What If Tinder Could Swipe Right on You?

With a video chat feature coming soon, the dating platform could use artificial intelligence to detect offensive images

T urning 18 is a big step in anyone’s life, one that I took a month or two ago. It usually comes with a number of rites of passage: buying your first lottery, having your first drink, entering into soul-crushing debt agreements.

Another is joining Tinder. We may be physically distanced, but the need to connect with others is stronger than ever. Absence makes the heart grow fonder, I guess?

That’s a lie, your high school relationship won’t last 2 weeks into university, JESSICA.

Tinder recently announced plans to introduce a video chat feature to the platform, and they’ve made it clear that some sort of artificial intelligence + object detection will be leveraged to prevent harassment and ensure safety during what I’m calling ‘sexy vertical Zoom meetings’.

What is object recognition?

Some definitions:

• AI (artificial intelligence): broadly refers to machine (computer) systems that perform tasks that humans can do.
• CV (computer vision): a type of AI that specifically understands the contents of an image or video.
• Object detection: adjacent to CV in that it’s looking for a specific object.

In human terms, the ‘ brain ’ is kind of like the overall concept of AI, your ‘ eyes ’ perform computer vision, and when you’re looking for something specific (your keys, a snack, true love) you’re doing an object detection task.

Most CV programs useCNNs(convolutional neural networks):

Each ‘layer’ of a CNN looks at parts of an image for certain features (via the pixel values) — starting with simple ones like edges and gradually getting more complex.

Those features are mapped to give a probabilistic location of where each one is. The highest probabilities from each section are pooled into a smaller image, where negative pixel values are turned into 0.

Then the process repeats . Eventually, all these connected layers can be compiled such that they match a certain label/outcome. Based on the probability of these outcomes, the CNN gives its best guess of what’s in the image.

It’s pretty ingenious, but also time-consuming. If Chaddington Woodsworth starts getting too… comfy on Tinder video chat, 20 seconds for processing can feel like a lifetime.

Luckily, some CNNs can operate in real-time . YOLO is an object detection algorithm that shows promise when it comes to autonomous vehicles, where real-time detection is a life-or-death matter.

The only time saying YOLO is acceptable in 2020

YOLO breaks the image up into a grid , where each grid square is responsible for predicting the size, class, and location of an object that is within its bounds.

It also gives a confidence score — how sure is it that this an object? Then, it tries to figure the class of the object (under the condition that it actually is an object). This returns a conditional probability : P(Car | Object), for example.

These confidence scores are associated with bounding boxes around the object in question, ultimately resulting in a lot of bounding boxes of varying weights. The algorithm then looks for boxes with lots of overlap — this signifies a high probability there’s an object there.

Those high-overlap predictions are output, and they’re kept if they meet a certain threshold probability score . YOLO can do this for every frame/second during a video making this the perfect tool if Tinder wants to monitor video chats for safety.

What this means for Tinder moving forward

Tinder has stated that when it comes to the balance between safety and privacy, they learn towards safety . Using YOLO real-time object detection for their upcoming video chat feature could make many people feel safer using it + deter any creeps.

The next wave of Tinder users is Gen Z — our whole lives are on the Internet, so I think we’ll be fine with Tinder watching our video chats.

How this feature might work is a different question. If Tinder sees an unwelcome/inappropriate object, would they end the call, report the user, or give you a warning?

Technically, it’s a whole other story. Classification of what counts as inappropriate changes from user to user, and throughout a relationship. Even if algorithms are customized to each user, this is going to be a computationally challenging problem.

Luckily I’m 18

3 miles away

And I love solving problems and long walks on the beach.