hs1501

what is ai

refers to machines performing tasks that ar commonly associated w intelligent beings (reasoning, discovering meaning, generalizing/learning from exp)

geoffrey hinton

professor @ university of toronto. one of the researchers who introduced back-propagation algorithm for training deep neural networks

won 2012 imagenet competition w their AlexNet for recognising objects in images w his phd students (alex and ilya)

yann lecun

prof @ new york uni

joined fb (now meta) in 2013 as director of ai research

known for his work on optical character recognition using CNN (convolutional neural networks)

deepmind

british startup → acquired by google in 2014

developed alpha go (beat 9-dan professional in 2016)

developed a program that learnt to play 49 diff atari games from scratch by seeing puzzles and score on the screen in 2013

andrew ng

chairman and co-founder of coursera. confounded google brain in 2012

machine learning

the practice of using algorithms to parse data, learn from it, then make a determination/prediction abt smt in the world

model

an algorithm in machine learning that is supposed to accomplish a particular task

contains many parameters - only a right choice wld allow model to produce desired outputs

parameters not hand-coded, but rather found from a large set of data

training (dvper pov) / learning (model pov)

model finding a right set of parameters algorithimically

3 types of basic machine learning paradigms

1. supervised learning
2. unsupervised learning
3. reinforcement learning

supervised learning

training data consists of labelled examples (each data point possesses an associated label)

unsupervised learning

learns patterns from untagged data

reinforcement learning

model learns desired behaviour from rewards/punishments it gets from its predictions through trial n error

neural network consists of

1. an input layer
2. one or more hidden layers
3. an output layer

a layer in a neural network is composed of

nodes (aka neurons)

neurons stores:

a number (except for neurons in input layer) — calculated using a weighted sum of the numbers stored in the neurons in the prev layers

activation function

eg: rectified linear unit (RELU), which only allows +ve numbers to go through

\
is used to determine whether weighted sum stored in neurons goes through to the next layer

bias

a fixed number that is added to the weighted sum to control the activation treshold

parameters of the model

weights and biases

a right set of weights and biases for each neuron is required for the neural network to give desired outputs

a right set of weights and biases can be obtained by:

repeatedly running labelled data through the neural network

by comparing outputs and labels → revises the weights to decrease the error (eg using a method called Gradient Descent w back-propagation) → trains the neural network

deep neural network

neural network that has multiple hidden layers

deep learning

a class of machine learning algorithms that uses deep neural networks

eg image processing

convolutional neural networks (cnns)

neural networks w special layers called convolutional layers (each weighted sum is calculated over only a small subset of neurons that are ‘adjacent’ to one another)

successfully in analysing images

recurrent neural networks (rnns)

neural networks in which connections bw neurons may form a cycle

capable of processing input of variable length

useful in natural language processing and speech recognition

long short-term memory (lstm)

specially designed RNNs that can learn longer-term dependencies

eg translation ai

hype cycle

coined by Gartner (a company)


1. technology trigger
2. peak of inflated expectations
3. trough of disillusionment
4. slope of enlightenment


1. plateau of productivity

what can ai do for businesses (PNS APPS)

1. perception
2. notification
3. suggestion
4. automation
5. prediction
6. prevention


1. situational awareness

ai’s impact on jobs

47% of total US employment in high risk category (ie potentially automatable by 2033)

75-375mil workers affected, will need to switch occupational categories

occupations that involve **complex perception n manipulation tasks, creative intelligence tasks, social intelligence tasks, and high-lvl cognitive capabilities** are unlikely to be automated soon (low-risk jobs are higher-order/creative)

busines transformations driven by ai

1. save money


1. replace mundane manual tasks
2. streamline processes
3. reduce cost of errors
4. improve product reliability
5. transform service concepts
6. using robots
2. make money


1. providing markets insights
2. powering product breakthroughs and innovations strategically
3. allowing consumer products to be personalised on a large scale
3. meet compliance and gov requirements


