What if our brains were actually just an artificial intelligence created using technology that is still unknown to us? Discover in this file three common points highlighted in scientific studies that risk destroying all your certainties!
L'Artificial intelligence is inspired by human intelligence. The ultimate goal of this technology is this Creating machines capable of thinking and think like people.
However, have you ever wondered? if we are not artificial intelligences ourselves extremely advanced?
As research in artificial intelligence and neuroscience advances, We realize that AI and the brain have much more in common than was previously assumed.
In this file You will discover several extremely disturbing similarities This is highlighted by various recent scientific studies…
AI can reflect on its mistakes to improve
A study published in March 2023 by researchers at Northeastern University and the Massachusetts Institute of Technology (MIT) shows that LLMs (Broad Language Models) are capable of learning from their own mistakes… like humans !
According to the study authors, it might be possible to teach AIs to develop this behavior Allow them to reach a new level in solving problems independently.
They explain: “Introspection allows people to find solutions effectively new problems through a trial and error process “.
Your suggestion called “Reflection” is an approach consisting of Equip an agent with dynamic memory and introspection skills to improve existing skills for action selection.
This framework would make it possibleTeach AIs about promptsto apply the technique of trial and error to their results. So, just like humans, they can Keep trying again if you don't succeed.
To test this method, the researchers Presented to solve GPT-3 and GPT-3.5 tasks problems and asked him to solve them. Whenever he made a mistake, the reflection technique was given to him as a stimulus to make him find his mistakes on his own.
According to them, this process helps the program develop like a human. The AI is then able to recognize its own hallucinations, avoid repetition in action sequences, or even create an internal memory map of a specific environment.
This enabled the modelSignificantly increase your success rateand could lead to complete AI automation…
Like our brains, AI also needs sleep to remember something
One of the main limitations of current AI is that it can only master one task and forget everything what they know as soon as they learn another.
Opposite to People who are able to accumulate knowledge In order to adapt and solve problems throughout its life, artificial intelligence is forced to replace its knowledge with each new learning.
However, researchers found this outa form of “artificial sleep” could help prevent this phenomenon from occurring.
As Sleeping helps us cement ourselves What we have learned throughout the day, AI could learn and remember how to perform multiple tasks by mimicking this biological behavior.
According to Maxim Bazhenov of the University of California, San Diego, “There is a big trend for this right now.” Bring ideas from neuroscience and biology to improve machine learning, and sleep is one of them.”
Therefore, together with a team of researchers, he… trained a neural network to learn two tasks differently, without having to overwrite the connections learned from the first task.
To accomplish this feat, they have interrupted by targeted training phases after rest-like periods.
In order to To simulate sleep, they activated artificial neurons of the network in a noisy pattern. They also made sure that this noise matched the patterns of the neurons during training sessions.
It is a way to reflect and strengthen connections learned from both tasks. A mechanism reminiscent of human dreams…
The network was trained on the first task, then on the second, then plunged into a phase of sleep. However, the scientists soon discovered that this order overwrote the connections learned from the first task.
The following experiments showed that this was the caseIt is not important to alternate training sessions quickly and sleep while learning the second task. This helps strengthen connections related to the first task.
Ultimately, this method proved effective in training an AI agent using two different patterns Look for simulated food leftovers and avoiding toxic particles.
The goal? Allow the AI Combine different experiences intelligently and apply what they have learned to new situations, such as people and animals…
AI reveals the secrets of the human brain
Artificial intelligence systems are directly inspired by the human brain. A A neural network consists of millions of processing nodes Helps AI learn when fed data.
THE Transformer type networks, invented by Google in 2017 and at the origin of the AI revolution like ChatGPT, get even closer to the human brain by trying to predict the next words in a sentence to find an appropriate answer.
It is The concept of “attention”, comparable to the way your brain works Complete a sentence when you only hear part of the words. That's why modern generative AIs are able to answer your questions.
However, AI can also unlock the secret of how the brain works…and the keys to improvement.
Such is the goal of a research project at Columbia University in New York in May 2023, funded with $20 million from the United States National Science Foundation.
According to Richard Zemel, a professor of computer science at Columbia University, that's the goal bring together the best researchers in AI and neuroscience for the greatest possible benefit of machines and people.
