Google has announced a breakthrough research in natural language processing called “Mind Chain Hints,” which takes advanced techniques like PaLM and LaMDA to what researchers say is extraordinary.
The fact that the mind chain hints that PaLM and LaMDA can be improved at such a dramatic rate is a big deal.
LaMDA and PaLM
The study conducted experiments using two language models, the Dialogue Applied Language Model (LaMDA) and the Pathway Language Model (PaLM).
LaMDA is a conversation-focused model, like a chatbot, but can also be used in many other applications that require speaking, conversation.
PaLM is a model that follows what Google calls the Pathways AI architecture, where language models are trained to learn how to solve problems.
Previously machine learning models were trained to solve one kind of problem, and they were basically relaxed to do one thing well. But in order to do other things, Google had to train a new model.
The Pathways AI architecture is a way to create a model that solves problems that have not necessarily been seen before.
as quoted google palmtop Interpreter:
“…We want to train a model that can not only handle many individual tasks, but can leverage and combine its existing skills to learn new tasks faster and more efficiently.”
what it can do
The research paper lists three important breakthroughs in thought chain reasoning:
- It allows language models to decompose complex multi-step problems into a series of steps
- The thought process chain gives engineers a peek into the process, and when something goes wrong, this allows them to identify where the problem is and fix it
- Mathematical word problems can be solved, common sense reasoning can be done, and according to research papers can (in principle) be solved any word-based problem a human can solve.
multi-step reasoning tasks
The study gives an example of a multi-step inference task for testing language models:
“Q: The cafeteria has 23 apples. If they use 20 for lunch and buy 6 more, how many apples do they have?
A: There used to be 23 apples in the cafeteria. They use 20 for lunch. So they have 23 – 20 = 3. They bought 6 more apples, so they have 3 + 6 = 9. The answer is 9. “
PaLM is a state-of-the-art language model and is part of the Pathways AI architecture. It’s so advanced, it can explain why a joke is funny.
However, as advanced as PaLM is, the researchers claim that the mind-chain cues significantly improve these models, which is what makes this new study so noteworthy.
Google explains it this way:
“Mind-chain reasoning allows models to decompose complex problems into intermediate steps that are solved individually.
Furthermore, the language-based nature of the mind chain makes it applicable to any task that people can solve through language. “
The research paper goes on to point out that the standard cues didn’t really improve when the model size increased.
However, with this new approach, scale has a significant and significant positive effect on how well the model performs.
result
Mind chain cues were tested on LaMDA and PaLM using two math word problem datasets.
The researchers used these datasets to compare results on similar problems across different language models.
Below is a graph image showing the results of using the idea chain prompt on LaMDA.
Results of scaling LaMDA on the MultiArith dataset show that it yields modest improvements. However, LaMDA scores were significantly higher when zoomed using the mind chain cues.
Results on the GSM8K dataset show modest improvements.
The PaLM language model is another story.
As can be seen from the graph above, the benefits of scaling PaLM with Chain of Thought Prompting are huge, and huge for both datasets (MultiArith and GSM8K).
The researchers call these results significant and a new state of the art:
“On the GSM8K dataset of math word problems, PaLM shows excellent performance when scaled to 540B parameters.
…combining MindChain Hints with a 540B parametric PaLM model achieves a new state-of-the-art performance of 58%, surpassing the 55% state-of-the-art achieved by fine-tuning GPT-3 175B on large-scale training by specially trained validators Set up potential solutions, then rank them.
Furthermore, follow-up work on self-consistency has shown that the performance of mind-chain prompts can be further improved by obtaining a majority vote over a series of generated inferences, leading to 74% accuracy on GSM8K. “
in conclusion
The conclusion of a research paper is one of the most important parts to check if the research has advanced the state of the art or is a dead end or more research is needed.
Google’s research paper conclusion section has a very positive note.
It states:
“We have explored the chain of thought as a simple and broadly applicable method to enhance reasoning in language models.
Through experiments with arithmetic, symbolic, and commonsense reasoning, we find that thought chain processing is an emerging property of model scaling that allows sufficiently large language models to perform inference tasks that otherwise have flat scaling curves.
Expanding the range of inference tasks that language models can perform is expected to inspire further research in language-based inference methods. “
This means that mind-chain hints may potentially provide Google with the ability to significantly improve its various language models, which in turn could lead to significant improvements in the kinds of things Google can do.
Citation
Read the Google AI article
Language models reason through thought chains
Download and read research papers
Mind-chain cues trigger inference in large language models (PDF)
