure - Deep Generative Model for Chinese Fonts.

Master


Click to expand

I then came to study with Martin Ester, and I have mainly worked on the topic of interpretability.

An active learning approach for model interpretation

Lu and Ester (2019a)Lu, Jialin, and Martin Ester. 2019a. “An Active Approach for Model Interpretation.” In NeurIPS 2019 Human-Centric Machine Learning Workshop.

Click to expand (The slide can be controlled by ← and → on keyboard )

Revisiting and debugging the recurrent attention model

Lu (2019)Lu, Jialin. 2019. “Revisit Recurrent Attention Model from an Active Sampling Perspective.” In NeurIPS 2019 Neuro AI Workshop.

Click to expand (The slide can be controlled by ← and → on keyboard )

Discovering modular groups of neurons to interpret DNN

Lu and Ester (2019b)Lu, Jialin, and Martin Ester. 2019b. “Checking Functional Modularity in DNN By Biclustering Task-specific Hidden Neurons.” In NeurIPS 2019 Neuro AI Workshop Workshop.

Click to expand (The slide can be controlled by ← and → on keyboard )

Domain knowledge-guided architecture for drug-response prediction

Snow et al. (2021)Snow, Oliver, Hossein Sharifi Noghabi, Jialin Lu, Olga Zolotareva, Mark Lee, and Martin Ester. 2021. “Interpretable drug response prediction using a knowledge-based neural network.” In KDD’21.

Click to expand

We utilize the hierarchical information on how proteins form complexes and act together in pathways, in order to form a specific architecture.

Neural Disjunctive Normal Form.

master thesis.

Click to expand (The slide can be controlled by ← and → on keyboard )

The key technical challenge here is how to devise algorithms to optimize discrete parameters by gradients. Here I made an introduction and survey on how we can do this3 On more interesting blocks with discrete parameters in deep learning .

Current


Click to expand

After working on the Neural Disjunctive form however, I now have two further directions stemmed from it.

Direction 1: general inference algorithm


Click to expand

Given a domain specific language, find out a good general purpose inference algorithm.

It can be but not limited to gradient-based optimization. Maybe mixed integer programming, blackbox variational inference, etc.

Direction 2: Symbolic grounding problem


Click to expand

In the framework of vertical integration, learning of neurally-extracted symbols that has human-aligned meaning is actually very hard.

SAT-Net ((2019Wang, Po-Wei, Priya Donti, Bryan Wilder, and Zico Kolter. 2019. “Satnet: Bridging Deep Learning and Logical Reasoning Using a Differentiable Satisfiability Solver.” In International Conference on Machine Learning, 6545–54. PMLR.), ICML best paper) utilizes a symbolic SAT solver layer on top of a convolutional neural network (CNN) for solving a satisfiability-based task of visual Sudoku The original work claims that the CNN can learn to detect MNIST digits and the SAT-layer can solve this sudoku problem. However, a recent NeurIPS paper (2020Chang, Oscar, Lampros Flokas, Hod Lipson, and Michael Spranger. 2020. “Assessing SATNet’s Ability to Solve the Symbol Grounding Problem.” Advances in Neural Information Processing Systems 33.) re-assess this work and find out that in this case, the learned neurally-extracted concepts are not really meaningful as it should be. That these concepts are not really corresponding to digits.

This phenomenon is also confirmed by a lot of relevant approaches/papers in this hybrid integration framework.

I currently have an idea and am working on analyse this problem from a noise-to-signal perspective.