Teams will specifically try to do this understand the concept learn flexibly robust ». In fact, most current AI can only handle one specific task, while the human brain is much more adaptable.
Despite it, AIs are very talented develop language skills. With just one or two examples, they learn much faster than humans. And this ability could help us better understand how to train the brain more effectively.
Also, Continuous learning is about how humans and AI can forget and remember information. Third area of study: the uncertainty principle.
Many AIs are not very good at knowledge whether they are confident about the information or should remain uncertain. A mistake they share with many people…
This research could lead to that the emergence of better brain-machine interfacesin particular to enable the development of AI-assisted prostheses to make it easier for people with disabilities to get around.
Artificial intelligence can learn through imitation
As you probably know, human intelligence depends heavily on it Acquiring knowledge from other people. This has allowed us to accumulate knowledge across generations and cultural evolution.
The Transference is related to our ability to imitate the actions and behaviors of those around us in real time. But in December 2023, DeepMind researchers discovered that AI can also learn in this way.
Previously, Imitation learning or learning through imitation It took many examples and huge amounts of data to successfully copy a human trainer.
DeepMind's new approach enables AI agents imitate body movements in real time without using pre-collected data.
At within a virtual world called GoalCycle3DThe researchers used AI agents that already knew how to navigate between the obstacles in this simulation to reach the goal.
Then, other agents without knowledge were deployed and quickly realized that the best way to achieve their goal was to follow and imitate the experts.
Scientists have recognized this AI learned faster through imitationand that she was able to apply the knowledge thus acquired to other unknown virtual paths. Furthermore, agents have proven that they can use these new skills even in the absence of experts.
According to you it is a authentic example of social learning humane. This discovery could play an important role in the development of general AI…
In addition to Reduce the amount of resources required For training artificial intelligence, this approach could enable AIAcquire the social and cultural elements of human thought. The full study is published in the journal Nature Communications at this address.
This AI organizes itself to mimic the brain: scientists very surprised
In November 2023, scientists at the University of Cambridge in the United Kingdom succeeded Create an AI capable of self-organizing and use the same tricks as the human brain to complete certain tasks.
The human brain and other complex organs develop under a number of competing constraints and demands. For example ours The neural network needs to be optimized to process information without using too much energy or resources.
That's exactly what it is why the brain is structured to create an efficient system that can function within these physical constraints.
Danyal Akarca from the University of Cambridge, one of the lead authors of the study, explains: “The Biological systems evolve frequently They want to make the most of the energy resources available to them, and the solutions they find are often very elegant and reflect the give and take between the different forces imposed on them.”
Based on this observation, the researchers and their team created with the help of neuroscientist Jascha Achterberg an artificial system with physical limitations imposed to model a simplified version of the brain.
Instead of using real interconnected neurons, they use interconnected computing nodes Everyone has a specific location within a virtual space. The goal ? Simulate how two distant neurons have greater difficulty communicating.
Afterwards he was asked the system to solve a maze. A task that requires multiple inputs and information processing.
This simple restriction forced the AI to produce complicated functions. However, these properties can also be found in biological systems such as our brain.
In the eyes of co-author Duncan Astle, a professor in the Department of Psychiatry at Cambridge, “that tells us something fundamental about reason why our brains are so organized.”
The different characteristics described in the study can be divided into two main categories. In one case, AI has an internal structure similar to the human brain, with connections between neurons similar to ours.
In the second case AI also has internal functions comparable to that of the brain. The signals produced by neurons to send information across connections are similar to the signals in the brain.
Every time, Everything seems to be designed for maximum efficiency in conveying information. The study could therefore contribute to the development of more efficient AI systems.
This is particularly the case when large amounts of information Constant changes must be managed with limited energy resources.
THE Robots used in the real world need brains similar to ours because they will face the same challenges. They must constantly process new information from their sensors while simultaneously controlling their bodies to move in space toward a target using limited electrical energy.
This discovery will not only enable, but could also enable, the development of more efficient neural networks in the field of machine learning better understand the human brain.
These artificial brains make it possibleperform impossible tests on a real brain. It is possible to impose various restrictions on them to check whether the AI is more similar to the brains of certain people.
