Fake photos of people of color won’t fix AI trends

Equipped with a Believing in the technology’s innovative potential, a growing group of researchers and companies aim to tackle bias in AI by creating artificial images of people of color. Proponents argue that AI-powered generators can bridge the diversity gap in existing image databases by supplementing them with composite images. Some The researchers are using a machine learning architecture to map existing photos of people into new races in order to “balance the ethnic distribution” of the dataset. Others, like Media is created And Qoves . Laboratory. say “the face dataset is really fair”. As they see it, these tools address data aberrations by cheaply and efficiently producing diverse images on demand.

The problem these technologists are trying to fix is ​​an important one. AI has many defects, unlock the phone for wrong person because they can’t distinguish Asian faces, false accusations people who commit crimes they don’t commit and confuse people with dark skin for the gorilla. These spectacular failures are not unusual, but an inevitable consequence of the data the AI ​​is trained on, which is largely male and white—making these tools immoral. accurate to anyone who doesn’t fit this narrow archetype. In theory, the solution is simple: We just need to cultivate more diverse training sets. In practice, however, it has proven to be an extremely laborious task due to the scale of input that such systems require, as well as the extent to which gaps are present in the data (e.g. , IBM research revealed that six out of eight the featured face dataset covers more than 80 percent of faces with lighter skin). Therefore, diverse datasets that can be generated without manual sourcing is an attractive possibility.

However, as we take a closer look at the ways this proposition might impact both the tools and our relationship with them, the shadow of this seemingly convenient solution begins to form a scary way.

Computer vision has has been developed in some form since the mid-20th century. Initially, researchers tried to build top-down tools that manually define rules (“a human face has two faces”). eye symmetry”) to define a desired image class. These rules are converted into a calculation formula, which is then programmed into the computer to help it look for pixel patterns that correspond to the patterns of the object being described. However, this approach proved mostly unsuccessful considering the sheer variety of subjects, angles and lighting conditions that can make up an image—as well as the difficulty of converting even simple rules into coherent formulas.

Over time, the proliferation of publicly available images has made a bottom-up process through machine learning possible. With this approach, a set of labeled data volumes is fed into a system. Because “supervised learning,” the algorithm takes this data and teaches itself to distinguish between the desired categories specified by the researchers. This technique is much more flexible than the top-down approach because it doesn’t rely on rules that can change under different conditions. By training itself on a variety of inputs, the machine can identify relevant similarities between images of a given class without being explicitly told what those similarities are, generating a much more adaptable model.

However, the bottom-up approach is not perfect. In particular, these systems are largely limited by the data they provide. As tech writer Rob Horning put it, technologies of this kind are “supposedly a closed system.” They have difficulty extrapolating beyond the given parameters, resulting in limited performance when faced with subjects for which they are not properly trained; differences in data, for example, led Microsoft’s FaceDetect to have a 20% error rate for dark-skinned women, while the error rate for white men hovers around 0%. The pervasive impact of these training biases on performance is why tech ethicists started lecturing on the importance of dataset diversity and why. Companies and researchers are racing to solve the problem. As the popular saying in AI, “garbage in, garbage out”.

This maxim also applies to image generators, which also require large datasets to train themselves in the art of photo-realistic representation. Most of today’s face generators use Network created by opponents (or GAN) as their foundational architecture. At its core, the GAN works by having two networks, the Generator and the Discriminator, work together. While the Generator generates images from noisy inputs, the Discriminator tries to sort the dummy images generated from the real images provided by the training set. Over time, this “rival network” allows Generators to improve and create images that Discriminator cannot identify as fake. The initial inputs serve as anchors for this process. in history, Ten thousand some of these images were required to produce sufficiently realistic results, showing the importance of a diverse set of training in properly developing these tools.


Goz News: Update the world's latest breaking news online of the day, breaking news, politics, society today, international mainstream news .Updated news 24/7: Entertainment, the World everyday world. Hot news, images, video clips that are updated quickly and reliably.

Related Articles

Back to top button