5. Čemu žurba sa umjetnom inteligencijom?
o Jeftinija procesorska moć
o Više podataka
o Bolji algoritmi
… jer možemo
6. o Automatizacija donošenja
odluka je sada ekonomska
nužnost
o Vođeno bržim, mikro-
upravljanim, međusobno
povezanim, automatiziranim i
optimiziranim svijetom oko nas
o Neprekidno autonomno
donošenje odluka je već tu
Čemu žurba sa umjetnom inteligencijom?
7. o Odluke se donose s obzirom na pozitivne
i negativne procjene ishode projekta
o Pozitivni i negativni ishodi svode se na
(naučene) težine
o Relativno prema nekom sustavu
vrijednosti
o Vode prema ili od ciljeva i problema
-
+
Čemu žurba sa umjetnom inteligencijom?
8. o T
ežine su kodirane namjere
o Temelje se na nekom pogledu na
svijet, vezane su na trenutak,
kulturu, pravo, ekonomske ciljeve
i filozofsku perspektivu
o Dakle, autonomni sustavi su
kodirani namjerom
Čemu žurba sa umjetnom inteligencijom?
9. o Povezani lanac od softvera do
namjere
o Kako možemo narinuti pravila
koja će izradu koda i učenje
usmjeravati prema pozitivnim
namjerama za naše prijatelje i
negativnim namjerama za one
koji nam škode?
Čemu žurba sa umjetnom inteligencijom?
19. “Imamo mogućnost u idućem desetljeću učiniti velike
pomake koje muče čovječanstvo. Umjetna inteligencija
će biti tehnologija koja će voditi ka cilju. Imamo
moralnu obvezu realizirati ova obećanja uz kontrolu
negativnih učinaka. Mi to možemo, jer smo takve
stvari već radili.”
Ray Kurzweil
22. “Uspjeh u kreiranju umjetne inteligencije će biti najveći
događaj u ljudskoj povijesti…”
“Nažalost, to može biti ujedno i posljednje što
ćemo postići, ako ne naučimo kako izbjeći rizike.
Kratkoročno, vojne industrije razmatraju
autonomne sustave naoružanja koji samostalno
mogu birati ciljeve koje će uništiti.”
“…ljudi, ograničeni sporom biološkom evolucijom,
ne mogu se mjeriti sa umjetnom inteligencijom
koja će ih prestići”
Stephen Hawking
23. “Ja sam pobornik onih koji su zabrinuti zbog moguće
super-inteligencije. Prvo će strojevi raditi puno toga za
nas i neće biti super-inteligentni. Ovo je dobro, ako se
uspijemo izboriti za dobro. Nakon nekoliko desetljeća,
inteligencija strojeva će se povećati toliko da bi se
trebali zabrinuti. Slažem se sa Elonom Muskom i
drugima oko ove teme
Ne razumijem kako neki ljudi nisu zabrinuti.”
Bill Gates
24. AI je “najveća prijetnja našem postojanju…”
“Sve više naginjem prema stavu kako nam treba
nadzor, na nacionalnoj ili međunarodnoj razini, samo
kako bi spriječili neku glupost koju ćemo učiniti”
“Mislim kako postoji mogućnost opasnog ishoda”
(referira se na Google’s Deep Mind u koji je
uložio novac kako bi imao nadzor nad onim što
se radi)
Elon Musk
26. Više od 16,000 istraživača i vodećih
umova potpisali su otvoreno pismo
Ujedinjenim Narodima sa pozivom za
zabranom kreiranja autonomnih i polu
autonomnih oružja,
38. Kad pametni ljudi brinu, nekako
imam potrebu obratiti pozornost!
Pitanje svih pitanja je tko je proglasio
neke ljude pametnima i prema kojim
kriterijima?
Bez znanstvenog pristupa i kritičkog
propitivanja postojećih saznanja
nema napretka.
!
39.
40.
41. Ja-ROBOT & Asimovsa3 zakona robotike
Robot ne smije naškoditi čovjeku ili svojom pasivnošću dopustiti da se
čovjeku naškodi.
