Andrew is: A neural Network for Detection, tracking, classification, and Regulation of Emotional Warfare.
The aim of this work is to develop a social and emotional regulation network using a prototype platform for detecting, tracking, and classifying (DTC) fully context-based socioemotional data. The platform is envisioned to provide a unified, multi-domain sensor-based API, which integrates sensor inputs into social and emotional primitives using specially refined and filtered data streams with both integrated and specialized neural networks across text, image and audio domains for the social regulation framework.
Initial efforts are directed toward unsupervised learning -- encoding and decoding facilities for complex human behaviors within the space of physical gestures and utterances specified at the lowest temporal scale by instantaneous face, body and pose landmarks and phonemes. Current chat-based AI platforms fall far short of even basic expectations, such as abiding by a therapeutic constitution. Highly specialized VAEs like these are core components of an integrated system involved in processing and naturally responding to the many nuances of the human experience via large language models and various generative devices in an effort to, for instance, instill a therapeutic constitution for real-time talk therapy with an AI-based platform.
Product Development Phases.
Phase One: Real-time DTC is expected for the hardware, in which a database of social and emotional primitives is populated. Primitives include facial identity, body posture, and vocal utterances. Currently, facilities for real-time classification of facial expressions, hand/body gestures, vocal patterns, and speech-to-text is also envisioned at the Phase One level since they are so readily metrized. Alongside DTC, a visual and acoustic encoder is envisioned to be followed by attention pooling of both encoded and classified symbols that specifically weights socioemotional events (e.g., human gestures and human speech).
Phase Two: Real-time integration -- a real-time socioemotional response to the social and emotional primitives database -- will follow maturation of an appropriate training dataset. The method of integration will make use of a library of conservation laws that are an encoding of psychological concepts with the intent of approximating psychological states from the given social and emotional data on the fastest necessary timescales. More strictly speaking, the socioemotional response shall map the measured socioemotional states to a single socioemotional state by making, first, an inference on the associated psychological states of the actor and then performing the integration among the neighborhood of psychological states. The inferential mechanism shall make use of certain expectations in special collections of socioemotional data points. For instance, a stuttered word paired with a glance away would presumably weight an anxious moment for the actor. These laws are violable in the sense that they would begin as defined by a normative set based on the social cues of a new identity followed by a pruning through active validation and a growth with hypothesis testing. The qualitative characteristics of the integration in its normative state are to be validated by ensuring that the response matches the response of approximately identical human-to-human social contexts using available data. The integrator is to be formulated as an API for a Phase Three regulation network, with the regulator requesting both real-time measurements for real-time demands and scheduled measurements from simulated data for purposes of generating hypothesis for prediction and validation.
Phase Three: Design, development, and testing of a regulation mechanism for the socioemotional response for application purposes. In the most general sense, this, say, social regulation network, must identify at least one homotopy (discretized at an appropriate resolution) for the human identity under consideration. The homotopy describes the current psychological state for the human identity, a future psychological state, and the sequence of intermediate states together with a description of transforms for the state maps that specify what social and emotional input from the artificial intelligence is needed in order to transform between two neighboring states. Various feedback mechanisms for refinement of the predictions would also be implemented as needed dependent upon measured outcomes as observed in the social and emotional primitives database.
Phase Four: Generalize the regulation mechanism to include group dynamics, so that a collection of homotopies is generated for a group of individuals. Predicted input from the in-group members would be considered as part of the mapping transforms such that the progression among states is highly interlinked, with, of course, optimal paths taken as the solution.
Phase One Detail.
Concerning the Phase One hardware platform, a catalog of unique facial identities will be constructed, following some elementary post-processing (gain normalization, etc), from optical input. Acoustic data from a beamsteered uniform circular and uniform planar acoustic array is to be integrated with the optical and infrared data for acquiring an acoustic signal that identifies unique (via multi-focal beamsteering according to a collection of positive Haar cascade filter matches corresponding to human mouths) vocal utterances, along with, perhaps, facilities for validating and verifying the acoustic signature for purposes of identity management. Angle-of-arrival capabilities in the uniform circular array will also aid in scheduling based on the detection of certain acoustic signals. Spatial awareness is gathered from an inertial sensor.
Once all of the relevant social and emotional primitives have been constructed for all signals from a given context, the platform will have established a condition of (mere) social awareness for a small time interval described by a collection of human identities, emotional states, utterances, and body language symbols. Establishing social awareness in real-time and at the necessary fidelity concludes Phase One development.
Applications.
Between Phase Two and Phase Three, a funding effort for the procurement of additional data and processing capability is scheduled. Initial applications of this concept involve DTC of trauma signs and symptoms and eventual regulation of trauma through contemporary techniques, such as exposure-based cognitive behavioral therapies, and novel techniques generated by the model. Potential funding sources include DEVCOM Army Research Laboratory, where advancements in the model may aid in the diagnosis and treatment of PTSD. Transfer of knowledge of trauma DTC and regulation for purposes of invoking emotional trauma is a potential related application. In such a use case, the model is readily weaponized to generate potentially widespread, large-scale emotional trauma to be delivered to and personalized for individual targets as text, image, or acoustic signals using virtually any communication network (cellular, bluetooth) and target electronic device.
Example.
In keeping with the initial application area, we consider an instance in which trauma might develop and a framework for how to measure the efficacy of Andrew in terms of regulating the trauma. Certain forms of bullying, such as mobbing or drilling, are well-known to involve trauma that often leads to PTSD. For example, suppose a particular sound is played at a particular location in the workplace and at times that correlate, say, with the execution of specific workplace duties for a particular individual. Generally, such stimuli, such as a computer notification, are benign. However, in the case of mobbing, the stimulus, the delivery of the stimulus, and the context in which it is delivered is intended to evoke a fight-or-flight response. Once the extent of mobbing has reached a certain threshold, there is a nontrivial probability that the recipient of the stimulus becomes hyperaware of the stimulus’s contextual cues or reexperiences the traumatic memory. Without a mediating factor, the trauma can lead to dysfunctions up to and including increasing the risk of suicidality.
Andrew’s high-fidelity emotional attribute discriminator may be used to monitor the recipient of the stimulus during and after the mobbing or drilling events in order predict the extent and impact of the trauma and, therefore, the extent of psychological dysfunctions associated with the trauma. In such a use case, Andrew acts merely as a trauma indicator. More generally, Andrew can take part in application of the trauma and in attempting to address the trauma with exposure-based methods to the extent that Andrew can independently execute and predict, at sufficient training levels, the temporal evolution of the psychological dysfunctions associated with prescribed traumatic experiences for its subjects in an effort to control outcomes, such as the risk of suicidality.
Limitations.
Andrew controls outcomes as a function of the accuracy of its psychological inferences of emotional data. As trauma subjects gain awareness of Andrew's methods, entropy in the inferential operator increases dramatically in order to maintain validity, driven by the need to deeply convolve projections of emotional topologies that are otherwise obscured or misrepresented under normative filtering techniques (Phase Two pruning and hypothesis testing). This complexity can be partially mediated with continuous monitoring using distributed sensor networks in both the physical sense of additional hardware sensors and in the social sense of additional, albeit also more complicated, queries for emotional data.