Category Archives: Week 4

Can a Neural Network Think?

Discussion Notes – Tianyi Zhao and Adey Zegeye

Case Study: What a Deep Neural Network thinks about your #selfie – Andrej Karpathy


ConvNet Training in Karpathy


  • Data collection: defining a quick script to gather images tagged with #selfie. (5 million)
  • Convolutional networks trained to pick images with at least one face. (2 million)

To decide if the selfie is good or bad: 

  • ranked the users by the number of followers
  • Divided into groups of 100, and sorted by the number of likes
  • Top 50 = positive, bottom 50 = negative
  • Train a ConvNet with the binary split.
  • The ConvNet can be well customized based on different demands of the trainers.
  • The database is so large and various. There should be a more specific explanation about the statistical preferences or a detailed discussion.
  • The patterns should be strictly selected and apply to any selfie
  • The accuracy of ConvNet depends on:
    • How well and precise the trainer defines patterns
    • Different neural network architectures (Caffe, VGGNet, ImageNet, etc.)
    • ConvNet in NLP, in machine translation for example:

According to Ethem Alpaydin, the neural machine translation ends the era of phrase-based statistical translation, because it translates an entire sentence at a time rather than cutting it into words. Recurrent neural networks (RNNs) are prevalent in this field. However, ConvNet has gradually replaced RNNs in language translation. On the one hand, ConvNet can make computation fully parallelized in GPU. ConvNet computes all elements simultaneously, while RNNs operates in a strict left-to-righting or right-to-left order (one word at a time) in which each word must wait until the network finishes the previous one. On the other hand, ConvNet processes information hierarchically, making it “easier to capture complex relationships in the data.” (Gehring & Auli, 2017) 

Case Analysis : Key Points and Issues 

Although ConvNets can be very useful in picking up pattern in large amounts of data, one of the main issues is that they don’t tell the whole picture. They can only select data based off of set parameters and rules – which does not translate into human problem solving or decision making.

  • A ConvNet is a large collection of filters that are applied on top of each other
  • They will learn to recognize the labels that we give them
  • Context is important when comparing what a neural network can do vs. a human brain
  • In the Karpathy article, the 1/3 rule problem resulted in the network choosing a “logical” but not accurate for the purposes of human perception (the way a human would determine whether the selfie is good or not)
  • “What we lack in knowledge we make up for in data” – Alpaydin
    • Still includes limitations: the problem of binary (two-value logic)
    • It is not always true/false , yes/no, we think and speak in more complex patterns
      • In some of the selfies considered “good” by the convnet, the entire person is cut out of the image. The network does not have the ability to use context in the way humans do, to know or understand that selfies can be taken in different “moods” or patterns recognized in human language / images
    • “we live and think with multi-valued(not true or false, but “all depends…”) reasoning and multimodal(different kinds of statements: hypothetical, imaginary, contrary to fact, ironic, etc.)” – Irvine, 2019
      • Cropped Selfies
    • Interactive machines can replicate “intelligence” by copying and learning / adjusting but it is not in itself “intelligent”

Another main issue is inaccurate language used to describe what a neural network CAN do. Neural networks don’t think, so the neural network doesn’t “think” your selfie is good or bad – it simply uses the information it is given within a set of parameters to decide if the image is good or bad (using two-value logic). This language proves confusing without the background readings that explain the process behind how a neural network works in comparison to

  • The parameters set are very important and heavily influence accuracy
    • This can lead to discrimination and ethical issues, who is deciding what features a classifier should be trained to extract?


Andrej Karpathy, “What a Deep Neural Network Thinks About Your #selfie,” Andrej Karpathy Blog (blog), October 25, 2015,

Ethem Alpaydin, Machine Learning: The New AI. Cambridge, MA: The MIT Press, 2016.

Gehring, Jonas and Auli, Michael. “A novel approach to neural machine translation. ” Facebook AI Research. May 9, 2017.

Geoff Dougherty, Pattern Recognition and Classification: An Introduction (New York: Springer, 2012). Excerpt: Chaps. 1-2.

Martin Irvine, Introduction: Topics and Key Concepts of the Course (Presentation), 2019.

Peter Wegner, “Why Interaction Is More Powerful Than Algorithms.” Communications of the ACM 40, no. 5 (May 1, 1997): 80–91.