Robot mora slušati ljudske naredbe, osim kad su one u suprotnosti s
prvim zakonom.
Robot treba štititi svoj integritet, osim kad je to u suprotnosti s prvim ili
drugim zakonom.
42. Ja-ROBOT & Asimovsa3 + 0 zakona robotike
Robot ne smije naškoditi čovjeku ili svojom pasivnošću dopustiti da se
čovjeku naškodi.
Robot mora slušati ljudske naredbe, osim kad su one u suprotnosti s
prvim zakonom.
Robot treba štititi svoj integritet, osim kad je to u suprotnosti s prvim ili
drugim zakonom.
Robot ne smije naškoditi čovječanstvu ili svojom pasivnošću dopustiti
da se čovječanstvu naškodi.
43. Ružno (autonomna vozila & the trolley
predicament)
Etička pitanja pojavljuju
se kada se programiraju
vozila koja moraju
donositi odluku kod
koje je neminovna
ozljeda ili smrt ljudskog
bića, naročito kada
takvu odluku treba
donijeti u djeliću
sekunde o tome koga
će se ugroziti .
44. Ružno (nedostatak ljudi vodi do mehaničkih
njegovatelja)
Rješenja temeljena na
UI mogu poboljšati
zdravstveni ishod i
kvalitetu života za
milijune ljudi u
godinama koje dolaze,
ali samo ako steknu
povjerenje liječnika,
sestara i pacijenata.
45. Ružno (obrazovanje upravljano strojevima)
Trebaju nam živi
nastavnici za
kvalitetno
obrazovanje, ali AI
obećava pojačati
obrazovanje na svim
razinama nudeći
prilagodbu...
47. Ružno (nema posla za tebe – prekvalifikacija
je nužnost obrazovanja
Poljoprivreda (sad ispod 2%)
Masovno obrazovanje kroz osnovne
škole!
Je li to dovoljno danas?
48. Safe exploration - agents
learn about their
environment without
executing catastrophic
actions?
Robustness - machine
learning systems that are
robust to changes in the
data distribution, or at
least fail gracefully?
49. Avoiding negative side
effects- avoid undesired
effects on the
environment?
Avoiding “reward hacking”
- prevent agents from
“
gaming” their reward
functions
50. Scalable oversight - agents
efficiently achieve goals for
which feedback is very
expensive? For example,
can we build an agent that
tries to clean a room in the
way the user would be
happiest with, even
feedback from the user is
very rare
51. A Terrifying Nonprofit to For-Profit
Transition: Open AI Not So Open Anymore
April 17, 2019; Vox
A month ago, the nonprofit Open AI
announced a radical structural change. In
order to fulfill its mission of ensuring “that
artificial general intelligence (AGI)…highly
autonomous systems that outperform humans
at most economically valuable work—benefits
all of humanity,” it needed to transform itself
into a for-profit organization.
52. …i tako
o AI je tu i ubrzava
o Ekonomija je imperativ
o Donošenje odluka je programirano u svim proizvodima
o Odluke i etika, što s autonomijom?
o Briga donosi regulativu
Evolucija ili….
56. No-Code Machine Learning
• Quick implementation. Without any code
needed to be written or the need for
debugging, most of the time spent will be on
getting results instead of development.
• Lower costs. Since automation eliminates the
need for longer development time, large data
science teams are no longer necessary.
• Simplicity: No-code ML is easier to use due
to its simplistic drag and drop format.
57. No-code machine learning uses drag and
drop inputs to simplify the process into the
following:
• Begin with user behavior data
• Drag and drop training data
• Use a question in plain English
• Evaluate the results
• Generate a prediction report
59. TinyML
In a world increasingly driven by IoT solutions, TinyML makes its
way into the mix.
While large scale machine learning applications exist, their
usability is fairly limited.
Smaller scale applications are often necessary.
It can take time for a web request to send data to a large server
for it to be processed by a machine learning algorithm and then
sent back.
