# AMLD notes
applied machine learning days

Keynote 1

Switzerland leads AI publications per capita by a factor of 2.5.
## Sergei Yakneen
CTO isomorphic labs (rational drug design with AI)

-alphafold 2

Moores Law vs Erooms Law
3B dollars per drug
90% failure rate (after 6 years in lab)

target discovery
target validation
it identification & hit to lead
lead optimisation
pre-clinical
clinical trials
drug

three bottlenecks
- understand desease biology
- compound design
- asserting compound quality

fails are LATE
successes are not transferrable

molecular docking simulation between protein and ligand

quantum mechanics
learned forcefield
molecular dynamics

orthogonal validation
computational biology (macro simulation, cell and upwards)

48 layer evoformer

ML reserach scientists
medical chemists
devopss
product managers
product engineers
data engideers
.... a village

### Q&A
- partners useful
- prediction aids toxicity, efficacy, and in biological stage: refine bio-markers per patient
- biggest bottleneck: compute is always insufficient, data too, biological data is hard to work with. how to make data digestible for ML Talent attraction.
- potential advantages of Lausanne location (new office in construction): science feeding industry. Labs, people, infrastructure, talent.

## Raphael Conz
- general manager, office of economic affairs and innovation (SPEI), state of vaud

### Lausanne: home of innovation video
vertical farms
3d prniting
exoskeleton
brain scan
smd electronics
robo farming
high altitude place
right innovation environment

EPFL: 400 AI researchers

Regional successes

# from strategy to execution: leveraging ai across industries

visium - ai consulting
applied ai transformation
236B units annually

cv for fault detection of 10M units daily
in-depth in-house ai infrastructure

scalable foundation
hub-and-spoke
people MUST want to use the tech

AI is a long-term investment
start early and learn while tuning into fitness

learn from learning

### alberto barroso, tetra pak
data maturity: advanced

industrial optimisations
new business models
environmental goals

hub-and spoke organisation
customer focus >> new business models

understanding incremental evidence in management
evidence-based decision-making

closing tighter loops around prosesses and customers
decision-making faster and evidence driven

# open-source ai
how free should OS be?
regulation or no?
not exactly os license anymore
should the user be enforcing compliances?
beyond licenses
governance challenges
regulation?
- OS AI may not be able to be free
- behavioural use restrictions
- legal, technical, ethical frameworks

### Q&A

## Leandro von Werra
Chief Loss Officer at hugging face
github/hf hub/x: lvwerra

fully open:
- bloom
- olmo
- starcoder2
- bigCode
- theStack v2
- am i in the stack (check if my code is included)

## Martin
meditron, open medical LLM

2012: ImageNet works
2023: nearly every task works

# explainable AI
## Alan - centai.eu

black boxes are much like brains. opening them up does not reveal stored information easily.
failure mode is obscure. what is confidence based upon?
canine classification: wolf vs. husky. snowy ground >> wolf!
bias reinforcement: doctors are boys, nurses are girls
amazon employee cv filter: no women,

#### EU AI act
biometric id prohibited
high risk models require auditioning
high-risk use cases need to argue for reasoning (XAI)

#### Explainability
- poorly defined concept
- no consensus on metrics for comparison
- gap with normative requirements
- different purposes and users
- plethora of competing solutions
- different application domains

## Rita P.Ribeiro
metro porto: equipment failure by detecting anomalies

sensor stream -> autoencoder -> (filter) -> detect anomalies -> alarm
add layer for explanations

stream + anomaly -> rules for explanation -> explanation

compressed air production unit (no redundancy)

two failures in three months of training data (zenodo)

using three auto-encoders (WAE GAN, TCN AE, LSTM AE)
TCN (early detection, hysteric)
WAE (early detection)

training on the first month

high reconstruction error, prioritized rules

## Tim Rattay - prediction of radiotherapy side effects ...
rule extraction with fidex and fidexGlo

tabular data
QSVM
hidden bias layers
three hidden layers to sigmoid (smoothen) the activation function.
combining models into ensembles > causes loss of interpretability.
heuristic decision model with hyperplane localization (adds explainability)
transform decision trees into rules
insert rules into DIMLP networks
random forest > DIMLP
"characterization of symbolic rules ambedded in deep DIMLP networks: a challenge to transparency of deep learning."
"a comparison study on rule extraction from neural network ensembles, boosted shallow tres and SVMs"

### Fidex
local rule example.
- discrimination of hyperplanes
- complexity: (dimensionality of classification problem) * (number of training samples) * (max number of rules) * (..)

