︎

















MENtal AmbIvalence



AIGC Experiment & GPT Architecture Visualization










Tool

Processing (Java)
Google Colab(Python)
Project

CDP Colloquium III

Role

Machine Learning
GPT Training
Immersive Art Design
Creative Coding
Conceptural Design
Time

Feb 2023 ~ May 2023

I.    Background


GPT has become a hot topic recently, as AI technologies are playing a variety of roles in people's lives. With the current trends, the integration of AI into our daily lives and even helping us manage stored memories is no longer a fiction scenario. Imagine a future where people have AI chips installed in their brains, using artificial intelligence to manage the storage to access our real memories. But before that, backing the tape up a bit, how does AI perceive, understand and learn from human emotional memories? This is the questions that should be explored and considered in this project.



What is the current AI technology's processing mechanism for human memory?


What can be specified as emotional text, image, sound, smell, touch, perception, etc.?


And what will be the future attitude of human beings to deal with third-party technology to interfere with human personalized privacy, emotions and memories?

II.   GPT


 So, what is GPT?

GPT is a neural network machine learning model trained using internet data to generate any type of text. In simpler words, it requires a small amount of input text and will generate large volumes of relevant and sophisticated machine-generated text. There are a variety of functionality in GPT-3 which includes, text completion, code, completion and chat assistance, etc.




The GPT-3's simulation of human speech intonation, content, and logic is derived from a large amount of data training. But it is not difficult to see from the conversation with it that the context of its dialect is indeed machine-like and designed to be rational. The official GPT-3 does not understand the details of the user's emotional input and can only respond with plain and general answers, which inspires me to think whether the result will be more human if we use more personalized and emotional text data to train GPT-3.

II.   Memory


Then how to define memory?

Quoted from Epistemology of Memory, memory is an approximation of some facts. In other words, memory generally alters significantly what enters it. When we recall some episodes of the past, those recollecting are not the retrieving, but rather the generating of representations of the past. It actually generates new beliefs about the past.


There are two common categories of memory, which are Semantic memory that indicates proposition of memory, and Episodic memory that refers to actual events of memory. Epistemic theory, therefore, mainly focuses on these two kinds of memory and illustrates an interesting concept, that is whether practical knowledge can be fully understood in terms of various propositions. In simpler words, the goal is to explore whether knowledge-how is reducible to knowledge-that, and how to use various knowledge-that to build knowledge-how.