Categories and Complexities: Machine Learning parameters

Proma Huq & Deborah Oliveros

What is a “good” selfie? What’s a “bad” one? – Under the lens of machine learning in terms of AI, the parameters for determining a good selfie from a bad selfie are debatable. As with any labeling of sets of data, Karpathy is aware of how skewed that categorization might be. What was interesting to us was how both sets of data were very similar to each other. When the author asked, “if I gave you any of these images, could you tell what category it belongs to?” I would not be able to do that in most of the cases (based on his pre-existing concepts of a good or bad selfie). While this may be subjective for the naked human eye, or tainted by biases, for the machine, it’s easier to differentiate them by means of creating a graphical model, since it recognizes the pictures while connecting all the given information to determine category, based on pre-existing parameters.

The following are some observations we had based on Karpathy’s article, in the context of the concepts from Alpaydin’s chapters:

  • Depending on what kind of data we provide and how we categorize that data, the machine’s responses and categories may be positive or negative in a societal context. To reiterate and simplify machine learning: we tell the machine what is “good” and what is “bad” and then the machine learns to differentiate based upon those parameters and subsequent learned patterns.

Case study example: Snapchat & Face Recognition

As we learned from Alpaydin, in the case of Snapchat (and Instagram, but Snapchat did it first!) face recognition for augmented reality filters, the “input” would be the image captured in the selfie, and the “category” or class to be recognized are the faces in the image in order to apply the filters to change them. The learning program therefore needed to learn to match the face image to the filter points in order to transform them to whatever augmented reality filter the user chooses that day.

There are several challenges with this seemingly simple application: Faces are 3D, often people do little videos of them with the filter, not just a photo. Other challenging factors can include accessories (hats, sunglasses), hair styles, facial hair, smiling or frowning, difference in angles and/or lighting etc. So how does the machine navigate all of this?

Facial recognition, in tandem with an “active shape model” learn to then categorize your face and apply the filter based on the “landmarks” of the terrain of your face.

ML and Ethical Issues

When it comes to examples of facial recognition and categorization, the concept of machine learning fairness comes into play. Some examples that come to mind are facial recognition systems mis-gendering African Americans, flagging Asian people in pictures as blinking, or most recently the case of police in the UK using facial recognition on one of the busiest streets in London over the holidays, to detect suspicious behavior and facial expressions, despite the fact that the data provided from criminal records have a long documented history of discrimination and targeting minorities.

–       Interesting fact: the top 100 selfies were all women, but none of them were noticeable women of color, which tells us (based on Karpathy’s grading systems of views and likes) that his first assessment of how to take a good selfie as “be female” should possibly be amended to “be a white female”. I doubt there’s some data determining white women take more selfies or their selfies are liked more than those of women of color. However, based on a very flawed categorization, that’s what the machine picked up on. It’s almost funny when he says, “I was quite concerned for a moment there that my fancy 140-million ConvNet would turn out to be a simple amount-of-skin-texture-counter” because he was worried it would detect the best selfies as the ones showing more skin. However, he fails to address that it resulted in a ‘amount-of-skin-shade-counter’ considering white as the best. Also, how about that cropping suggestion that leaves the girl out of the picture with the cars and then the clearly non-white guy out of the picture with Marilyn Monroe? This illustrates my previous point.

By design, these systems theoretically designed to be objective. That is because of how these convolutions or filters react to different ‘stimulations’ that are so minute they don’t necessarily have to rely on environmental stimulus it to determine category. Conversely, when it all comes together, that objectivity is lost due to human interaction, the categorization of that stimulus into labels. That’s where the ethical questioning has a crucial role because we need categorization, we just need to be aware of how representative of reality that categorization is and what is the possible outcome and impact in decision-making as a result of that categorization in terms of machine learning.

Works Cited

Andrej Karpathy, “What a Deep Neural Network Thinks About Your #selfie,” Andrej Karpathy Blog (blog), October 25, 2015,

Ethem Alpaydin, Machine Learning: The New AI. Cambridge, MA: The MIT Press, 2016.

Aubrey, A. “How Do Snapchat Filters Work?” Dev.To. 2018.

Is Supervised Machine Learning Standardizing Our Selfies?