1. cybersecurity
2. privacy legislation
3. risk management

ai applications in retail industry

* facial recognition, speech and text analytics — analyse sentiment, emotion, tone, context behind customer behaviour → identify customer patterns
* create personalised product reccs
* deep learning algorithms used to improve search n label images
* chat bots: communicate w customers
* predict customer behaviour
* The Magic Mirror - using Augmented Reality (AR) tech → virtual dressing room

ai applications in banking industry

* facial recognition — track people
* automate services and tasks
* analuyze transactions and detect fraud

DBS bank case study

* piyush gupta (ceo of dbs since 2009) — led digitization of bank
* 2014: announced they planned to apply Watson (ai computer system developed by IBM) to their wealth management business to improve advice and experience delivered to rich customers
* launched digibank in india and indonesia in apr 2016 and aug 2017 respectively
* aug 2017: announced it was investing 20e il over 5 years in a program to skill 10000 sg employees in digital banking and i merging tech
* jun 2018: employed use of ai chat bot in interview process

smart nation case study

* since nov 2014
* lifesg app connects to more than 40 gov services
* singpass (national digital identity initiative) — provides convi n secure platform for citizens n businesses to transact w gov and other priv service providers
* vid analytics to automate analysis of police cam footage — police can respond faster
* smart nation sensor platform uses sensors to collect essential data that can be analysed to create smart solutions

FaceNet

an ai system developed by google in 2015, capable of verifying, recognising n comparing faces

facial recognition

since covid, researchers developed ai systems that recognise ppl even w masks on

can be used to identify ppl acting suspiciously and ppl on blacklists

1;1 verification and 1;n identification (finding identity of someone)

can be deployed in cashierless stores

ai chat bots and voice n text communication

eg virtual voice agents (google assistant, siri, alexa)

* they work seamlessly w smart speakers
* can control lights and devices, play music n vids, get answers, place orders
* but, still not v good at voice recognition esp when diff languages and dialects are involved. info avail is also limited

how to implement ai

1. choose/invent model
2. program model
3. train and run model

training ai models requires…

* massive amts of computation
* can take a shorter time if u use GPUs (graphic processing units)
* now, can train ai models using gpus on the CLOUD w/o having to physically buy and maintain gpus (save cost)
* \

programming ai models through…

1. tensorflow (developed by google brain)
2. pytorch (developed by meta ai / aka fb ai)

tensorflow vs pytorch

similarities


1. both open-source and allow computation on one or more cpus and gpus
2. both can be used w python (the top programming lang for ai)
3. both allow ai models programmed in them to run on multiple devices (desktop, phones, cloud, web)

differences


1. pytorch is easier to use
2. tensorflow has interfaces (Keras and AutoKeras) to make the implementation more user-friendly

\
with these software libraries, simple ai models can be programmed w less than 10 lines of code

ai walking case study

* google’s DeepMind used reinforcement learning to teach a computer to walk and navigate complex env

ai use in traffic

* DataSpark (SingTel) uses data from mobile phone networks to track people mvmt, to:
* coordinate traffic lights in a Smart Nation
* show ppl which direction to move in crowded areas n big events — maintain safety
* optimise services
* plan evacuation due to emergencies
* understand foot traffic in shopping malls…??

ai use in cost and energy saving

* Fuel Dashboard (Boeing) can track >600 parameters derived from operational data for flight operators to identify potential efficiency problems.
* also allows comparison bw actual performance n plan to optimise flight plans
* ai algorithms learning to predict delays → airports and airlines have better chance to avoid delays
* real-time updates on traffic and calculations on eta → produces an optimal speed → flights can avoid wasteful burning of fuel as they wait their turn to land
* Verdigris (a company) combines sensors and AI to identify motor problems using excess energy and to verify energy efficiency upgrades
* DeepMind (Google) applied machine learning to reduce amt of energy Google’s data centres use for cooling by up to 40% in 2016

ai use in weather monitoring

* Helios platform (L3Harris) delivers real-time insights abt hyperlocal weather that trad weather sources typically struggle to identify
* through its large terrestial camera network, Helios uses AI to detect occurrence and impacts of weather in specific locations on critical ground infra (eg roads)
* this fast n accurate local ground weather intelligence can be used to support weather forecasting, emergency response, vehicle safety and other weather-dependent decision making

