Source: Deep Learning on Medium
Dank Learning & Information Gerrymandering
Meme Generation With Deep Neural Networks and Gerrymandering in Social Networks
Besides having the most amazing title of an article written on article that I have ever seen Dank Learning: Generating Memes Using Neural Networks is a fascinating as well as troubling read. I will look into this article and the concept of information gerrymandering which is a practice intended to establish a political advantage for a particular party or group by manipulating district boundaries.
Meme Generation System
The article on arXiv published the 8th of June called Dank Learning: Generating Memes Using Deep Neural Networks caught my attention. Supposedly the meme generation system can produce a humorous and relevant caption (a big promise). It is not only defined on an image, rather it is in addition to this defined by a label relating to the meme template.
In a practical sense:
- Pretrained Inception-v3 network to return an image embedding.
- Passed to an attention-based deep-layer LSTM model producing the caption (inspired by ‘Show and Tell Model’).
- A modified beam search to encourage diversity in the captions.
- Evaluate the quality of our model using perplexity and human assessment.
- Can they be differentiated from real ones? Apparently not.
What is a meme?
According to the paper: ’A meme is an idea, behavior, or style that spreads from person to person within a culture often with the aim of conveying a particular phenomenon, theme, or meaning represented by the meme.’
What is a dank meme? According to the Urban Dictionary, which I thought to be more accurate: “The term “dank” is often used to describe a meme in which the comedy is excessively overdone and nonsensical, to the point of being comically ironic. Dank memes often contain intentionally added visual artifacts. e.g. excessively poor image quality, color bleeding or extremely saturated and modulated colors, or patterns that indicate that the image has been compressed and decompressed extensively.”
The training data was labelled and given a set of captions. We are talking of a sizable dataset that consists of approximately 400.000 image, label and caption triplets with 2600 unique image-label pairs, acquired from using a python script that they wrote. Each image is associated with a caption.
From this images were generated.
Almost as a side note at the end of the paper an important mention is made:
“Lastly we note that there was a bias in the dataset towards expletive, racist and sexist memes, so yet another possibility for future work would be to address this bias.” As such this does not mean all memes due this, yet there seems to be a tendency towards a specific type of opinion across this selected dataset.
“An analysis shows that information flow between individuals in a social network can be ‘gerrymandered’ to skew perceptions of how others in the community will vote — which can alter the outcomes of elections.”
In Nature on the 4th September 2019 there was an article written by Bergstrøm & Coleman.
- After Writing in Nature, Stewart et al.1 there are experiments and computational models to uncover a previously unrecognised obstacle to democratic decision-making. When social networks become primary conduits of information, the pattern of network connections influences what voters believe about others’ voting intentions. This influence matters, because people shift their own perspectives and voting strategies in response, either through behavioural spread known as social contagion2 or on the basis of strategic considerations.
- “…social media is starting to compete with, or even replace, nationally visible conversations in print and on broadcast media with ad libitum, personalized discourse on virtual social networks”
- In information gerrymandering, the way in which voters are concentrated into districts is not what matters; rather, it is the way in which the connections between them are arranged (Fig. 1). Nevertheless, like geographical gerrymandering, information gerrymandering threatens ideas about proportional representation in a democracy.
Social-network structure affects voters’ perceptions.
In these social networks, ten individuals favour orange and eight favour blue. Each individual has four reciprocal social connections.
a, In this random network, eight individuals correctly infer from their contacts’ preferences that orange is more popular, eight infer a draw and only two incorrectly infer that blue is more popular.
b, When individuals largely interact with like-minded individuals, filter bubbles arise in which all individuals believe that their party is the most popular. Voting gridlock is more likely in such situations, because no one recognizes a need to compromise.
c, Stewart et al.1 describe ‘information gerrymandering’, in which the network structure skews voters’ perceptions about others’ preferences. Here, two-thirds of voters mistakenly infer that blue is more popular. This is because blue proponents strategically influence a small number of orange-preferring individuals, whereas orange proponents squander their influence on like-minded individuals who have exclusively orange-preferring contacts, or on blue-preferring individuals who have enough blue-preferring contacts to remain unswayed.
This is interesting for several reasons, the most striking is perhaps;
“These online social networks are highly dynamic systems that change as a result of numerous feedbacks between people and machines. Algorithms suggest connections; people respond; and the algorithms adapt to the responses. Together, these interactions and processes alter what information people see and how they view the world.”
At present, online social networks are not subject to substantive regulations or transparency requirements.
Memes can be believably faked on a large scale and people can be influenced at scale.