Does AlphaZero generalize?

In my experiments using AlphaZero methodology, I taught an AI via self-play to excel in my Cosmic Tic-Tac-Toe game below. You can play it too here! The game works just like Tic-Tac-Toe, except that there is a hole brick in one of the squares where no one can place a piece in. The first player …

Practical RAG Tips

On 16 Nov 2023, I went for Singapore's Machine Learning meetup, and the topic was on Retrieval Augmented Generation. Practical Tips There are quite practical tips being shared. Here's a summary for what to do to build your RAG database: Don't take in duplicate documents (this helps to have better variety for cosine similarity later) …

Memory for Decision Making

This is an opinion piece and I shall just describe how we may use memory to help improve our decision making in a model. Given that we do not predict the world via probabilities, we need some form of learning. Whenever we have a certain experience, we can activate our analogy-finding via memories to then …

We don’t calculate probabilities, we predict multiple outcomes via analogy

I have been thinking about how we can possibly calculate probabilities in our minds without using the notion of probability. See Part 1 on a discussion of whether the world is fundamentally deterministic or probabilistic. TL;DR: We don't know, but the world is definitely uncertain. How can we plan and act well enough to survive …

Is the world probabilistic or deterministic?

Ever since the great debate on quantum physics (and possibly even earlier), we have been asking ourselves whether the world is probabilistic - that is - is there no way to pre-determine what would happen? Einstein was adamant that "God does not play dice" and the world should fundamentally be certain. Advocates of quantum theory …

A Modern Take on “Godel, Escher, Bach: An Eternal Golden Braid”

Recently, I came across a book by Douglas Hofstadter, which talks about the fundamental concepts underpinning intelligence, while at the same time welding in ideas from three people: Godel, a mathematician who famously invented the Godel Incompleteness Theorem, Escher, an artist who melds in reality and fantasy into his paintings, Bach, a musician who composes …

How do we generalize?

Comparing us humans to current machine learning/AI technology, we can perhaps conclude that humans at least do one thing better - we can generalize to unseen tasks, without needing loads of information. For example, if I know how to drive a car on a right-lane road, I can generalize what I need to do on …

Recurrent Neural Networks (RNN) vs Markov Models, which is better for Text Generation (Part 2)

Comparing on the News Dataset This is Part 2 of comparing RNNs to Markov models. Do refer to Part 1 for more on the methods used. The notebook used for the code can be found on my Github link. https://www.youtube.com/watch?v=YykFbnVdqKE&list=PLcORGO1bTgjUrFJJr1nOnUV7zp0esSHp_&t=0s Part 2: RNNs vs Markov Models (News / Coding Dataset) In order to have a …

Recurrent Neural Networks (RNNs) vs Markov Models: Which is better for Text Generation? (Part 1)

I was interested to see how an RNN would compare to Markov models to generate text. I have read some blogs stating that a Markov model could work pretty well too, so I was curious whether it really does. So, here's my attempt to compare both kinds of models and compare their performance on text …

How do babies think? (My proposed) Monte-Carlo Simulation Voting Hypothesis

Recently, I went to visit my nephew, and I realize that he has taken to calling his mother "bicycle" when the actual name is "Jessica". Then, I came to wonder, this is probably an instance of best-matching, where the baby simply speaks whatever sounds closest to the desired sound. He probably tried out various pronunciations …