nlp stands for…

natural language processing

what is nlp

the study of techniques that enables computers to use (both written and spoken) human lang the ways human beings can

nlp capabilities

* topic modelling (uncovering hidden topics from large collections of documents)
* text categorization (sorting text) & text clustering (grouping text based on similarities in context)
* info extraction (finding meaningful info in unstructured text)
* named entity recognition (NER) (extracting names of ppl/places/companies etc, and classsify them into predefined labels)
* entity resolution (identifying records in data sources that refer tot he same real-world entities and identifying rships bw these records)
* sentiment analysis (detecting polarity \[eg pos or neg\], emotion and intention \[interested/not interested\] in text/speech)
* summarization
* translation
* speech recognition (speech → text) and speech synthesis (text → speech)
* natural lang generation (NLG) (transforming structured data into human lang)

info extraction

* finding meaningful info in unstructured text
* ai nlp allows extracxtion of useful info from big data that are too large to be analyzedc using trad methods
* eg ai can analyze patent data to visualize rships bw patens → help investors make more informed decisions
* question-answering ai: can generate answers to questions by querying a knowledge base
* closed-domain question answering deals only w questions under a specific domain (eg med and law) vs open-domain question answering deals w factual questions abt nearly everything

NER and entity resolution

* NER = named entity recognition
* NER useful in entity resolution
* financial org and public-sector org can use entity resoltion to detect fraud, improve risk-assessment, improve investigative outcomes help ensure compliance, improve customer insights, and reduce false positives n negatives
* Senzing (a company) dvped a software using AI that is capable of performing real-time entity resolution

\

sentiment analysis

automatically analyzing customer feedback (eg opinions in survey responses n socmed convos)

using sentiment analysis allows brands to better understand their customers → can tailor products and services to meet their needs

summarization

* EXTRACTIVE models: “copy n paste” operations -select relevant phrases of input doc and form a summary
* more robust bc they use existing phrases, lack flexibility since cannot use new words / connectors. also cannot paraphrase
* ABSTRACTIVE models: create summary based on “abstracted” content - can use words not in original input
* moer potential to produce fluent and coherent summaries

NLG

* natural lang generation
* nlg is challenging bc of the complexity, ambiguity and variety of expressions in human languages
* nlg uses: writing reccomendation, suggesting replies to emails n complaints, collating audit findings
* recent model: gpt-2 — can adapt to the style n content of prompt, allowing user to generate realistic and coherent continuations abt a topic of their choosing

speech synthesis

* WaveNet model (created by Google’s DeepMind in 2016) can generate realistic-sounding human-like voices that were better than what google had from its other speech synthesis systems
* WaveNet can also be used to synthesis other audio signals such as musicn
* \

translation

* classical approach to machine translation is RULE-BASED (based on dictionaries n grammars) → requires a great deal of manual effort
* another method: STATISICS-BASED (picks out the most likely translation according to sample data given)
* 2016: google translate started using RNN (recurrent neural network) models like LSTM (long short term memory) which give superior performance compared to statistics-based models
* Their rnn models make use of ATTENTION MECHANISM that allows algo to focus on diff regions of the input during course of translation
* Transformer model is becoming more popular in machine translation

Transformers

* popular in machine translation
* 2017: google brain introduced the transformer language model
* model is based solely on attention mechanisms, w/o recurrence
* can outperform recurrent and convolutional models
* appliesn a self-attention mechanism → directly picks up rship bw all words in a sentence and choose the best meaning
* neural networks typically contain an encoder reading the input sentence sentence and generating a representation of it. a decoder then generates the output sentence word by word while consulting representation generated by encoder
* Transformer starts by generating initial representations, or embedding for each word. these r represented by unfilled circles. then, using self-attention, it aggregates info from all the other words → generating new representation per word informed by the entire context → this step is repeated multiple times for all words → suggests fully generating new represenations
* decoder works similarly, but generates one word at a time . attends to both previously generated words and final representations generated by encoder
* design of Transformers allows dvpt of pretrained models (google;s BERT and OpenAI’s GPT), that can be fine-tuned w smaller datasets for more specific tasks
* fine tuning of general-purpose models → purpose-specific models = TRANSFER LEARNING
* training of BERT involves MASKING some words randomly n asking model to predict masked word based only on context

