Human Data Lab
The lab is run by Dr.Deokgun Park at University of Texas at Arlington. The core area of interest is Artificial General Intelligence(AGI). The goals of the lab are explained in these videos .I have been working in HDILAB over the course of Spring 2019. This was my first ever experience in research work and I quite enjoyed the workflow. Documentation was an alien concept that I was introduced to which seemed like the missing piece of the puzzle for me to further improve and reflect upon what I have learnt. It had been my lifelong dream to be developing games at some calibre and I finally took a step into this world learning animation and VR control mapping in Unity3D. I was exposed to the current state of AI and deep learning approaches. I got experience giving a live seminar over Multi-Layer Perceptron implementation with numpy in python3. I read about many ways of creating virtual environment and how reinforcement learning played a key role in the current approaches.
Dr.Park and I had a discussion about tasks that I was assigned and how I executed them. My main problem was that I kept concentrating on multiple things and was not able to focus on one area. I need to improve my documentation skills. We discussed my future plans and what was my vision for my career. I tend to have high expectations as it was pointed out and need to document even the failures. Finally, we decided the following tasks for me to pursue during the summer 2019.
My goals
I have two main goals that I am working on over the summer 2019.
- Building virtual environment for incubation of AI.
- Understanding and implementing the Hierarchical Temporal memory(HTM) algorithm.
Virtual environment for AI:
Using Unity3D, learn develop a simulation/ a game that can be played by an AI similar to Obstacle Tower challenge . I am familiar with Blender for creating animations and have started to learn animation with unity3D. Some of the tasks include animation from motion capture data, both face and body. Faceware Live is the face motion capture software that has a plug-in for unity3D. The challenge would be in creating a seamless virtual human behaviour within the game. Marching cubes algorithm is a handy tool for creating various terrains. This can help build a robust environment that is auto-generative. Generative adversarial Networks come handy when you have to replicate a real-world 3D object in the game. The idea is to make 3D simulations that have parameters that can be fed into an AI algorithm and be controlled to play the game. Additional feature would be giving access to us to play and see what the AI would see, a first or third person view using a Virtual Reality headset.
Temporal Memory Algorithm
Hierarchical Temporal Memory is a hypothesis by Jeff Hawkins which is based on the working of the neocortex, the outermost part of the human brain. This theoretical framework encapsulates the essence of intelligence and is a new approach to building truly intelligent machines. One of the key factors for the working of this algorithm is Sparse Distributed representation (SDR) which, is proposed and the way our brain stores information. The other concept is spatial pooling, which can be simply put as neurons that fire together work together. My goal is to understand the algorithm completely and implement it in a simple use-case, prediction of characters in a sentence.
- This is a book being authored by Numenta, an AGI research company co-founded by Jeff hawkins good read...
- On Intelligence is a book authored by jeff hawkins that gives the ideology behind the thinking behind the human brain