Artificial Intelligence (AI): tries to design computer models that accomplish the same cognitive tasks that humans do
Machine Learning: a subset of AI that allows computers to “learn” (i.e., progressively improve performance on a specific task) by creating new algorithms to produce a desired output based on structured (or unstructured) data that is provided
Deep Learning: a subset of Machine Learning involving numerous layers of algorithms
Computer does not need to be provided with structured data
Neural Networks: Networks of algorithms that are similar to the neural networks present in the human brain
There are a number of different artificial intelligence formulations
Strong AI: General purpose AI
Machines that possess artificial general intelligence (AGI)
Would be just as smart as humans across the board with the ability to understand and learn any task that a human can
Applied AI: AGI isn’t going to be created any time soon, but machine learning has made it possible for machines to learn how to master complex tasks (expert systems), including
Playing the ancient Chinese board game Go
Identifying human faces
Translating text into practically every language
Driving cars
Computer replication: Understand how the mind works and replicate its functions in machine or organic form
Problem: How can a computer learn to distinguish between pictures of dogs and cats?
Approach: You label pictures of dogs and cats with specific defining characteristics (e.g., length of ear, color of nose), then feed this structured data through the computer
You feed the computer pictures labeled as dogs or cats without additional structured data
Data is sent through different layers of the artificial neural network corresponding to different layers of abstraction, from low level (e.g., does this part of picture contain a brown spot?) to more complex, e.g., (is there a nose in this part of the picture?)
Supervised, Unsupervised, and Deep Reinforcement Learning
#
Supervised learning: network receives explicit feedback on how successful it is
Unsupervised learning: network does not receive explicit feedback; instead it learns to detect patterns in data
Reinforcement learning: distinct from both of the above in that…
It does depend upon a feedback signal
However, the feedback signal does not tell the network exactly what it has done wrong; instead, the network is driven by a reward signal
The job of the network is to maximize the reward, but it is not told how to do that
It has to work out for itself which outputs are most profitable
In 2016 and 2017, AlphaGo program created by Google’s Deep Mind research group beat the world’s leading human experts at the game of Go
These victories were widely recognized as historic achievements for AI
Go is an abstract strategy board game for two players, in which the aim is to surround more territory than the opponent
It is one of world’s most complex games
Chess has $10^{123}$ possible moves; Go has $10^{360}$
AlphaGo used a mixture of supervised learning and reinforcement learning
Was initially trained on a database of 30 million moves from an online server using supervised learning – received explicit feedback on how successful it was
Once AlphaGo had achieved a relatively high level of playing strength, training shifted to reinforcement learning
In reinforcement learning, network will not increase its reward simply by repeating what has worked in the past
It needs to engage in trial and error to discover new reward-generating strategies
Reinforcement learning was achieved by getting the network to play games of Go against former versions of itself
New version, AlphaGo Zero, incorporated zero supervised learning
After three days (and 4.9 million games played against itself), it was able to beat the version of AlphaGo that had defeated the leading human expert in 2016 – 100 games to 0!
Within 40 days, it was able to beat all existing versions of AlphaGo
Computers found to be as good or better than doctors at detecting tiny lung cancers on CT scans
Potential hazard: A radiologist who misreads a scan may harm one patient, but a flawed A.I. system in widespread use could injure many
Determining optimal treatment, including
Type and dosage of drugs
Best diet for individual, e.g., to avoid glucose spikes after eating
This may vary, depending on patient’s unique gut microbiome
Precision surgery without human artifacts like handshaking
Study comparing computer-controlled robots with human surgeons in performing intestinal surgery on a pig found that the robot sutures were much better—more precise and uniform with fewer chances for breakage, leakage, and infection (Shademan, Decker, Opfermann et al., 2016)
Computer chip is implanted in motor cortex and communicates directly with external device
Allows animal (or person) to directly control a robotic arm with their thoughts
A cap with electrodes can now be used instead of implants, but an extensive calibration process is required
Direct brain-to-brain communication in humans is also now possible
Two research participants are are positioned in two different buildings on campus
The sender, left, thinks about firing a cannon at various points throughout a computer game
That signal is sent over the Web directly to the brain of the receiver, right, whose hand hits a touchpad to fire the cannon (Rao, Stucco, Ryan et al., 2014)
Diagnosis/identification of psychological disorders
AI system that analyzed Facebook posts of consenting patients in an emergency department was able to generate predictions of depression risk that were as accurate as standard depression screening tests (Reece & Danforth, 2017)
Indicators included references to sadness, loneliness, hostility, rumination, and increased self-reference, e.g., words like “alone,” “ugh,” “tears,” and higher frequency of use of “I” and “me”
Length and timing of posts were also considered (Eichstaedt, Smith, Merchant et al., 2018)
AI has also been used to analyze Instagram photos to successfully screen for depression
Photos posted by depressed individuals tended to be bluer, darker, and grayer
The more comments Instagram posts received, the more likely they were posted by depressed participants, but the opposite was true for likes received
Depressed participants were more likely to post photos with faces, but had a lower average face count per photo than healthy participants
The screening models created from the data were able to outperform general practitioners in correctly diagnosing depression without the assistance of assessment instruments
A.I.-driven voice analysis
Researchers are currently working on developing voice analysis programs that can help identify psychological disorders
In depression, speech is generally more monotone, with reduced pitch range and lower volume; there may be more pauses
In anxiety, speech tends to be faster, and there may be evidence of difficulty breathing
Programs may also be used to predict other mental illnesses like schizophrenia and PTSD
Use of deep-learning algorithms can uncover patterns that might not be evident even to trained experts
“The technology that we’re using now can extract features that that even the human ear can’t pick up on.”