chat bots

* an AI-powered NLP tech
* softwares used to conduct written/spoken convos in natural lang
* music chat bots exist too (AI Duet)
* some chat bots can now be developed without code
* while ai can now retrieve info from internet/database to answer open-domain questions, many chat bots still have rather limited scopes

generating images from text

nlp techs allow ai to generate high-quality images from text descriptions

current challenges in nlp

1. multilingual speech/text
2. ASEAN lang lack corpus
3. ai has no true understanding of lang - poses limites to question answering abilities. ai cannot understand questions w complex structures! and can’t find ans to ambiguous questions
4. sentiment analysis affected by many parts of text
5. transformers is a major breakthrough but take lots of computer power to train and are slow
6. many nlp systems require additional training and cannot work outs of the box (they r pretrained models)

digitization of nlp ai

* makes nlp tech more easily n widely accessible
* smart nation initiative — many official procedures becoming paperless
* IMDA implemented nationwide E-invoicing network (2019) to help business improve efficiency, reduce cost, enjoy faster payment and stay green
* MAS driving e-payments (SGQR) , including in hawkers
* singpass, the national digital identity initiative, provides a convi n secure platform on personal computers n handphones for digital signatures and biometric authentication
* digitization has the need to maintain records securely wo central ledger → use of BLOCKCHAIN (a piece of tech that achieves this)
* w.o involving central authority in a so-called self sovereign identity system
* can be used to authenticate n authorise users and devices over a computer network
* certify authenticity and maintain ownership records via NFTs
* instructions to perform certain actions can b inserted into blockchain such that they r automatically executed when triggering event happens = SMART CONTRACTS

nlp ai in healthcare

* ai can detect n diagnose diseases from medical exam results and patient record
* helps detect diseases earlier n make advanced medical care cheaper n more widely avail

nlp ai in finance industry

* auto track n extract useful info from documents etc
* can make predictions n reçus based on info
* track n extract info form internet etc to
* continuously assess risks associated w company
* suggest what a fair value is for a merger/acquisition based on patents etc
* make predictions for future performance, price/market trends
* fully automated investments
* track transactions to detect fraud
* automate workflows, customer relations
* build reputation and engage customers n stakeholders via socmed, chatpots, apps
* challenges faced
* provide more personalised n user0frieldly financial advice
* provide natural lang customer interface w 3d displayed
* improve cybersecurity
* ensure compliance w new legislation n industry standards

nlp ai and hr industry

* screen candidates using chat bots
* predict who will stay n who will job hop
* predict job performance
* match applicants …
* \

industrial rvolution

1st: steam engine powering mechanical production facilities

2nd: division of labour, mass pdct w help of electricity

3rd: electronics and it → automates pdct

4th: cyber-physical systems (fusion of phy, digital and cloud tech)

smart factories

ie: a digitization manufacturing facility (enabled by ai, cloud computing and internet of things) that operates using connected devices, machinery and pdct systems that continuously collect n share data

\
key principles of a smart factory: fewer shutdowns, improved planning and pdct processes, optimised facilities

\
integration of ai and automation in a smart manufacturing facility is realised through:


1. real time predictive maintenance
2. time-series forecasting
3. factory monitoring

\
eg. introduction of machine learning to manufacturing process allows possible scheduling of preventive n predictive maintenance based on accurate real-life info to avoid pdct line shutdowns

\

amazon go

Just Walk Out (checkout free shopping exp)

powered by the tech used in self-driving cars (computer vision, sensor fusion and deep learning)

\

AiOps

artificial intelligence for IT operations

a multilayered approach to manage complex IT operations that optimisés service availability n delivery.

