6. TODAY’S ARTIFICIAL INTELLIGENCE IS
❑ POWERFUL
❑ ACCESSIBLE TO ALL
https://www.healthcareitnews.com/news/new-ai-diagnostic-tool-knows-when-defer-human-mit-researchers-say
https://www.infosearchbpo.com/3d-lidar-annotation.php
12. ❑ Philosophically no body, no childhood
and no cultural practice, computers
cannot acquire true intelligence:
https://www.nature.com/articles/s4159
9-020-0494-4#Sec8
❑ Technologically, self-programming
AI demands vast data across
multiple branches, tremendous
computations, and complex math
models. Not currently feasible.
❑No reliance on human programming to
learn and do.
❑Like humans, general AI can adapt to its
environment.
General AI are self-programming AI
Properties
Availability
13. Narrow AI for a single task, where knowledge gained will not
automatically be applied to other tasks
Rule-based AI Machine Learning AI
Some experts argue
that this is not AI!
https://www.ck12.org/book/ck-12-basic-geometry-concepts/section/2.3/
Types
❑Search-driven AI
❑Expert System
https://www.researchgate.net/pu
blication/330217507_Application
_of_machine_learning_in_rheum
atic_disease_research
Its Trendy Subfield: Deep Learning
Found in
❑Gaming
❑Management Systems
❑computer algorithms that can improve automatically through
experience and by the use of data
14. RULE-BASED AI
Gaming AI – gfycat/witcher
By Nbro, https://commons.wikimedia.org/wiki/File:Animation_of_alpha-beta_pruning.gif
15. Search-driven
Simple Chess AI - freecodecamp
❑ Decision-making by solving a search problem
based on heuristics or mathematical
reasoning.
❑ Navigate through trees of possibilities to find
the best possible outcome
→ partial game trees to make
computation feasible
https://giphy.com/explore/pathfinding
16. Expert Systems simulates the behavior and judgement of
human experts.
❑ Knowledge base: Knowledge, rules and form procedures of
the domain.
❑ Rules engine: Function to get relevant data from the
knowledge base, interpret it, and to find a solution.
❑ User interface: Function to allow a non-expert user to
interact with the expert system and find solutions.
❑ Knowledge acquisition and learning module: Function to
acquire more data from various sources.
https://www.mygreatlearning.
com/blog/expert-systems-in-
artificial-intelligence/
❑ For non-experts to gain information.
❑ Used in various areas of medical diagnosis,
accounting, coding, gaming and more.
Properties
Components
https://www.javatpoint.com/expert-systems-in-artificial-intelligence
19. Machine learning uses an algorithm to learn and generalize
from historical data in order to make predictions on new data.
Machine Learning Rule-based AI
is probabilistic is deterministic
adapt in accordance with
training information streams
require manual data analysis
and modification of rules
needs full demographic data
details of the domain
needs experts to set up
objective rules
changing parameters hard-coded rules
20. Supervised learning
Formulation:
Given an input set X and the corresponding output set Y,
supervised learning involves learning a function F
such that F(X) = Ẏ matches Y as much as possible.
Types:
https://www.ceralytics.com/3-types-of-machine-
learning/
21. Object Detection
❑Regression: Fitting bounding boxes
to image points.
❑Classification: Identifying object in
vehicle
https://alexeyab84.medium.com/yolov4-the-most-accurate-real-time-neural-network-
on-ms-coco-dataset-73adfd3602fe
https://www.youtube.com/watch?v=nw1GexJzbCI&ab_channel=TzuTaLin
Intense Labeling
Great results
24. Reinforcement learning
Robotic simulation
https://gfycat.com/gifs/tag/sethbling
Advanced Gaming AI
https://www.freecodecamp.org/news/a-brief-
introduction-to-reinforcement-learning-
7799af5840db/
Simplification:
Given an environment E and a set of allowed actions A,
the reinforcement learning model M learns to maximize a
cumulative reward function F.
It does so by producing a sequence of actions (trial) 𝐚𝟎, 𝐚𝟏, 𝐚𝟐, …
Whenever a trial fails, F is penalized such that M is tuned to
produce a better trial. Otherwise, F accumulates rewards.
❑ Learning to take suitable actions to
maximize reward in a particular situation
through trials and errors.
❑ Involves actions, states and reward
functions more than just inputs, outputs
→ Balancing exploration and exploitation
25. Rewards can be exploited
https://boingboing.net/2020/01/11/optimizers-curse.html
https://gfycat.com/gifs/search/reinforcement+learning
Not safe to test how tough your vehicles are!
27. Deep Learning
The Universal Approx. Theorem
a feed-forward network with a single hidden layer containing a finite number of
neurons can solve any given problem to arbitrarily close accuracy as long as
you add enough parameters.
Thanks to
❑ Neural Networks → Indefinitely Flexible
❑ Gradient Descent → The tractable optimizing technique
❑ GPU → The actual computing technology that allows parallelization on Big Data
https://www.montreal.ai/ai4all.pdf
Forward Inference Backward Propagation
is inspired by neural networks of the brain
to build learning machines
28. F. Wang, M. Zhang, X. Wang, X. Ma and J. Liu, "Deep Learning for Edge Computing Applications: A State-of-the-Art Survey," in IEEE Access, vol. 8, pp.
58322-58336, 2020, doi: 10.1109/ACCESS.2020.2982411.
Structures of different deep learning models.
29. https://www.montreal.ai/ai4all.pdf
Rotation and translation of a GAN-generated car using GIRAFFE (created by author using https://github.com/autonomousvision/giraffe, MIT License).
Deep Learning –
An Example
Advantages
Disadvantages
❑ Approximating complex functions
❑ High accuracy
❑ Many existing frameworks and codes
❑Needs a lot of data for training
❑Domain changes requires more data
❑No clear mathematical understanding of
parameters yet
❑Needs much GPU capabilities
31. Know your
direction
Image by Jash Rathod https://pub.towardsai.net/branches-in-artificial-intelligence-to-transform-your-business-f08103a91ab2
32. Know your
language
The majority of AI
applications can be
easily written in
Python
Thanks to their flexibility and great efficiency,
you can push certain boundaries with C/C++
33. Know your
framework
❑ Great Google Community
❑ Strong API
❑ Fast Inference
❑ Research-driven
❑ Very Pythonic
❑ Many Easy-to-Understand Tutorials
35. Know your
trade-offs
Due to domain complexity,
there has always been a major dilemma
between speed and accuracy
https://www.researchgate.net/publication/328509150_Benchmark_Analysis_of_Representative_Deep_Neural_Net
work_Architectures