Instead, a more desirable approach might be to use ML
programs on edge devices.
https://hackaday.io/project/174575-solar-scare-mosquito-20
60. AutoML
Similar in objective to no-code ML, AutoML aims to make building
machine learning applications more accessible for developers. Since
machine learning has become increasingly more useful in various
industries, off-the-shelf solutions have been in high demand. Auto-ML
aims to bridge the gap by providing an accessible and simple solution that
does not rely on the ML-experts.
Data scientists working on machine learning projects have to focus on
preprocessing the data, developing features, modeling, designing neural
networks if deep learning is involved in the project, post processing, and
result analysis. Since these tasks are very complex, AutoML provides
simplification through use of templates.
https://auto.gluon.ai/stable/index.html
https://dynamicallytyped.com/stories/2021/openai-dall-e-clip/
61. Machine Learning Operationalization
Management (MLOps)
Machine Learning Operationalization Management (MLOps) is a practice
of developing machine learning software solutions with a focus on
reliability and efficiency. This is a novel way of improving the way that
machine learning solutions are developed to make them more useful for
businesses.
62. Understanding the ML systems lifecycle is
essential for understanding the importance
of MLOps.
1. Design a model based on business goals
2. Acquire, process and prepare data for the ML model
3. Train and tune ML model
4. Validate ML model
5. Deploy the software solution with integrated model
6. Monitor and restart process to improve ML model
https://mobidev.biz/blog/when-why-how-use-kubernetes-app-
development
63. Full-stack Deep Learning
What is full-stack deep learning? Let’s imagine you have highly
qualified deep learning engineers that have already created some
fancy deep learning model for you. But right after the creation of the
deep learning model it is just a few files that are not connected to the
outer world where your users live.
As the next step, engineers have to wrap the deep learning model into
some infrastructure:
• Backend on a cloud
• Mobile application
• Some edge devices (Raspberry Pi, NVIDIA Jetson Nano, etc.)
64. Generative Adversarial Networks (GAN)
GAN technology is a way of producing stronger solutions for
implementations such as differentiating between different kinds of images.
Generative neural networks produce samples that must be checked by
discriminative networks which toss out unwanted generated content.
Similar to branches of government, General Adversarial Networks offer
checks and balances to the process and increase accuracy and reliability.
It’s important to remember that a discriminative model cannot describe
categories that it is given. It can only use conditional probability to
differentiate samples between two or more categories. Generative models
focus on what these categories are and distribute joint probability.
65. Unsupervised ML
Unsupervised ML focuses on unlabeled data. Without guidance from a
data scientist, unsupervised machine learning programs have to draw their
own conclusions. This can be used to quickly study data structures to
identify potentially useful patterns and use this information to improve and
further automate decision-making.
One technique that can be used to investigate data is clustering. By
grouping data points with shared features, machine learning programs can
understand data sets and their patterns more efficiently.
66. Reinforcement Learning
In machine learning, there are three paradigms: supervised learning,
unsupervised learning, and reinforcement learning. In reinforcement
learning, the machine learning system learns from direct experiences with its
environment. The environment can use a reward/punishment system to
assign value to the observations that the ML system sees. Ultimately, the
system will want to achieve the highest level of reward or value, similar to
positive reinforcement training for animals.
This has a great deal of application in video game and board game AI.
However, when safety is a critical feature of the application, reinforcement
ML may not be the best idea. Since the algorithm comes to conclusions with
random actions, it may deliberately make unsafe decisions in the process of
learning. This can endanger users if left unchecked. There are safer
reinforcement learning systems in development to help with this issue that
take safety into account for their algorithms.
67. Few Shot, One Shot, & Zero Shot Learning
Few shot learning focuses on limited data. While this has limitations, it
does have various applications in fields like image classification, facial
recognition, and text classification. Although not requiring a great deal of
data to produce a usable model is helpful, it cannot be used for extremely
complex solutions.
Zero shot learning is an initially confusing prospect. How can machine
learning algorithms function without initial data? Zero shot ML systems
observe a subject and use information about that object to predict what
classification they may fall into. This is possible for humans. For example,
a human who had never seen a tiger before but had seen a housecat
would probably be able to identify the tiger as some kind of feline
animal.