Potential problems:
It can be difficult to know why your vocal levels fluctuate, e.g., you may just be trying to speak quietly
Issue of bias: need to ensure that programs work for all patients, regardless of age, gender, ethnicity, etc.
Deep learning algorithms work in ways that even the developers themselves can’t fully explain, that is, knowing which features are being used to make the predictions
Apps that can administer Cognitive Behavioral Therapy for disorders like depression or social anxiety, e.g., Woebot
Virtual therapists, e.g., Ellie
These programs may analyze tone of voice, breathing pattern, smartphone keystrokes and communication, and/or physical movements in making diagnoses and generating responses
Pros:
Easy accessibility and affordability
Research has indicated that people would rather share their innermost secrets with an avatar than a human being
2018 was the target date proposed in 2015 both by Elon Musk of Tesla and by Google for introduction of self-driving cars
But a series of widely publicized crashes, some fatal, have delayed release
Are self-driving cars really a viable option?
The great successes of deep learning have all been in relatively circumscribed domains, including chess and Go, and even things like image recognition, which primarily involves identifying patterns in a data set, then projecting those patterns onto new exemplars
Some believe that self-driving cars (and other autonomous vehicles, like drones) need more than sensitivity to patterns and the ability to learn from experience
They need to be able to deal with the unexpected – completely unpredictable behavior from other drivers, pedestrians, cyclist, and even wild animals
Also, human drivers are constantly exploiting their knowledge of how physical objects move and behave (folk physics), as well as their knowledge of other drivers and road-users (mindreading)
Perhaps a key challenge for designers of self-driving cars is how to equip their vehicles with this kind of general knowledge
Deliberate programming of AI to be hostile, e.g., by terrorist group
Sci-fi scenario: If human behavior contradicts one of AI’s preprogrammed goals, AI could turn malicious
In 2017, two Facebook computers started communicating with each other in a language they had developed on their own
Simulation of government leader’s image and voice issuing unauthorized orders, e.g., military action
Generation of fake emails, phone calls, video chats
Loss of privacy
Loss of jobs
Some AI experts predict that AI will replace or eliminate 40% of jobs within 15 years
Greatest impact will be on jobs involving tasks that are repetitive and can be automated
Other AI-related risk: Development of computer technology has meant an increase in the amount of time we spend looking at computer screens, and that may be harmful
Recent study published in JAMA Pediatrics found that after controlling for age, gender and income, 3-5 year old children with higher use of screen-based media
Had lower measures of structural integrity and myelination
Scored lower on cognitive tests (Hutton, Dudley, Horowitz-Kraus et al., 2020)
Similarly, in older adults, increased television viewing was found to be correlated with cognitive impairment and poor verbal memory (Fancourt and Steptoe, 2019)
In fruit flies, daily blue-light exposure (such as is used in computer screens) causes brain neurodegeneration, as well as shortening of lifespan (Nash, Chow, Law et al., 2019)
American Academy of Pediatrics recommendations:
Children under 18 months should avert their eyes from TV and screen media at all times
For children 2 to 5, screen time should be limited to no more than 1 hour per day
However, three years later, Google still had not really fixed anything – the company simply blocked its image recognition algorithms from identifying gorillas altogether
Baboons, gibbons, and marmosets were all correctly identified, but gorillas and chimpanzees were not
Try typing things like “why Black/Latinx/Asian women/men/boys/girls …” on Yahoo or Bing
You’ll get really biased, racist and discriminatory auto-suggestions
These are now blocked on Google, but not on other search engines
