Thought leadership
In the midst of all the uncertainties that 2020 threw at us, science, technology and innovation were some of the few things that managed to thrive and advance. The pandemic proved to the world that the use of technology to remain operational is indisputable, now more than ever. Whether it is online video conferences, 3D-printed respirators, or AI in hospitals: the last year showed us the remarkable ways in which technology can help society at large.
It’s too early to say how 2021 is going to play out; but it’s definitely safe to say that we’ll see even more exciting new innovations in science and technology this year. At Mobius Labs, we believe that AI and computer vision technology will be one of those fields that will revolutionize multiple industries. One of the reasons for this transformation is that computer vision is no longer limited to a scientific black-box used only by scientists and experts. The democratization of this technology has opened up various avenues for its universal application.
We are excited to share all the new features of computer vision technology that the Mobius Labs team is working on. Here’s a sneak peak!
Superhuman Search™ is a game-changing development that allows users to search images without the need of tags and metadata. This feature of Mobius Labs’ Superhuman Vision™ technology is aimed at all the businesses and industries that need to find relevant visual content on the go. One would simply have to search for the content using simple language, and this feature would display the best results, keeping in mind fine nuances like characteristics, actions, occasions or even emotions.
The first reaction to this would understandably be, “Why not just Google search it?” The aspect that sets Superhuman Search™ apart from other image galleries like Google Images is that the latter relies on image meta-data, while the former does not. For instance, when a search query is put into Google Images, the system looks for all images that have pre-assigned (by a human or a machine) metadata or tags related to the search query and consequently displays the results. In contrast, Superhuman Search™ does not require any metadata at all; it instead looks into the contents and nuances of the images themselves. So it analyses the images, and not the textual data associated with them.
The right for data privacy is something that we strongly believe in, and all the upcoming features of our technology keeps this in mind. Superhuman computer vision technology is delivered as an SDK (Software Development Kit) which can be directly deployed on the client servers instead of a third-party cloud. Furthermore, it is built for edge devices (like mobiles, tablets, laptops) and can be easily deployed on them in a matter of minutes. As a result, client data stays with the client, who are also able to customize the technology according to their specific needs.
Federated Learning is a key example of how seriously we consider data privacy. It is learning how to improve models in a fully privacy-preserving way. When it comes to Superhuman Search™, the underlying guarantee is the same. We never ask users to share any data with us, thus keeping control in their own hands.
Our vision is to cease treating data as a product; instead, our mission is to provide business with the technology that allows them to make the most of their visual data while retaining complete ownership of it at all times.
Most of the AI applications today can handle object detection fairly easily. However, most machine learning algorithms still require a huge visual data-set in order to efficiently and accurately detect objects.
Mobius Labs’ Few-shot Learning detects not only objects but also new concepts with considerably less input data. The underlying idea behind this approach is to use an extensive variety of visual concepts to pre-train the model.
Usually [unless we’re talking about our Few-shot Learning mentioned above], the accuracy of a machine learning model depends on the quantity and quality of the data used to train the model. In conventional machine learning, this data is sent to external servers for further processing and analysis. Consequently, the data is stored in a third-party cloud server, which inadvertently raises privacy concerns for the people who own this data.
Federated learning is one of the innovative technologies that Mobius Labs is currently researching on, which is a secure alternative to the above. Instead of uploading the raw user data to the third-party servers, the edge devices download a “global” model from the server. Each edge device now retrains this global model on their own local data to generate a “local” model. After local training is complete, simply the model parameters (and NOT actual data) are shared with the server.
The collected local models from different users are then aggregated (averaged) to produce a single global model. Finally, this global model is shared back to the users and then used for inferencing.
Facial recognition technology has been creating quite a buzz in recent times: but training machines how to recognize human emotions still remains a challenging task. This is one of those finer nuances that Mobius Labs is especially trying to bring in our next-gen AI technology.
At the outset, it’s important to note that “emotions” and “facial expressions” do not always correlate. We might have a “smile” on our face, but feel something completely different emotionally. Therefore, our team of scientists decided to train a model that can distinguish between different facial expressions, rather than teaching a machine how to classify a limited set of happy/sad/angry emotions; the ability to distinguish between actual facial expressions goes beyond classifying these into a simple set of facial expression tags. By teaching a machine how to recognize what similar facial expressions look like, our team is trying to differentiate expressions on a considerably more fine-grained level. As a result, the machine will be able to distinguish between different ranges of, let’s say, ‘happy’: slightly smiling, smiling with teeth, laughing out loud and the likes.
