27
SCS vision meets reality supercomputing systems 22 nd ieee real time conference ai : a gentle introduction for «smart dummies» by christof a. buehler 10/16/2020 1 may its foce be with you ai :

A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

S C S

vision meets reality

supercomputing systems

22nd ieee real time conference

ai : a gentle introduction for «smart dummies»

by christof a. buehler

10/16/20201

may its foce be with you …ai :

Page 2: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

supercomputing systems : one slide profile

founded in 1993 by

prof. dr. anton gunzinger

130 employees

service provider for

industry & academics

ips go to customer

our origin

Page 3: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/20203

ai: where are we?

frankenstein 1818 ex machina 2015

«ai enables machines to mimic human

behavior»

2020 thalamocortical system

Page 4: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/20204

frontiers in ai : where are we?

marketing

reality

Page 5: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/20205

reading zip codes

ai : the rise of deep learning

facts of dl • today’s state-of-art ml methods

• ≥ human accuracy in many tasks

o image recognition

o med. diagnostic

o nlp

o games

2012

ImageNet challenge

deep learning support vector machine random forest

sear

ch in

tere

st

copy from brainbreak through of dl: 2012

machine learning

deep learning

artificial intelligence

Page 6: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/20206

dl : what is deep learning?

brain≈100b neurons

dl ≈ deep neural nets +big data +efficient algorithms

brain neural network vs. dnn

≈100m

≈ 200+ layers

Page 7: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/20207

neurobiology inspired

• net of interconnected neurons

• building block (neuron / perceptron)

dl : what is deep learning?

f

learning adjusting weights (i )

x1

x2

xn

f f ( xi i )i = 1

n

1

2

n

out

artificial neural network

input

layer

hidden

layer

hidden

layer

hidden

layer

output

layer1 2 3

Page 8: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/20208

how do nns learn?

neural net : learning

input

layer

hidden

layer

hidden

layer

hidden

layer

output

layer1 2 3

forward path

• input data to net:o images, movies

o voice

o text ..

• propagate data through net

forward

Page 9: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/20209

how do nns learn?

input

layer

hidden

layer

hidden

layer

hidden

layer

output

layer1 2 3

backward path

• back-propagate loss

prediction - truth

• adjust weights ()

gradient descent

• learning adjusting weights () & min. loss

backward

neural net : learning

Page 10: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/202010

how d-neural nets see …

hierarchical features

low level edges, corners

mid level circles, rectangles

high level face, flower …

from detail to whole

learning abstractions

Page 11: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/202011

classify challenge

14 mio images

20k image categories

• human, cat, dog

• plane, car, knife

• ...

ImageNet - huge image pool

how good are dnns?

Page 12: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/202012

how good are dnns?

ImageNet challengebreak through of dl: 2012 classify challenge

• 14 mio images

• 20k image categories

• nn surpass humans in 2015

top nets today • cnn, gcn, gan, xai

• transformers

• but …

Page 13: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/202013

ai conclusion : programming becomes training

icons by www.flaticon.com/authors/turkkubhttps://twitter.com/matthewfitch23

Page 14: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/202014

ai : a word of caution …

are powerful …

• non-linear modeling

• discovering concepts

• mimicking the brain… at all!

… it’s often a branding term

deep neural nets

https://www.pinterest.ch/pin/377950593709450799/

but ...

Page 15: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

scs : selected customer projects

ai

Page 16: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : patient monitoring system

partner e. keller, usz zürich, eth - & uni zürich

icu-cockpit : its focusicu : intensive care unit

high-resolution raw data

for medical research

robustness & safety

Page 17: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : patient monitoring system

key functionalities

• data visualization

• build-up research db

• alarm handling

o set thresholds

o classify

o motion detection

reduced false alarms …

partner e. keller, usz zürich, eth - & uni zürich

Page 18: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : patient monitoring system (prelim. results)

ai• deep architectureo lstm, transfer, multi-task

• gt pbt02

gt

false alarms false positive rate

ROC

det

ecte

d a

larm

s

tp-r

ate

minutes before incident

mean alarm probability

det

ecte

d a

larm

s

tp-r

ate

predict brain hypoxia results

partner e. keller, usz zürich, eth - & uni zürich

Page 19: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : deep eye ophthalmology

the oct microscope

• michelson interferometer

• 3d-imaging @ µm

optical coherence tomography segment the eye

b-scan

a-scanb-scan

a-scan

b-scan

partner dr. peter m. maloca - iob in basel

Page 20: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : deep eye ophthalmology

convolutional neural net

cnn u-net

train gt

f1score vitr. & retina 98%

choroid 92%

choroid

choroid

vitreous

segment the eye solution

partner dr. peter m. maloca - institute of molecular and clinical ophthalmology in basel

Page 21: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : deep eye ophthalmology

f1score tumor 65%

note: very small gt-set!

choroid

tumor

segment the eye solution

convolutional neural net

cnn u-net

train gt

partner dr. peter m. maloca - institute of molecular and clinical ophthalmology in basel

Page 22: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : deep eye ophthalmology

cnn vs. expert

ophthalmologists

• our net performs on-par

with experienced experts

paper published ✓

experts cnn vs. experts

vitreous retina choroid sclerainte

rsection-o

ver-

unio

n

reading center retina experts

machine better than experts? results & answer

partner dr. peter m. maloca - institute of molecular and clinical ophthalmology in basel

Page 23: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : ai-assisted hip screening

from analog to digital

we have a dream …

ai-assisted dhh diagnosis

𝜶

β

partner dr. med. st. essig (svupp), dr. med. th. baumann (smopp)

Page 24: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : ai-assisted hip screening

ai-approaches

α,β lines landmarks end-2-end

α 64°

β 71°

angle detection via

landmarks

lines

cnn net

α 64°

β 71°angles

bone segm.

partner dr. med. st. essig (svupp), dr. med. th. baumann (smopp)

Page 25: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

case : ai-assisted hip screening

results

• u-net + psp & resNet50

• learning direct angle for α

top-2 dnn architectures

u-net with

psp-net

resNet 50

[1] D. Golan et al., Fully Automating Graf’s Method for DDH Diagnosis Using Convolutional

NNs, 2016

[2] E.A. Simon et al. Inter-observer agreement of ultrasonographic measurement, 2004

u-net achieves expert

accuracy with end-2-end

best deep approaches

partner dr. med. st. essig (svupp), dr. med. th. baumann (smopp)

Page 26: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/202026

happy experimenting with ai-tools

«labelling cats &dogs»

http://ai-demo.scs.ch/#/dashboard

http://ai-demo.scs.ch

Page 27: A gentle Introduction to AI · 2020. 10. 22. · •u-net + psp & resNet50 •learning direct angle for α top-2 dnn architectures u-net with psp-net resNet 50 [1] D. Golan et al.,

10/16/202027

ai : a gentle introduction

q & a ?

«gentle introduction to ai»