Group Name: The Georgetown AI 3

Group Members: Zach Omer, Beiyuan Gu, Annaliese Blank

Alpaydin – Machine Learning Notes – Pattern Recognition and Neural Networks

Chapter 3:

  • Captcha corrupted image of words or numbers that need to be typed to prove that the user is a human and not a computer (pg. 58).
  • Semi-parametric estimation model that maps the input to the output but is valid only locally, and for different type of inputs use different models (pg. 59).
  • Localizing data in order to increase complexity
  • Its best to use a simple method, similar inputs have similar outputs

Generative Models

  • These represent how our beliefs can be reflected through or based off the data we generate (pg. 60).
  • Character recognition identity and appearance
  • Invariance size does not affect the identity
  • The generative model is CAUSAL and explains how the data is generated by hidden factors that cause it (pg. 62).
  • Facial recognition
  • Affective computing adapts to the the mood of the user
  • Biometrics recognition of people with their physiological and behavioral characteristics


  • Inputs are used for decision making (pg.73)
  • Dimensionality reduction learning algorithms, both the complexity of the model and the training algorithm depends on the number of input attributes.
  • Time complexity – how much calculation to do?
  • Space complexity – how much memory we need
  • Decreasing the number of inputs always decreases the time and space, but how much they decrease depends on the particular model and learning algorithm
  • Smaller models are based on small data, which is trained to with fewer data
  • Archive dimensionality can be done in two ways: feature selection and feature extraction
    • Feature selection: process of subset selection where we want to choose the smaller subset of the set of input attributes leading to maximum performance
    • Feature extraction: new features that are calculated from the original features
  • Decision trees

Chapter 4:

  • Neutral Networks and Deep Learning:
    • Perception model
    • Learning algorithms adjust the connection weights between neurons (pg. 88).
    • Hebbian learning rule: the weight between two neurons get reinforced if the two are active at the same time – the synaptic weight effectively learns the correlation between the two neurons
    • Error function : the sum of the difference between the actual outputs the network estimates for an input and their required values specified by the supervisor
    • If we define the state of a network as the collection of the values of all the neurons at a certain time, recurrent connections allow the current state to depend not only on the current input but also the on the previous time steps calculated from the previous inputs.
    • SIMD, NIMD
    • Simple cells vs. complex cells
  • Deep Learning
    • Deep neural networks- each hidden layer combines the value in its preceding layer and learns complicated functions of the input.

In relation to the Karpathy article, this piece helped us understand how the data we produce and generate through a “selfie” can be unpacked and mechanically understood from an IT standpoint. 


Case Study – Karpathy – What a Deep Neural Network Thinks of Your #Selfie (notes)

Convolutional Neural Networks

  • The house numbers and street signs in the graphic remind me of those “prove you’re not a robot” activities that you have to do when logging in or creating an account on certain sites. Are those just collecting human input to enhance their ConvNet algorithms?!
  • ConvNets are
    • Simple (one operation, repeated a lot)
    • Fast (image processing in tens of milliseconds)
    • Effective (and function similar to the visual cortex in our own brains!)
    • A large collection of filters that are applied on top of each other
      • Initialized randomly, and trained over time
        • “Resembles showing a child many images of things, and him/her having to gradually figure out what to look for in the images to tell those things apart.”

The Selfie Experiment

  • Interesting experiment, but not all selfies?
    • Full body portraits, couples photographed by a third party, mirror selfies (are they the same as front-facing camera selfies?), soccer players (?), and King Joffrey from Game of Thrones


    • Was this human or computer error? Also, why was Joffrey ranked lower than the soccer players? His forehead is even cut off in the shot (which was one of the tips)!
  • Didn’t give a working definition of selfie in the article
  • Didn’t need an algorithm for most of the selfie advice (be female, put a filter/border on it, don’t take in low lighting, don’t frame your head too large, etc.)
    • Is this natural human bias showing through in the implementation of an algorithm that ranks selfies (a very human idea)?
  • Reflections on supervised learning
    • Supervised learning is a learning in which we train the machine using data that is well classified. According to its definition, the selfie experiment is an application of supervised learning as the researcher fed the machine with 2 million selfies that were pre-labeled as “good” or “bad” by his criterion.  
    • In our opinion, supervised learning is not very applicable in this experiment because “good” or “bad” is an ambiguous concept, which makes it more difficult to categorize 2 million selfies to these two categories. We believe the classification of the training data for supervised learning should be uncontroversial. For example, if you are planning to train a machine to recognize a toad, you need to feed it with a great number of pictures with toads. The pictures present either toads or non-toads. But in the selfie experiment, the classification of selfies as “good” or “bad” is not convincing. It is based on individual judgement, which shows significant uncertainty. Also, as pointed out above, there are some errors in the data. Therefore, it makes us reflect on the accountability of supervised learning in the cases where the training data is not well classified. And if it is true that “the more training data, the better,” does the quality of data matter?  
    • As we see it, unsupervised learning would be better in this selfie experiment. Unsupervised learning is a type of machine learning algorithm used to draw inferences from dataset consisting of input data without labeled responses.   

