Google has announced a breakthrough in creating artificial intelligence architectures that can handle millions of different tasks, including complex learning and reasoning. The new system is called the Pathways Language Model, or PaLM for short.
PaLM is able to surpass the current state of AI technology and beat humans in language and reasoning tests.
But the researchers also point out that they cannot escape the inherent limitations of large-scale language models that can inadvertently lead to negative ethical outcomes.
Background information
The next few sections are to clarify the background information of the algorithm.
Small sample learning
Few-shot learning is the next stage of learning beyond deep learning.
Google Brain researcher Hugo Larochelle (@hugo_larochelle) said in a presentation titled, Use meta-learning to generalize from a few examples (video) explained that with deep learning, the problem is that they have to collect a lot of data, which requires a lot of human effort.
He pointed out that deep learning may not be the path to AI that can solve many tasks, because with deep learning, each task requires millions of examples to learn every ability that AI learns.
La Rochelle explained:
“…the idea is that we’re going to try to solve this problem very directly, this problem of few-shot learning, which is a problem of generalizing from a small amount of data.
…the main idea I’m going to introduce is that instead of trying to define what this learning algorithm is in terms of N and using our intuition to determine what is the right algorithm for few-shot learning, actually try to do it in an end-to-end fashion.
This is why we call it learning learning or what I like to call meta-learning. “
The goal of few-shot methods is to approximate how humans learn different things and can combine different knowledge to solve new problems that have never been encountered before.
The advantage of this is that a machine can use all the knowledge it has to solve new problems.
In the case of PaLM, an example of this ability is its ability to explain a joke that has never been encountered before.
Pathway artificial intelligence
In October 2021, Google published an article laying out the goals of a new AI architecture called Pathways.
Pathways represents a new chapter in the development of AI systems.
The usual approach is to create algorithms that are trained to do specific things well.
The Pathways approach is to create a single AI model that can solve all problems by learning how to solve them all, avoiding the inefficient approach of training thousands of algorithms for thousands of different tasks.
According to the Pathways documentation:
“Instead, we want to train a model that can not only handle many individual tasks, but can also leverage and combine its existing skills to learn new tasks faster and more efficiently.
That way, what the model learns by being trained on one task—for example, learning how aerial imagery predicts the elevation of terrain—can help it learn another task—for example, predicting how floodwater will flow through the terrain. “
Pathways defines the path forward for Google to take AI to the next level to close the gap between machine learning and human learning.
Google’s latest model, called the Pathways Language Model (PaLM), is the next step, and according to the new research paper, PaLM represents a major advance in the field of artificial intelligence.
What makes Google PaLM compelling
PaLM extends the few-shot learning process.
According to the research paper:
“Large language models have been shown to achieve superior performance on a variety of natural language tasks using few-shot learning, which greatly reduces the number of task-specific training examples required to adapt the model to a specific application.
To further understand the effect of scale on few-shot learning, we trained a 540 billion parameter, dense activation Transformer language model, which we call the Pathways Language Model (PaLM). “
There are many published research papers describing algorithms that perform less well than the current state of the art or achieve only incremental improvements.
But that’s not the case with PaLM. The researchers claim significant improvements over the current state-of-the-art models, and even better than human benchmarks.
That level of success makes this new algorithm compelling.
The researchers wrote:
“We demonstrate the continued benefits of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks.
On many of these tasks, the PaLM 540B achieves breakthrough performance, surpassing state-of-the-art fine-tuning on a set of multi-step inference tasks, and surpassing the average human performance on the recently released BIG-bench benchmark.
A number of BIG-bench tasks show discontinuous improvements in model size, implying a dramatic increase in performance as we scale to the largest model. “
PaLM outperforms state-of-the-art on English NLP tasks, which makes PaLM important and compelling.
In a collaborative benchmark called BIG-bench, which consists of more than 150 tasks (related to inference, translation, question answering), PaLM outperformed state-of-the-art, but underperformed in some domains.
Notably, humans outperform PaLM on 35% of tasks, especially math-related tasks (see Section 6.2 BIG-bench of the research paper, p. 17).
PaLM is better at translating another language to English than translating English to other languages. The researchers say this is a common problem that could be solved by prioritizing more multilingual data.
Nonetheless, PaLM outperforms other language models and humans almost across the board.
reasoning ability
Of particular note is its performance on arithmetic and commonsense reasoning tasks.
Examples of arithmetic tasks:
question:
Roger has 5 tennis balls. He bought two more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls does he have now?answer:
The answer is 11.
