See the science of behavior & psychology of team dynamics with AI

The complexity of human emotion as it relates to conveying a message, compounded by multiple channels of communication, is a challenge for individuals & organizations. This has implications on conversations that guide critical decisions made and crucial actions taken. Recent developments in the field of AI have created an opportunity to augment communication with insights found in Emotional Intelligence (EI). We are very excited to apply machine self-learning, memorization and other techniques to elevate the way we communicate.

Emotional Intelligence

The Science of Emotion

Emotion is intertwined with cognition. It is generally the case that emotional thoughts precede and strongly influence our internal decisions and external conversations. The emotions we experience are often what move us to action. Emotions, even when kept private, can be publicly perceived. They are translated in voice, facial expression, posture and specific behaviors. Beyond translation, physical manifestation of an emotional state can be found in an effected heart rate, blood pressure, skin conductance and endocrine responses.

The Role of Emotion in Communication

Psychological theories emphasize that a key role for emotion is as a tool for communication. It is rare to speak openly about one’s internal state, in part because development of the ability to communicate through emotional cues largely precedes higher order explicit communication processes (Buck, 1984), and because emotional cues are sent frequently without awareness (Buck, 1984; Gross, 2001). Communicating via emotional signals is valuable for coordinating interdependent activities and teamwork. It can provide usable information regarding the reactions, intentions, preferences and likely future behaviors of others in the workplace. Humans attribute meaning to the expressive displays of those around them, regardless of whether they interpret correctly or unknowingly misunderstand. Sound judgement of one's own feelings and those of others is crucial to maintaining and improving the quality of both personal and professional relationships.

Emotions in Speech, Text & Video

Emotions are expressed within speech, text and physical signals. They are the fundamental aspects of human communication. Speech is an acoustically rich signal that provides information about a speaker's emotional states. While text alone is a more difficult medium to analyze for communicated emotion, it is the format most commonly used. Beyond speech and text, facial expression and physical movements captured in video are extremely reflective of human emotions.

Artificial Intelligence

Diving Deep

Communication takes shape in many forms, and each format provides its own unique set of nuances that need to be evaluated and understood when building predictive models. This is only the first of many challenges. Beyond this initial complexity, there is anywhere from 10x to 100x less data, making architecture of efficient deep networks more challenging. Procuring such a large labelled dataset is impractical and also expensive. Another equally important challenge is the ever-increasing data. As people continue using our API's, we get more data-streams (usually unlabelled) which need to be used quickly to improve existing models.

To tackle these challenges, we've built a complex grid of pre-trained networks, which ingest data from various formats and learns the latent data representations. The system has also been designed to reduce the turnaround time in receiving new data and serving the corresponding updated model. 

Transfer Learning

Given a source domain DS and learning task TS, a target domain DT and learning task TT, transfer learning aims to help improve the learning of the target predictive function fT (·) in DT using the knowledge in DS and TS, where DS ̸=DT, or TS ̸=TT. In various stages of our training pipeline, we use variants of Transfer Learning, namely, Inductive Transfer Learning (labelled data in DS) and Self Taught Learning (unlabelled data in DS). While using Transfer Learning, the route is not always easy to navigate. Negative Transfer is a classical issue when the source domain and source task are causing a reduction in performance of the target task. Thus, it requires a lot of expertise for using these collections of pre-trained networks, primarily in answering (1) what to transfer, (2) how to transfer & (3) when to transfer.

Continuous Learning

With our set of rich API's, large amounts of data get ingested into our training pipeline. Upon running models, we provide the needed flexibility to give feedback on model results. We make use of this streaming data and its corresponding feedback to continuously refine our networks. These refined models are instantly plugged into the serving pipeline. The serving pipeline performs A/B testing. Depending on the results, the new model is deployed for all users. This ensures that our API's are always serving the most up to date version of all the models.