Meetup STEM4health-Berlin 18/04 - Deep Learning: Success is guaranteed?

gistfile1.txt

# Meetup 18/04 - Deep Learning: Success is guaranteed?

Roland Vollgraf, Zalando
Loris Bazzani, Research Scientist @ Amazon
Djork-Arné Clevert, Bioinformatics @ Bayer

Recording for internal purposes.

grants4apps.com/berlin
50,000 euros funding for project
2 minutes to pitch your project in front of everyone

## (Loris Bazzani) Deep Learning: a computer vision perspective

lorisbazzani.info

Machine Learning box. Deep Learning as machine learning box
Train model with annotations.
Very popular between computer vision people.
Computer vision applications

Image recognition = give an image and receive a label
Object detection inside the image. find the dominant objects. For finding, you have to train the algorithm.
Image captioning. give me a description of an image. e.g. “man in black shirt is playing guitar"

Deep learning is not very new. Neural networks exist for at least 30 years. GPUs are useful because are a good way of efficiently multiplying matrices.

Large Annotated Dataset - ImageNet collects annotated images. image-net.org
You can train your model using these images and GPUs
Research at spring.com shows that ImageNet’s algorithms is bringing very similar results to human eye.
The advent of ConvNets was the major improvement in the last decade

there are many open source softwares for using it

Deep Learning = Learning Hierarchical Representation. Detect compositions of an image. Remove color, taking patterns

LeNet
2 Convolution
2 MaxPool
3 Fully-connected
Output

MSRNet
1 Maxpool
6 Convolution
1 Maxpool
6 Convolution
1 Maxpool
6 Convolution
...

To go from image to video you go from 2d to 3d convolutions
example of algorithm detecting what’s (sport) being shown in the video

in the first, the “current” frame and the one before
in the second model, he considered clips of videos (pieces of many pictures)

with: deep learning

- less supervision. e.g. more unsupervised learning of networks
- online learning and update of the network. e.g. reinforcement learning.
- attentional models for filtering useless information

he has a paper of usages of deep learning.

challenges

- currently, we’re able to train an algorithm to do something very specific.
- perform machine learning techniques consuming less energy, like the human brain (and not many GPUs like the computer who won the Go game)

humans learn by comparing and checking for similarities

## (Roland Vollgraf, Zalando) Fashion DNA a Coordinate System in the Space Fashion

they wanted to recommend products to customers
check what similar customers bought. chicken-egg problem because you find how similar they are by finding similar products

we can leverage meta information for finding similar products to new ones. similar products are bought by the same kind of person
they analyze the image to find the meta information

input -> omage
AlexNET

recommendations for user who buys just dark clothes: more dark clothes
for user who buys shoes and shirts with figures of animals: more of the same

Fashion DNA map used tSNE mapping of F
example of clustering of products. shoes are in the same cluster. inside, children shoes. close to children clothing
strange things may happen: images of shorts photographed with the same legs were grouped. shorts without legs were also grouped.

were inputted as parameters too:

- price cluster
- comomodity group
- color code
- brand pattern code

in the end, considering attributes was better than considering images themselves.
when combined, were the better one. from ~80% to almost ~90%
FDNA is a meaningful feature representation for fashion items

when a user buys one black t-shirt, he doesn’t want another. was not addressed in this study, but attribute-based analysis would help to mitigate the problem.

in clothing you can’t many groups. t-shirts of super heroes is one of the few examples.

## (Djork-Arné Clevert) Decoding Biological Data with Tectified Factor Networks

Deep Learning is key technology for image/speech processing
Google, Facebook… are using Google Voice, translate
Nature and (another magazine) covered deep learning in 4-page articles
is the most searched algorithm/method of ML

Microsoft, Amazon, Google and Facebook are learning larger databases

startups using it used to raise 17M in 2011. now is 1500M
older than 30 years. probably 60

what is? find low level representations of images. try to find more complex patterns

Rectified Factor Networks
new data readly coded

receive input, factorize and output
Google trained an unsupervised algorithm to detect a face, a horse, a car
finding similar pictures of the same car
finding similar pics of horse may find a unicorn

RFN code is publically available GitXiv

@ Bayer

with convolution networks

- predicting dug activity based on profiles
- classification of histopathological images
- learning molecular fingerprints
- bioactivity prediction in structure-based drug discovery

unsupervised

- detection of drug off-target effects
- biclustering to identify deregulated pathways

supervised

     (3 things)

deep learning won the tox21 challenge
not so deep. had just 3 layers
99% neurons associated
97% neurons associated
90% neurons associated

determined structure in the molecule was found to be related with toxicity

training with genomics were able to separate (somehow) africans, european caucasian and asian (these, even between chinese and japanese)