The team hopes that its AI will be further developed Highlight how limitations exist Certain conditions contribute to differences between human brains, particularly in people with cognitive difficulties or mental disorders.
The researchers behind this work say that they were “very surprised” by the results.. The full study is published in the journal Nature Machine Intelligence at this address!
Brainoware: the first computer with a real brain
Created by scientists at Indiana University Bloomington an organoid neural network (ONN) based on brain organoids connected to microelectrode arrays.
This new type of computer was inspired directly by the brain and was called “Brainoware.” It could allow Overcoming current AI hardware limitations such as energy consumption and heat generation through natural solutions.
The study's lead author, Hongwei Cai, and his colleagues wanted to find a biological way Solve the problem of “reservoir computing” : a unique way to process and learn complex sequential and temporal data.
It enables the extraction of patterns and connections from temporal processes, e.gIt ensures a quick workout while emphasizing energy efficiency. This makes it a viable solution for the environment.
This approach has showed promising results for various applications such as time series forecasting, speech recognition, language modeling and addressing complex nonlinear dynamic systems.
Its unique architecture and training methodology offer an innovative alternative to efficient data management. This makes it a valuable tool for machine learning and AI, especially speech recognition.
Additionally, Brainoware could provide the required complexity and variety Finally allow AI to adapt to the human brain and get closer to it.
This system is not it not yet ready to replace our computers classic because it depends on incubators and cell culture technicians. However, this future technology could be the future of computing…
L'full survey “Brain Organoid Reservoir Computing for Artificial Intelligence” was published in Nature Electronics at this address.
When the architecture of the brain inspires deep learning
A study published in January 2024 by Israel's Bar-Ilan University shook the world of artificial intelligence. It seems to prove that Brain architecture can compete with the performance of deep learning models, despite their shallowness.
This work, published in the journal Physica A, focuses on the fascinating contrast between the simple structure of the brain and the complex, multi-layered architectures of modern AI systems.
When comparing the two, researchers led by Professor Ido Kanter found that the The brain can handle complex tasks despite this flat architecture.
There could be a reason for this revelation big change in the way AI is approached. Likewise, the current design of GPUs, which favors depth at the expense of broad architectures, could be revised.
In reality, the broader and higher architectures represent two complementary mechanisms. In order for the brain's learning mechanisms to be mimicked by AI, GPU properties should be changed to imitate the dynamics of the brain…
AI and Brain: How Similar Are They Really?
As AI continues to evolve, the observed similarities with the brain raise the question: Will it one day be possible to reproduce artificially? the complexity and capabilities of the human brain?
A fundamental difference between the two affects computing power. AI systems like deep learning networks can perform complex calculations at lightning speed and outperform humans in areas such as data analysis, pattern recognition and computation.
The parallelism of AI algorithms gives them enough power to process large quantities data immediately. That is her great strength.
He in turn The human brain functions through a network of interconnected neurons the information is processed sequentially. Its effectiveness lies in the holistic processing of information and the integration of various sensory impressions and emotions.
As for the learning mechanisms, AI relies primarily on data. Large amounts of labeled data are required to train the models, improving their performance across iterations.
It is this supervised learning process that makes this possible Recognize patterns and make predictions. In comparison, our brain's learning process is much more complex.
It includes innate abilities, learning through experience and social interactions. We are special good at learning with little datageneralize our knowledge in different areas and adapt to new situations.
A Another big difference from AI is our consciousness. Current AIs perform certain tasks without an emotional state. Replicating this complex phenomenon currently remains a real challenge.
Furthermore, the The brain only uses about 20 watts Energy for operation. This is a real advantage over AI, which consumes enormous amounts of electricity and computing power.
However, as we have highlighted in this file, The researchers are gradually managing to close the gap that separates human and artificial intelligence. In an increasingly distant future, an AI capable of competing with humans on all counts could see the light of day…
The question inevitably arises: Are we actually robots? Does it come from technology far superior to ours? Are we currently there? Repeat the process little by little Who led to the development of this technology? And especially, Who is really the original creator?…This Questions reflect the fascinating theory of the simulatorWe invite you to discover it in the video below.