In this case, unsupervised learning can help to find interesting patterns in selfies so that we can see the distinctions between different clusters. It would be more inspiring.


  • “takes some number of things (e.g. images in our case) and lays them out in such way that nearby things are similar”
  • Visualized pattern recognition
  • Reminded of artwork that is made up of smaller images (mosaic art)

  • t-SNE is a kind of unsupervised learning.

Concluding Thought(s)

  • Everyone has their unique way of taking selfies. It’s a manifestation of our personality, our digital presence, our insecurities, our “brand.” While it’s fun to run algorithmic tests for pattern recognition and even to collect information on different ways of taking selfies, if a computer starts dictating what makes a selfie ‘good’ (a subjective term to begin with) we’re taking steps toward standardizing a semi-unique form of expression in the Digital Age. If everyone’s selfies start looking the same in an effort to ‘look better’ or get more likes, the world will lose some of its charm.
  • Can facial recognition security really be trusted if there are tens, hundreds, or thousands (for some) of our selfies out there on the web being data-mined for their facial properties? Maybe so, but that seems more accessible to hackers or identity thieves than fingerprints or passwords at this point in the Digital Age. 



Alpaydin, E. (2016). Machine learning: the new AI. Cambridge, MA: MIT Press.
Karpathy, A. (2015, October 25). What a Deep Neural Network think about your #selfie [Personal Blog]. Retrieved from

Facial Recognition: How it Works + Why it Matters

Face recognition and convolutional neural network

By Beiyue Wang, Shahin Rafikian and Kevin Ackermann


Face recognition is currently very common in our lives. An increasing number of smart phones replace passwords with face recognition to increase security.  Besides, law enforcement agencies are using face recognition more and more frequently in routine policing. Once criminalsfaces were captured by street cameras, the police are able to immediately compare that photo against one or more face recognition databases to attempt an identification. For instance, last year AI security system had been launched in almost every metro station in Shanghai to track hundreds of wanted criminals. The technology can scan photos from the national database and identify a person from at least 2 billion people in seconds. It was reported that within 3 months, the technology helped the police successfully catch about 500 criminals.

As we know, humans have always had the innate ability to recognize and distinguish between faces, yet computers only recently have shown the same ability. Face recognition needs to be able to handle different expressions, lighting, and occlusions. From this weeks reading, we know that the realization of face recognition must be attributed to convolutional neural network.


For me, it is very hard to fully understand this kind of technology and I hope to get more information in class. Based on the reading, convolutional neural network where the operation of each unit is considered to be a convolutionthat is, a matchingof its input with its weight. (Machine learning) For instance, in CNN network, starting from pixels, we then get to edges, and then to corners, and so on, until we get to an image. The whole process contains mathematics and statistics. The picture below presents the process of classification or recognition, from sensing an image, preprocessing, segment the foreground from background and labeling, feature extraction, post processing, classification to decision.

Indeed, face recognition brings us lots of benefits. However, it also has many shortcomings and problems.


First, with the number of faces into database going up, face recognition is prone to error, because many people look alike in the world. As the likelihood of similar faces increases, matching accuracy decreases. It has been proved that face recognition is especially bad at recognizing the minority, young people and women. Actually, my cell phone always couldnt recognize my face to open. In my view, solving the problem of accuracy still has a long way to go.

Second, a study purporting to infer criminality from a dataset based on existing prisonersand nonprisonersfaces has serious endogeneity problems. Prison itself may advance aging, or affect routine expressions, or even lead to disproportionate risk of facial damage. Besides, many people questioned that  training data consisting of prisonersfaces is not representative of crime, but rather, represents which criminals have been caught, jailed, and photographed. Indeed, how a classifier operates leaves it vulnerable to a critique of the representativeness of its training data.

Third, as we discuss in last class, face recognition no doubt brings some discrimination problem. For example, the police now use machine learning to get a demographic pattern of criminals so they are likely to watch those people more than others, which causes some discrimination problems.

Facial Recognition and the Inner Workings of Neural Networks

Have you ever stopped to think what it is that determines when you recognize a face? You could break down the constituent elements – eyes, ears, mouth and nose – but why is it that when you see a face, you can immediately recognize it as a face? What’s more, what are the minute details and changes that separate one face from another?