#### FidexGlo
10000 samples = rules
discard superfluous ones

##### Q&A
rule-fit algorithm vs. DIMLP
linear combination of rules, less interpretable, since proportional mixes do not relate to decisions

## Customer Reviews vs. Repurchasing decisions
sentiment + projection
bank load example (explanation for decision)
explain for a group of nodes (customers)

## multi-modal GAI
suffers from inexplainability
sequential fusing?
specialized pipelines
robust to missing data
flexible addition of new encoders
inherently interpretable

github.com/epfl-iglobalhealth/multimodn

## Timur Sattarov - bank order field corruption repair
Deutsche bundesbank
detect when which field breaks predicted value. relate to nearest neighbours
add artificial noise and let the model learn the difference to real-world noise
boost in performance
UMAP auto encoder

"walk in the data" graph representation anomaly grouping?

## Claudio Fiandrino - Explainable AI for 6G mobile networks
IMDEA networks institute
5G optimized for internet of things
6G (2030) is optimizing for connected AI
- ai-native air interface
- PHY (connection modulation, self-configuration, error-correction, resource allocation, ...)
- MAC
- ai-native architecture
- automation
- self-management
- SHAP
- decision trees

# AI Literacy

partici.fi/69446318

## AI literacy to Protect Public Power
Katharina Schüller
Tania Duarte (We and AI, Consultancy)

## Building AI literacy as a driver for innovation examples from a german authority

typical software procurement - one size fits nobody - local development - functional deviance

what if we turn the weakness into strength

"the management is working against us"
"we build our own roque solutions"
"job is great, apart from IT"
"we get information, but not how we need it"

moving innovation to local contexts: huge potential
local development encouraged

data health dashboards
report intervals
self-organized maps for DB health

## AI Literacy in the education field
Lesley Wilton

AI collision avoidance system malfunction

# Bridging the Gap: Solving Real World Problems with Open Research
Antoine Marot (RTE)
Irene Sturm {DB InfraGo}
Daniel Boos (Swiss Federal Railway)
Erik Nygren (Flatland)
Julia Usher (Zurich Uni of applied sciences)
# Appeal of open research: a community perspective
Jeremy Watson

operations research project with AI
Flatland Community
trains running on tracks
reduce problem with abstract perspective (train scheduling)
looks simple
solution is hard

### Community model
conferences, meet-ups instead of prizes
fewer deadlines, longer timescales, thoughtful
collective endeavor

right size teams
hunter-gatherer specialization nirvana

## Translational Research
Julia Usher (ZHAW)

open research, why?
enables dissemination
tech advancement = biological evolution

## Human in the loop: How to integrate AI in daily operations
Daniel Boos

AI impact scales
- society
- organization (talk topic)
- human

task at seams between them

automation vs hybrid

past:
- manual
- SCADA
- HI (rules, control) "Human machine collaboration" real-world impact
- full auto (boundary control)

explainability per stake holder (collaborator)
system automation vs. crisis coping quality by humans (irony)

HI
- occupational perspectives
- knowledge domains
- how to design targeted explanations for those users?
- how to mitigate risk with proper management of control, accountability

gap: how to align operations with transformational mutations?

