Trinity is a state-of-the-art virtual assistant service that is being developed by Interlink AI as a Home virtual assistant.
Artificial intelligence is simply enabling the ability for machines to mimic natural human intelligence for complex task handling. The modern Artificial intelligence is only capable of specific task handling based on prior training/learning.
Machine Learning:
Artificial Intelligence is included with several subfields. One of the major subfields of modern AI is machine learning. Machine learning is a major subfield of AI that enables computers/machines to learn by analyzing data and detect data patterns to perform specific tasks by themselves without human intervention. The core principle of machine learning is to build and implement algorithmic structures (ML algorithms) for data analyzing and make decision based on data pattern results.
There are several types of ML techniques we are using build more advanced features for Trinity, our newest home AI-assistant. Using ML techniques will enhance the intelligence, personalization, accuracy, and adaptability of our AI-home assistant.
Supervised Learning:
supervised learning is a ML technique that use labeled datasets train ML algorithms to detect data patterns that uncover the relationships between input and out data. The primary goal is to create a model that can predict accurate output results on new data based on labeled data.
supervised learning is being used on trinity for multiple purposes.
Natural language processing:
Trinity using trained models to understand user voice commands. Automatic speech recognition models is being used for more robust recognition results. The fine-tune process is developing more accurate NLP models and using more advanced sentimental models for emotional detections based on voice tone.
Supervised learning-based NLP is also very useful natural language understanding. For example, user wants to set the alarm next day 6 AM. There are several phrases user can use in English language to make demand such as, “make alarm tomorrow at 6 A.M” or “set the alarm for tomorrow 6 O’ clock.” etc. With supervised learning datasets, Trinity can understand the both sentences require the same action.
You can watch this demo for intelligence interaction using NLP: YouTube
Facial recognition and emotional intelligence:
real time facial recognition is one of the major advanced feature in Trinity. Trinity can use frontal camera unit stationed near the front door to detect humans and recognize/identify them in real time. For example, if are registered as a resident in the home unit in trinity system, when you comes from home, Trinity can greet you in advanced and start a conversation with you.
With more advanced Algorithms, Trinity will be able to report or recommend things specific to you based on your preferences. For another example, Trinity can give you a report that is specifically made for each user such as emails, messages etc. With these features will make Trinity not a just another virtual assistant.
Here is a sample little demo version on facial recognition-based interaction: YouTube
Trinity is designed as an emotional companion. Since Trinity is a AI home assistant, It is important to have the ability to detect human emotions and respond to them. Trinity has two emotional detection modes. They are based on visual input data and voice command data. For example, Trinity can detect your emotions based on previous two methods and provide you with companionship such as recommend a movie or play a song for fixing user’s mood. Following figure shows how trinity will detect human emotions based on real time data analyzes.
Personalization:
Personalization a major factor for AI-virtual assistants. Trinity is being designed to give a personalized experience-based on user preferences and past interaction for each user. For example, user is asking about specific stock prices every morning. Based on this Trinity will make a report and give to user even though he didn’t ask for it in the morning.
Smart security:
Trinity is also fully capable of providing 24/7 security for the home unit. We are using Supervised ML algorithmic structures for real-time data analyzing and decision making. Trinity smart security mode contains detection, facial recognition, tracking, and responding algorithms for activities including:
- abnormal activity detection
- intruder detection
- object detection
- movement tracking
- alerts and warnings
- live remote monitoring
- emergency contact activation (such as law-enforcements)
- emergency lockdown activation
Semi-Supervised learning:
This is another common ML technique that uses both labeled and unlabeled datasets for training the ML model. This process allows the model to train from a small amount of labeled data while leveraging the information present in a larger pool of unlabeled data, effectively bridging the gap between supervised and unsupervised learning methods. This can also increase the accuracy than a fully-supervised ML model.
Trinity is using this method for better performances in following areas.
continuous learning based on user interactions:
Trinity is being designed to use semi-supervised learning to understand the user behavior/feedback and improve its responses, such as scheduling new tasks, mimic users word slangs, use new nick names etc.
Adaptivity for situation:
This is also a major improvement for an AI home assistant service since it can change the responding approach based on the current situation. For example, if the user is having a major health problem, Trinity will only recommend health foods and always remind of medicines.
Cost-effective improvements:
With semi-supervised learning models, Trinity is able to extract more data and improve performance without human interventions as a state-of-the-art AI home assistant. This is helpful for both users and service providers.
Reinforcement learning:
Reinforcement is also a major player in building powerful AI assistants. RL can be applied to your AI virtual assistant in multiple ways to enhance its adaptability, learning capabilities, and decision-making. Since you want your assistant to learn new facts via direct user confirmation instead of a reward system, we can implement a custom RL-like approach that adapts based on user feedback.
Trinity is using RL for several purposes:
updating data by using human feedbacks as direct reinforcement:
Trinity is learning by user interactions. such as when user says “yes” to some fact it will learn that the fact is true. If not user can say “no, here is the correct answer.”, Trinity can update the database with corrected facts/data. Then in a later occasion when the same topic is in a conversation, Trinity can use the stored data for better communication or performance.
Conversational optimization:
Trinity is being designed to learn from user feedback to improve answers over the time. for example, positive and negative feedbacks will be used to determine which answer should be used in next time.
This can also used to shape the responses. for example, when user asks “what are my plans for today?”, trinity can give a generally long response. But if user prefer shorter response, it can be adjusted by just saying “next time only tell plans within next 12 hours”.
change voice interactions:
RL can help optimize how the assistant modulates its voice, speed, or tone based on user preferences. For Example, If you frequently ask it to “speak slower,” it learns that you prefer a slower speech rate.
Implementation of Reinforcement learning to Trinity:
For implementing Reinforcement learning-based operations into Trinity’s core functionalities, We are using a hybrid approach rather than traditional strategy. We are using positive confirmation such as “That’s right”, “correct”, “yes” as the reward and negative confirmation or correction as punishments. Trinity can update the database based on this confirmation and correction method for accurate information and communication.
Deep Learning:
Deep learning is advanced machine learning algorithms that use neural network architectures. Trinity is designed to use deep learning algorithms such as YOLO for computer vision-based tasks including object detections.
As discussed before, we are using several subfields of AI including machine learning algorithmic structures to build a state-of-the-art AI home assistant.
Watch the full demo on the Trinity prototype: YouTube