uses machine learning, predictive analysis and ai to automate, enhance n improve it operations

meta learning ai

learning algo that learns from other learning algo (learning to learn)

able to learn w a few examples

meta learning in machine learning most commonly refers to learning algo that learns from OUTPUT of other machine learning algo

learning algo

learn from Historical data n make predictions given New examples of data

meta learning algo

learn from OUTPUT of learning algo n make a prediction given Predictions

ai predictive maintenance

ability to leverage large volumes of real0time asset data → anticipate n address potential issues before they lead to breakdowns in operations

\
from data, ai systems can


1. detect abnormally
2. statistically analysis
3. embedded intelligent edge processing

\
limitations:


1. insufficient/wrong type of sensors
2. analysis must b done at machine edge
3. manufacturers don’t share the same vision n prefer status quo

ai vision capabilities

1. object n facial recognition
2. image/video captioning
3. pose n action recognition
4. image processing

uses of ai vision

1. healthcare sector (detect diseases/colour-tagged cancerous cells)
2. security (detect smuggled weapons, drugs, intruders)
3. workplace safety (detect faulty machinery parts or violations to workplace safety protocols)
4. satellite imagery (provide activity signals, analysis of soil)

YOLO

You Only Look Once: unified, real0time object detection

a deep learning model for detecting the position n type of object from the input image

can classify objects in one of the 80 categories avail and compute bounding boxes for those objects from a single input image

one-step process for object detection, works on a neural network model that requires just one pass of an image

\
several advantages: can be used within real0time applications as the object detection systems embedded in autonomous vehicle systems

facial recognition

an advanced form of biometric authentication capable of identifying and verifying a person

uses ai algo and machine learning (ML) to detect human faces from bg. starts by searching for human eyes → eyebrows → nose etc

\
applications:

* identify celebs in their coverage of sig events
* providing secondary authentication for biometrics n apps
* auto indexing image n vid files for media n ent companies
* security for events
* identify offenders
* rescue victims

\

image/vid captioning

process of generating a textual description of what is happening from input image/vid

action recognition

model analyses input vid n returns. a list of categorised actions

enhancement of Machine Vision capable to handle sequences of actions n mvmt

networks must look at several consecutive frames

\
use cases:


1. automate testing
2. identify offenders
3. safety enforcement n supervision
4. assist in behavioural analysis n rehab

image processing

ai can enhance resolution of an image n add colours to black0and white images

GPUs

for neural networks as they require computation to be done in parallel

\

CPUs

for sequential calculations like sims\`

slower than gpus

tensor

array of numbers

register

a location on circuit to store info as memory

SIMD

single instruction multiple data

singe instruction burt requires many numbers to be stored during computation

floating point

numbers w decimals

\

FLOPS

floating point operations (calculations using floating point numbers)

\

GEMM

generate matrix manipulation

algo for matrix calculations

TPUs

tensor processing units

similar to GPUs but optimised for AI calculations

FGPA

field programmable gate arrays — integrated circuit that can be bought. allows developers to reconfigure and are “programmable” in the field

\

DSP

digital signal processors

used for processing analogue signals (recorded audio)

implementing with AI algorithms is sensible

applications:


1. alw on wake up: devices permanently on, listening for keywords (hey siri). ai enables speech recognition for vocal commands
2. contextual awareness - devices can recognise if they r in a quiet or noisy env and dynamically adjust their speaker volume, similar to how screen brightness dynamically adjusts to external lighting
3. sentiment analysis: recognising if someone is laughing/shouting nearby

\
properties of ai run on dsp


1. low energy: can be used in mobile devices w/o draining battery
2. ai lagos running on de vice itself = better data privacy bc data not transmitted to cloud for processing
3. lower cost bc no need to communicaten w cloud services for ai

cloud services

smaller scale, more affordable for small n medium enterprises

creates new business models n allows small companies to access powerful ai capabilities to drive their products

\
characteristics


1. elastic
2. on-demand
3. multi-tenancy (can be shared)
4. metered service(pay as u use)

\
implications


1. anonymity of geography n provider
2. transient customer provider rship
3. new cybersecurity challenges