To explicitly tell a machine all of the rules and definitions that make one face unique would be challenging, time-consuming and virtually impossible. Enter convolutional neural networks. Using data pools of thousands of faces, the machine learning program begins to “learn” what differentiates one face from another. Then, once the machine has learned how to recognize a face, it can apply this knowledge to faces that it sees in the future.

Let’s take Apple’s Face ID as an example to further illustrate how this process happens. Face ID is a form of facial recognition built into iPhone models past the iPhone X. The basic concept of Face ID is that the iPhone can recognize its owner’s face, and then use that recognition as a password on the device. According to Apple, “Face ID uses advanced machine learning to recognize changes in your appearance. Wear a hat. Put on glasses. It even works with many types of sunglasses” (iPhone XS – Face ID).

To recognize a face, the first step is to “see” or gather input. To do this, the iPhone projects 30,000 infrared dots on a person’s face, and then uses an infrared camera to take a picture of the facial dot map. This facial map is sent to a chip in the iPhone that uses a neural network to perform machine learning. Basically, what this means is that the chip is able to view the patterns of dots that make up someone’s face and learn these dot maps to recognize the face. The chip is learning to perform a task – recognize a face – by analyzing training examples – the initial Face ID setup (Hardesty, 2017).

To clarify, neural networks, which draw inspiration from neurons in the brain’s structure, could be described as the architecture of the machine. Machine learning is a method of “learning” within a neural network (Nielsen, 2015).

Computer vision is the “broad parent name for any computations involving visual content – that means images, videos, icons, and anything else with pixels involved” (Introduction to Computer Vision, 2018).

So, how exactly does a neural network work to learn and make decisions? Speaking broadly, imagine three sections to the neural network: an input layer, hidden layers, and an output layer. If we’re working with an image, the input layer might be made up of each individual pixel as a node to the input layer. In the case of the iPhone’s Face ID, we can assume that each infrared dot might be a node on the input layer (Machine Learning & Artificial Intelligence…, 2017). Once input data is entered into the neural network, weights are assigned to nodes within the hidden layers. These weights are multiplied together and added in complex ways depending on the input data. Each node within the hidden layer has a certain threshold that, if the threshold is met or exceeded, will “fire” just like an actual neuron. This data, fed into the input layer travels as it fires through the hidden layers “until it finally arrives, radically transformed, at the output layer” (Hardesty, 2017).

As I was reading the way in which every human-technology interaction is a programmed computer function, I began to think about the ways in which future programmed technology interactions (computer and beyond) can become both more personal and intelligent enough to adapt to our needs. But then I realized that our technology and technology programmed systems are already there — smart home devices keep a history of our interaction data for future smart predictions and vocal recognitions, mobile phone keyboards are capable of making keyboard predictions for users, streaming services recommend subscribers with certain shows/movies based on viewing history. It’s all data collection of user experiences. The Alpaydin reading can be in a dialogue with the Karpathy blog post, in that Convolution Neural Networks and scanning algorithms are trained similarly as if you were to train a child new actions and informations. The more you allow for a child to learn and experience something in particular, the more they will be familiar with it. Similarly with the predictive text  keyboard functionality on the iPhone, the more an iPhone users sends text messages, the more data the iPhone will store in order to make smart predictions.

Similar e-behavior (courtesy of algorithms designed to collect data and make such predictions) can be seen in the iPhone’s Face-ID, where Apple’s technology is able to scan the registered face in various mediums (e.g. bearded, with makeup, new hair style). It can be explored, however, that similar Convolution Neural Network data-imaging processes can be applied to facial recognition technology to even further strengthen facial recognition capabilities. In regards to Alpaydin’s discussion on social media data, is it possible that there is a breach in security with facial recognition due to how accessible imaging data is on the internet? And with access to technologies such as 3D-printing, it doesn’t seem far from impossible to be able to break into someone’s phone by 3D-printing a face based on predictions of one’s face/head structure, and access technologies that are locked via the 3D-printed face?


Andrej Karpathy, “What a Deep Neural Network Thinks About Your #selfie,” Andrej Karpathy Blog (blog), October 25, 2015,

Ethem Alpaydin, Machine Learning: The New AI. Cambridge, MA: The MIT Press, 2016.

Frank Pasquale ,When Machine Learning is Facially Invalid,, 2018

Geoff Dougherty, Pattern Recognition and Classification: An Introduction (New York: Springer, 2012). Excerpt: Chaps. 1-2.