Common sense reasoning example:
question:
Sean is in a hurry to go home, but the light turns yellow, what is he forced to do?
Answer choices: (a) take time (b) mess around (c) walk slowly (d) ocean (e) slow downanswer:
The answer is (e) slow down.
generate explanation
Another remarkable example is how advanced PaLM is in generating interpretation tests.
Here is the explanation from the research paper:
“The goal of this section is to demonstrate the capabilities of PaLM in interpretive language generation for prompts requiring complex combinations of logical reasoning chains, world knowledge, abstract language understanding, and lexical language understanding (e.g., puns).”
explanation of the joke
Here is an example of a language model explaining a joke:
joke:
I plan to fly to visit my family on April 6th. My mom said, “Oh, great, it was your stepfather’s poetry reading that night!”So now I’m taking off on April 7th.
Model Description:
The joke is that the speaker’s mother tried to get them to their stepfather’s poetry reading, but the speaker didn’t want to, so they changed the flight to the day after the poetry reading.
chain of reasoning
This is an example of a chain of reasoning, where multiple concepts and entities need to be properly connected and considered to create an answer.
“enter:
Guido is writing a Python program. He’s been confused about variable scope and looked it up on StackOverflow.Is it possible that Guido in the prompt refers to Guido van Rossum?
Model output:
Guido van Rossum is the creator of Python. He doesn’t need to look up variable scopes on StackOverflow. The answer is no”.”
Here’s another example:
“enter:
Jennifer looked out the window and saw a very cool cloud below her. She unbuckled her seat belt and walked to the bathroom. Could Jennifer be traveling more than 300 miles per hour relative to Earth?Model output:
300 miles per hour is about 480 kilometers per hour. That’s about the speed of a commercial airplane. Clouds are usually below the plane, so Jennifer might be on the plane.The answer is “yes”. “
The next generation search engine?
The above examples of PaLM’s complex reasoning capabilities demonstrate how next-generation search engines can leverage knowledge from the Internet and other sources to answer complex answers.
Achieving AI architectures that produce answers that mirror the world around us is one of Google Pathways’ stated goals, and PaLM is a step in that direction.
However, the study’s authors stress that PaLM is not the final word on AI and search. They made it clear that PaLM was the first step in what Pathways envisioned as the next search engine.
Before we go any further, there are two words, arguably jargon, that it is important to understand what PaLM is.
- Way
- generalize
the word”Way“Refers to how things are experienced or the state of their existence, e.g. texts read, images seen, things heard.
the word”generalize“In the context of machine learning, it’s about the ability of language models to solve tasks that they haven’t been trained on before.
The researchers noted:
“PaLM is just the first step in our vision to build Pathways into the future of ML expansion at Google and beyond.
We believe that PaLM provides a solid foundation for our ultimate goal of developing a large-scale modular system that will have broad generalization capabilities across multiple modalities. “
Real-World Risks and Ethical Considerations
What makes this research paper different is that the researchers cautioned against ethical issues.
They say large-scale language models trained on web data absorb many of the “toxic” stereotypes and social differences propagating across the web, and they say PaLM is not immune to these harmful effects.
The research paper cites a 2021 Research Papers Explores how large-scale language models facilitate the following harms:
- Discrimination, exclusion and toxicity
- infohazard
- The dangers of misinformation
- malicious use
- Human-computer interaction hazards
- Automation, Access and Environmental Hazards
Finally, the researchers point out that PaLM does reflect toxic social stereotypes, and make clear that filtering these biases is challenging.
The PaLM researchers explained:
“Our analysis shows that our training data, and therefore PaLM, do reflect various social stereotypes and toxic associations around identity terms.
However, removing these associations is not trivial…Future work should focus on effectively addressing these undesirable biases in the data and their impact on model behavior.
Meanwhile, any practical use of PaLM in downstream tasks should perform further contextualized fairness assessments to assess potential harms and introduce appropriate mitigation and protection measures. “
PaLM can be seen as a peek at the next generation of search methods. PaLM claims to outperform state-of-the-art technology, but the researchers also say more work remains to be done, including finding a way to mitigate the harmful spread of misinformation, toxic stereotypes and other undesirable outcomes.
Citation
Read the Google AI blog post on PaLM
Pathways Language Model (PaLM): Scaling to 540 Billion Parameters for Breakthrough Performance
Read the Google research paper on PaLM
PaLM: Extending Language Modeling with Pathways (PDF format)
