Scribd
Upload
10 views35 pages

Neral Introduction

Uploaded by

João Paulo Dellasta do Nascimento
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
10 views35 pages

Neral Introduction

Uploaded by

João Paulo Dellasta do Nascimento
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 35

ITCS 6156/8156

Machine Learning
Instructor: Hongfei Xue
Email: [email protected]
Class Meeting: Mon & Wed, 4:00 PM – 5:15 PM, Denny 109
How to Automate Solutions to Computational Problems

• Spam email classification:


• Binary classification of emails:
Spam vs. Ham (Legitimate message)

• Expert Systems approach (Rule-based)


• A group of experts write rules determining
whether an email is spam or not.
• A programmer implement the rules into
computer code

• Example rules:
• Classify the email as spam in “Money” appears in the text.
• What if the email is sent by your parents?
How to Automate Solutions to Computational Problems

• Cons of Expert Systems approach (Rule-based)


• Cognitively demanding: Difficult for humans to reason with
many useful but imprecise features that are indicative
(signals) of spam or not spam:
- Words, phrases, images, meta-data, time series, …
- Need to combine a large number of signals, figure out their relative
importance in determining spam vs. ham label.
• Brittle: Always going to miss some useful
features or patterns
- Spam filtering is adversarial, new features need to
be added over time.

Expert (Rule-based) Systems


Why Machine Learning?
• Machine Learning algorithms can automatically learn
the weights to combine features.

• A typical Machine Learning (ML) approach:


1. Acquire a large enough dataset of labeled examples:
• Each email is an instance, the label is spam (+1) vs. not spam (-1).
2. Represent emails as feature vectors:
• Each feature has a weight, the sign of the weighted sum of features should match
the label.
• Traditional ML: Engineer the features.
• Deep ML: Learn the features
3. Learn the weights so that the model (weighted combination of features)
does well on labeled examples.
What’s Machine Learning?
• Machine Learning is to construct computer programs
that learn from experience to perform well on a given
task.
• Supervised Learning: discover patterns from labeled
examples that enable predictions on (previously unseen)
unlabeled examples.
Topics of This Class
Machine Learning
• Function is everywhere!
• Function à 𝑓 ; Input instance à 𝑥; Output Targe à 𝑦

𝑓 à

𝑓 à
• Machine Learning Task:
• learn an (unknown) function 𝑓: 𝑋 à 𝑌 that maps input
instances x ∈ 𝑋 to output targets 𝑓 𝑥 ∈ 𝑌.
Linear Regression

• Given the floor size in square feet, predict


the selling price:
§ Input 𝑥: the floor size of the house
§ Output 𝑦: the selling price of the house
§ Need to learn a function ℎ such that ℎ 𝑥 ≈ 𝑓(𝑥)

• Assumption: the prediction


results are the linear
combination of input attributes
(features).
Logistic Regression

• Objective: To accurately determine the likelihood (probability) of


each data instance being classified into a designated category.

𝑥"
3 1

2
0.5
1

0
1 2 3 𝑥! -3 0 3
Support Vector Machine

• Objective: to find the optimal hyperplane that maximally separates


classes of data points in a feature space.
Neural Networks
• Assumption: complex patterns and relationships in data can be
approximated and learned through a series of interconnected layers
of simple computational units (neurons).
k-Nearest Neighbor

• Assumption: similar data points are likely to belong to the same


class, with proximity measured by a distance metric in the feature
space.
Decisions Trees

• Assumption: complex decisions can be modeled through a series of


simple, hierarchical decision rules based on the input features.
Generative Models for Classification
Clustering

• Objective: partition a dataset into k distinct clusters


without supervision.
Principal Component Analysis
• Objective: to reduce the dimensionality of a dataset by
transforming it into a new set of variables (principal components).
Autoencoders
• Objective: to learn a compressed, low-dimensional representation
of input data by encoding it into a latent space and then
reconstructing the original input.
Ensemble Methods

• Assumption: Combining multiple diverse models or algorithms can


yield more accurate and robust predictions than any single model
alone.
Netflix Prize 2007

• The Netflix Prize was an open competition for the best collaborative
filtering algorithm to predict user ratings for films, based on previous
ratings without any other information about the users or films.

• Rewarded $50,000 in 2007.

19
Reinforcement Learning
• Assumption: An agent can learn optimal behavior or policies by
interacting with an environment, receiving feedback in the form of
rewards or punishments for its actions.

Play Dota Chip Design

AlphaGo Nuclear Fusion AWS Deep Racer


Convolutional Neural Networks

• Assumption: Input data, particularly images, can be effectively


processed and analyzed using spatial hierarchies of patterns
through localized and parameter-efficient convolutions.
Graph Neural Networks
• Assumption: Data represented as graphs can be effectively
analyzed by capturing dependencies and relationships between
nodes through message passing or aggregation mechanisms.
Recurrent Neural Networks
• Assumption: Sequential data has an inherent order, and
understanding each element requires knowledge from its previous
elements in the sequence.
Attention & Transformers
• Assumption: It's beneficial to selectively focus on certain parts of
the input while processing others, thereby dynamically prioritizing
information based on its relevance.
Deep Generative Models

• Variational Autoencoders: • Generative Adversarial Networks:


Deep Generative Models
• Diffusion Models:
My Research:
Wireless Human Sensing
28
Research Overview
Human Activity Recognition 3D Skeleton/Mesh Reconstruction
[SenSys’22] Multi-subject 3D
Human Mesh Construction
Using Millimeter-wave
[UbiComp’20] Multi-View Deep Learning for Environment-independent
Device-Free Haman Activity Recognition [MobiSys’21] 3D Real-Time
Human Mesh Construction
Using Millimeter-wave

[MobiCom’20] 3D Human
Pose Construction Using
[MobiHoc’19] A Deep Learning Framework for the Fusion Commercial WiFi Devices
of Heterogeneous Sensory Data

Deep Learning

Commercial Wireless Devices in IoT Systems

Acoustic Wi-Fi mmWave

29
Why we need Wireless Sensing
Advantages of Wireless Sensing Systems:
• Easy to deploy using existing IoT
wireless devices for communication.
• Function well under the conditions of
bad lighting and occlusions.
• Have little privacy issue since the wireless
signal data is not human-readable.
• Exempt people from extra burdens and
discomfort caused by wearable devices. Ubiquity of Wireless Infrastructures

Bad Lighting Occlusions Privacy

Challenges of Vision-based Solutions Challenges of Using Wearable Devices

30
Applications
Human-Robot Interaction Smart Cities Rescuing

Smart Homes

Gaming/VR/AR Elderly Monitoring

31
Experiment on Wireless Dataset

§ Device-Free Human Activity Dataset

§ Heterogeneous Dataset

4 Wi-Fi Sensors 3 Acoustic Sensors

Wi-Fi Transmitter Acoustic Transmitter


Acoustic Transmitter
Wi-Fi Receiver Acoustic Receiver
§ Results

Acoustic

32
[MobiHoc’19] DeepFusion
WiFi-based Human Skeleton Reconstruction

§ YouTube Link:
§ https://www.youtube.com/watch?v=puU4EvBTPxA

33
[MobiCom’20] WiPose
mmWave-based Real-time Human Mesh Reconstruction

§ YouTube Link:
§ https://www.youtube.com/watch?v=xW4ZlmxEu9Q

34
[MobiSys’21] mmMesh
Questions?

You might also like