Humans vs Machines

II. Machine Learning Scope

After the release of GPT-3.5 late last year, Machine Learning (ML), particularly Generative AI and foundation models, have taken center stage in almost all IT discussions. Many research papers, blogs, and news articles emerge every day, covering topics from ML technologies to governance and trust.

As I read some of these papers, blogs, and news articles, a sense of “there is no new thing under the Sun” (Ecclesiastes 1:9b) resonates with me. Following this thought, I think that the scope of machine learning can expand.

From Wikipedia, ML was defined as:

Machine learning (ML) is an umbrella term for solving problems for which development of algorithms by human programmers would be cost-prohibitive, and instead the problems are solved by helping machines “discover” their “own” algorithms, without needing to be explicitly told what to do by any human-developed algorithms.

I agree with this basically, but how to “help machine discover” is what I want to talk about.

Learning from human brain and neural system

This is most of current research engagements from universities and companies, learning from human brain and neural system, see how human learning and using skills and knowledge, and do similar things or simulate the logic using machine.

Majority of current research engagements from both universities and companies focus on the emulation of human brain and neural systems, aiming to comprehend how humans learn, acquiring skills, and utilizing knowledge, and to simulate this logic using machines, leading to advancements in machine learning and artificial intelligence.

There has been notable progress in this domain, and arXiv.org is an excellent resource for accessing various papers. Some key papers include:

1. About a year ago, Yann LeCun released a paper about his vision: A Path Towards Autonomous Machine Intelligence.
2. A recent PNAS paper about building transformers from neurons and astrocytes.
3. Based on neural inference to create optimized chips
4. The largest map of the human brain opened doors for enhanced AI models trained not just on computational data but on a richer, biological context.
5. Study for neural networks from biological to artificial and vice versa.

However, to learn from human neural system, several areas I think we shall make better progress:

Single vs Multiple Dimensions

Human brains receive data from internal (body) and/or external world by many channels (or dimensions), typically from eyes (vision), ears (hearing), months (taste), nose (smell), skins (feeling), etc, and integrate these information to build a holistic understanding of the environment — cognitive thinking.

But, most of current machine training data are single dimension — text, image, or speech.

This may be a reason that, even machines can give us many comprehensive answers (asking ChatGPT for a business strategy, it can give you a good list which is generally better than some executive presentations), but they can not give a cognitive vision of environment like humans.

Developing ML technologies to effectively synthesize information from multiple dimensions will certainly extend its business use cases.

Note: I am using “multiple dimensions” not “multimodal”. The reason is that: multimodal is about being able to “understand” various types of data (text, image, video, audio, etc) with a certain ML model (as Google Gemini claims), but my focus here is about human connecting with environment (nature and human society) by a cognitive approach through multiple dimensions.

Forget vs Ignore

If you ask your friend a question that you believe he/she knows the answer, but he/she do not answer you, there are generally two reasons: either he/she somehow forgot it (purposelessly, for example an vacation he/she did together with you many years ago), or he/she want to avoid/ignore the context and discussion (purposely).

Machines can “learning”, but forgetting and/or deliberately ignoring are interesting topics. As I discussed in my prior blog, machines can potentially surpass humans in terms of retaining and learning vast amounts of knowledge. However, the ability to intentionally ignore or avoid certain topics is a skill humans doing routinely and machine cannot. Recent, there is a paper addressing machine “forgetting”, which, I do not think it is a truly “forgetfulness”, but just instructions guiding the machine to provide different answers.

Life-long Continual Learning

Human starts to learn after born, learning environments, learning knowledge, learning skills, and never stop — human learning is a continuous, lifelong process. There are many effort in this area for machine learning, including some recent technologies:

1. In-context fine learning and prompt tuning
2. Retrieval augmented generation
3. Self-supervised learning
4. Graph lifelong learning
5. … …

However, if you dig deep into these technologies, most of them are not truly like human life-long learning.

Break through in this area will be significant for ML.

Learning from human body (other systems)

In addition to the brain and neural system, the human body includes many other organs and systems. Biological science demonstrates that these organs and systems collaborate, function, and support one another, keeping the human body in a well balanced (resources, energy, chemicals, and more) state, and giving us a wonderful life.

There is a similar topic for ML model network, specifically for LLM — how to optimize the resource usage, or coordinate resource need of many components of a ML system in a limited environment, for example edge deployment.

I believe we can learn from how human body works, and spark ideas to be used for ML model network, for example:

1. graceful extensibility
2. Study how human organs organized and collaborate with each other to guide how LLM system operation with resources.
3. Simulate how human immune system for computer or ML security.
4. A Robust Adversarial Immune-inspired Learning System

Learning from human behavior

This is a vast field, considering that human behavior is the most intricate phenomena. I provide a couple topics below to stimulate contemplation and research in this domain.

The Prompt Engineering vs Human Learning

While exploring the techniques of Prompt Engineering, I find it exhibits many similarities to certain human behaviors.

1. zero-shot prompt: Does it like students learn and use knowledge for test or homework? For example, first grade students learn 1+1=2, and they are given problem like 1+1=? for homework.
2. Few-Shot Prompting: think of a middle term test problem with couple given hints in fourth or fifth grade.
3. Chain/Tree-of-Thought Prompting: this is what students trained at middle school or later, to resolve a (math) problem with logic steps.
4. Retrieval Augmented Generation (RAG): This is fun, what I think of is, a student doing a test, do not know a certain answer, and seek help (with help line hints, or else :-)
5. Analogical Prompting: Inspired by human analogical reasoning.

Artificial Hallucination vs Human Talking Nonsense

human talk nonsense with at least two reasons/purposes:

1. someone does not know the answer, so talk nonsense (thinking of my kid did not know her physics test question clearly, but still give a half-page answer)
2. someone want to trick or confuse others for certain purpose.

Machine hallucination kind of like first one, but far not able to act like 2nd yet.

Let this blog be a prompt

This blog isn’t intended to list all the potential aspects that machines can learn from the human brain, body, and behavior. Instead, it aims to encourage study on how to leverage familiar aspects of our nature and human functions to guide research in machine learning.