cho.sh

researchmini

3,249 notes

260722260722 I prefer TUIBB9A08 MarkRight8B0501 에이전트 런타임 서버 셋업1983DC Bun을 Rust로 재작성하다2BDED3 260721260721 Agent Runtime Server SetupC38C2C 260719260719 AutoBuilder63849A 260718260718 Slack Bots are the new CLIs3E662F 260717260717 WristcomputerD98BDF 260716260716 SalpakanA2940F AGENTS.mdDFE522 260714260714 260713260713 260712260712 260711260711 인간의 보완재와 AX41F1E5 260710260710 tokenmaxxing cli635294 260709260709

Project Florence · cho.sh

Project Florence

Project Florence

Warning

This post is more than a year old. Information may be outdated.

Project Florence is a research project on VFL
Project Florence Literature Review

240329

Action items

compare mlp: what's the baseline performance, with full images?
"joint embedding"
autoencoder
- try to baseline. have the middle layers to have 40 pixels
- how good can we get to reconstructing the images?

Can we train an autoencoder, in an federated way?

do the full images and see how the reconstruction autoencoder goes?
use that autoencoder to see if it can reconstruct partial images
try to train the autoencoder in a federated way.

240320

Action Items

compare mlp

Backlinks19

240329 240326 240320 240307 240221 Project Calenzano 240118 Project Impruneta Project Malmantile 231003 231004 Project Fiesole 230922 230914 230912 230907 230828 230818 230602

Comments

do max everywhere → Result: Max did not work well at all

do avg(with-zero) everywhere

do avg(without-zero) everywhere → Result: close to avg-with-zeros.

random: slightly worse
squares: slightly better

use dropouts (with normal avg incl. zeros)

different neurons are being dropped out randomly
framework might handle this

max-avgnonzero doesn't work well

240307

CNN → filtered we 3 by 3
random, start with a bigger filter 7 by 7
play with some functions, such as max pool
we could do dropouts → forces the network to distribute more info → it may be more resilient
catastrophic forgetting → learning different tasks make the NN to forget the first one
Replay → include good examples from A when training B.
find the necessary weights of the NN after learning the first set,
Training the second one makes it very hard for the weights to move.
What about we do sum?? Max??
Sum/max in the first two layers and average in the final two?
Or what if we don't do the average over the zeros?

240221

MLP seems to do well

conv-pool-conv-pooling may be better

abstraction of mlp can ignore the noises?

action items:

cnn conv-pool-conv-pool
mlp and run it for 200 epochs
visualize scatters

240118

Randomize each sites to have scattered pixel data (instead of quadrants)
Autoencoder
Pivot project?

231116

8C3CC0

Project Impruneta

231017

Project Malmantile

231003

Project Fiesole

230922

Approaches to vertical federated learning
- Train a network that performs well when given all 20 input parameters
- Train a network that can make good predictions even with only 10 input parameters
- Start by testing a simple approach: train on half the data from each site, average the gradients, and repeat
- Run experiments on MNIST data by splitting images in half and training on the halves separately
Challenges
- Networks trained on half the data may not learn interactions between the two halves
- Training on single examples at a time is impractical; need to use batches for efficiency
Experiments proposed
- Train a network on full MNIST images to get a baseline performance
- Train networks on half MNIST images to see performance drop
- Train networks by alternating between the two halves of examples, averaging the weights, to see if performance improves over training on halves alone
The next steps if this approach works
- Try training on batches instead of single examples
- Experiment with training for multiple epochs instead of single iterations

230914

Can't we just ensemble them-?

Federated learning and split learning
- Discuss split training versus federated learning
- Federated learning converges to a better optimization point than ensembling independently trained models
- Split learning involves training parts of the network at different sites
Vertical partitioning of data
- Vertical partitioning of features across different sites can lead to poor individual predictors
- Training a model that combines the data in a more sophisticated way may perform better
- Focus on implementations that do not require training parts of the network at a central node
Next steps
- Look into existing implementations of split learning and vertical partitioning
- Focus on approaches using deep learning rather than classical models
- Assume the record linkage problem is solved and focus on the training approach
Action items
- Search for relevant papers that meet the criteria
- Filter out papers using classical models instead of neural networks

230828

Vertical Federated Learning
- The goal is to train a model using data from multiple sites without sharing the raw data.
- Each site may have different features/columns in their data, but some overlap.
- The challenge is training parts of the network using the data available at each site.
Record Linkage
- Matching records across sites to identify which records represent the same entity.
- Can be done using properties like name, address, phone number, and string similarity.
Inference
- Once the model is trained, inference is done globally using all available data for an entity, not just at one site.
Potential Conferences
- NeurIPS in May 2024 is a good target conference. Earlier deadlines may be too soon.
Meeting Plans
- Thursdays at 2 p.m. at ISI or remotely if needed.

Comments0

No comments yet. Be the first to share your thoughts.