Hardesty, L. (2017, April 14). Explained: Neural networks. Retrieved February 6, 2019, from

Introduction to Computer Vision. (2018, April 2). Retrieved February 6, 2019, from

iPhone XS – Face ID. (n.d.). Retrieved February 6, 2019, from

Machine Learning & Artificial Intelligence: Crash Course Computer Science #34 – YouTube. (2017). PBS Digital Studios. Retrieved from

Nielsen, M. A. (2015). Neural Networks and Deep Learning. Retrieved from

Pattern Recognition and its Applications

Group Study Project

By: Linda Bardha, Dominique Haywood, Yajing Hu

For this project we have divided our presentation into three parts: Key Points from the readings, Case Study of Karpathy and Applications of Pattern Recognition.

Key Points from the readings

  • A pattern is a particular configuration of data; for example, ‘A’ is a composition of three strokes. Pattern recognition is the detection of such patterns.
  • In computer science, there are many tasks for which we have tried to devise “expert systems” programmed with manually specified rules and algorithms.
  • For example, facial recognition. By analyzing different face images of a person, a learning program captures the pattern specific to that person and then checks for that pattern in a given image.
  • The learning algorithm finds those unchanging discriminatory features and the way they are combined to define a particular person’s face by going over a number of images of that person.
  • The sensory results from an e-tongue are dealt with in a similar way to the human case. Electronic signals produced are perceived in terms of pattern recognition – attempting to match a new set of sensor readings with a database of memorized taste spectra.

Case Study of Karpathy

  • Kaparthy designed a script to identify images tagged #selfie to accumulate the images. After initializing the images, convolutional networks are trained to filter the images to accurately identify images that are pulled.
  • The Covnet designed by Kaparthy gathered 5 million images tagged #selfie. Higher numbers of images enables a more accurate filtering of images. Ultimately, the number of selfies were brought down to 2 million, however, the demographic statistics of the data was not discussed. It is unclear whether the demographics of both the sample that trained the covnet and the ultimate pool of selfies impacted the final conclusions.
  • Kaparthy ranked the users which the selfies belonged to based on the number of followers. This ranking was indicative of the reach of the post and allowed Kaparthy to establish the value of the picture. Another approach to establishing the value of the post would be to calculate the percentage of followers which liked each selfie. This would likely be a better indication of the photo’s value because users with higher number of followers may have purchased their followers. Calculating the percentage of followers which liked each selfie would provide an accurate view of follower engagement with the selfie, and it would be a better measure of the photo’s value.
  • Many of the features identified by the Covnet as being contributors to a “good” selfie are the same features of any photo that is pleasing to viewers. Primarily, lighting, borders and filters enhance the visibility of the subject and the background where the subject is taking the photo. It is still unclear how the demographics of the users impacted the patterns of a good selfie, however, the pattern that selfies of women were more likely to be ranked as good selfies likely has to do with the overwhelming number of women on social media applications like Instagram.

Applications of Pattern Recognition 

Case Study: Facebook is using AI and pattern recognition to try to prevent suicide. Guy Rosen, Facebook vice president of product management talks about this initiative in an interview for the Washington Post. (I’ve attached the link in the references section)

Note: Facebook started testing this program only in the United States in March 2017. After proven successful it is now used in many different countries.

Screenshots of the suicide prevention tools in Facebook.Courtesy of Facebook

  • Facebook is using AI and pattern recognition as a tool to stop suicide broadcasts.
  • The aim of this tool is to find and review alarming posts.
  • It uses pattern recognition to scan all posts/comments for certain phrases to identify whether someone needs help.
  • In the case of live video, users can report the video and contact a helpline to seek aid for their friend. Facebook will also provide broadcasters with the option to contact a helpline or another friend.
  • Users are also given information to contact law enforcement if necessary.
  • The algorithm determines not only what posts should be reviewed but also in what order humans should review them.

My attempt to the Algorithmic Description


Geoff Dougherty, Pattern Recognition and Classification: An Introduction (New York: Springer, 2012).

Peter J. Denning and Craig H. Martell. Great Principles of Computing. Cambridge, MA: The MIT Press, 2015.

Andrej Karpathy, “What a Deep Neural Network Thinks About Your #selfie,” Andrej Karpathy Blog (blog), October 25, 2015,

Peter Wegner, “Why Interaction Is More Powerful Than Algorithms.” Communications of the ACM 40, no. 5 (May 1, 1997).

Hayley Tsukayama, “Facebook is using AI and pattern recognition to try to prevent suicide”, Financial Post, November 28, 2017,