mockups & prototypes, frameworks & toolkits, interactoinal expertise & personas (UCD)

toolkits
- MS hacks (UX exploration)
- ...many more

## Panel discussion
Q how combine research domains? how to acquire domain expertise easier next time?
- communicate with experts, read (erik)
- common language difficult to introduce ML terminology. how to coordinate. communicate, abstract, simplify (irene sturm)
- project dimensions. find common ground and language. derive generic scenarios. balance abstraction and detail (antoine)
- what they care about. what si written down, what left out. visualise, make tangible. (Daniel)
Q as a researcher, what makes a question interesting
- use case. KPIs. what the problem is often is ill-defined (julia)
- issues may be interpretative and definition may be inherently hard (daniel)
- negotiate it out of the stakeholders (irene)
- progress in research: evaluation, operationalization, milestones. tradeoff may be vague. (antoine)
Q limits to publication (specific problem may be suppressed)
Q open research and crowd-sourcing. companies profit a lot. how does it feel?
- gratification in solving. validated. (erik)
- want ot have positive invluence on humanity. go private if it is for the money (julia)
- side projects. personal commitment. rewarding, not cheap. want awareness and recognition in management (irene, industry)
- side projects. need publicity to free resources. broken continuity. structural benefit OS/research community carry torch between industry involvement. collaborative platform across organisations. (antoine)
- competence building through interaction and tech trials. (daniel)
Q why are they side projects, even if valuable?
- digital environment tinkering. mature enough for integration into operations. (antoine)
Q is it challenging to work with partners treating your project as side-project?
- github etc lengthen lifecycles considerably. outlasts employment contracts. sunk cost is good in open research (erik)
Q UCD user stakes vs. collaboration stratefy
- empathy
- being embedded, mingle. subtle clues have high value (erik)
- define goal first. accept different stakes (irene)
- more critical. too deep, too high. gauge for level of discussion. someone should facilitate you. and vice versa. moderate collaboration continuously. beware meeting cultures. (daniel)
- takes a village to make things great. bring humans and disciplines together. its more than an algorithm. there are different human types too. "applied communication". explore HCI concepts. (antoine)
Q human-in-the-loop muddles initial research design
- situational awareness. what is improving. system goals. explore before committing to design. control can be defined, evaluation is explorative. (daniel)
- training is similar. evaluate if progress when training users. interesting challenge, research needed. (antoine)
Q bring people towards problem definition. avoid tech positivism (solution-first). how to think backwards with stake holders?
- participatory design. workshops. create inventory. break down problem into weighted domains. not easy to apply and time intensive. brainstorm first. forget operations. be creative for system review. (antoine)
- involve fresh minds. mix industries and backgrounds avoid status-quo staff only. (irene)
- who is open. be selective. ask around. find real users. (daniel)

# AI-Powered Projects: Your Innovation Journey

## Basem & Baher Higazy

### ConsultoPATH: developing the world's first AI for patient pathway design
optimize patient journeys for complex patients (80% of resources (40ppt wasteful) for 33% of patients)

llm limitations, use knowledge graphs to minimize RAG source hallucination, generate pathway (UML flow diagram)

pathways can be specialized and iterate through domain experts. wicked problem eveolves under solution.

Q: specialized models better than GPT
- mistral small better than GPT4 if specialized.
- avoid generalism
Q: transferrability?
- graph for other contexts. ecology of agents plus humans in the loop.
- allow feedback and transparency
- debug in production

## Reinventing Advertisement with GenAI
Leila Delarive

what if advertising was information?
empower local economy to communicate values and efforts.

incorporate weather and life-data into generated ads
asks for local assessment of AI generated ads. (reinforcement of your interests)

Q: how to ensure standing out of other genAI ads?
- fine tune your advertising
- customization (human touch)

## AI-Powered Translation Software For Legal, Tax and Finance
Paula reichenberg

### Innovation Journey
what was the spark?

observe and advice. felt like pushing them each time.
advising vs. acting

custom models for (parts of) documents
best candidate selection
best quality with human in the loop

model agnostic

national hosting with best international models

mind distance between customer journey and SOTA tech (join the wave, but select mature foundation)

8 people, 3 engineers, hiring!
## Uncharted territories: Learning from a challenging AI collaboration
Daniel Dobos

goal to build collaborations
applied research implementation

## Shaping your innovation in the context of Generative AI
Robert van Kommer

# keynote
cerebras - wafer-scale chips
jean-philippe fricker "JP"

350 people