How We Learn

How We Learn by Stanislas Dehaene delves into the science of the brain and how its function allows people to learn early in life and later within structured education. It was reminiscent of Uncommon Sense Teaching except with a stronger focus on the science behind learning opposed to teaching practices. This makes sense with Dehaene’s background in neuroscience.

Dehaene starts by going through different types of machine learning and detailing their strengths and weaknesses. Neuroscience has had a great impact on how learning models are made, with researchers attempting to get models to act more similar to the human brain. Humans beat out machines in efficiency when it comes to learning. While machines need extensive data to recognize an object or even a single character, the human brain needs remarkably fewer resources and can sometimes learn something in just one interaction. Humans can also explain what they’ve learned and how they came to that understanding. AI is notoriously a black box that’s reasoning behind its output cannot be understood by even itself.

Previously, it was believed that babies possess no knowledge and that their brains are a blank slate until they come into the world and are exposed to stimuli that they can learn from. More recent advances have debunked this misconception by finding that babies seem to know a lot more about the world than people thought. There is even evidence that, in the womb, babies are already ready to recognize faces. Other innate knowledge includes understanding objects, basic number sense, probability, and awareness of a difference between animate and inanimate objects. The brain has a template that humans have adapted over time that self-organizes itself, starting in the womb, and continues to refine and enrich itself over time with experience.

Babies are scientists that test hypotheses that they adapt the results of to form their world model. Dehaene explains that the idea that babies did not develop object permanence until later in life is poppycock. In the experiment used to develop this claim, babies were ‘trained’ to expect an object in a certain place. Even though they see the object go somewhere else they appear to expect it in that same place. The reason for this is because they have not developed higher level functions that would catch this error before they make it. Their hardwired ability to learn is evolutionary advantage of developing a natural pedagogy.

Like in Uncommon Sense Teaching the difference types of memories are explained, this time with slightly different terminology.

• Working Memory- Active memory, what is being currently processed.

• Episodic Memory- Hippocampus stored memory of recent events.

• Semantic Memory- Known knowledge/memorization. Develops while sleeping.

• Procedural Memory- Implicit/Unconscious knowledge.

Neuroplasticity can achieve remarkable things like remapping the brain of a person who underwent a hemispherectomy at a young age allowing them to seemingly retain all their faculties, but it does have limitations, especially as people age. The incredibly efficient human genome that has the blueprint for our brain has limits to what can be restructured. Babies start to learn phonemes of the language around them before they are a year old. After these phonemes are set, it becomes practically impossible to change them. This is why people who speak Japanese are commonly mixing up the R and L sound which in their language have no distinction. When babies and children and not able to properly build these systems when they are young, the negative effects can be life long and vary in severity.

Dehaene’s hypothesis of neural recycling states that neural circuits that were previously used for something else in our evolutionary history are reused for new yet similar cultural functions. The area in the brain that is used for basic number sense like approximate values early in life is recycled for mathematics later in life and is even used for very advanced math in the brains of leading mathematicians. The part of the brain that activates when a sighted person reads is the same part that activates when a blind person reads braille.

What is learning? Developing an internal model of the external world. There are four pillars of learning. Focused attention, active engagement, error feedback, and consolidation. The ability to learn and develop intelligence is determined significantly by environmental factors like socio-economic status and quality of education.

Attention is the brain’s ability to select information, amplify it, and channel to deepen an understanding. There are three systems of attention- alerting, orienting, and executive attention. The first two relate to external stimuli. A person must be aware that there is something to direct their attention to and then direct it. Executive attention is the internal call for available mental operations to process that stimulus.

Active engagement is being engaged with what a person’s attention is directed toward. Engagement is in the head, not the body. Simply being engaged, like doing a physical activity unrelated to what attention should be direct at is not advantageous. When interacting with material, there should be structure and guidance for higher levels of understanding. Discovery learning, forcing students to find the conclusion themselves, is ineffective engagement and makes it harder to learn concepts. Contrary to popular belief, people do not have learning styles, i.e. visual, auditory, or hands-on, because everyone learns functionally the same way. There are only best practices.

Error is the surprise from the difference between expectation and reality. This is used to correct assumptions and learn from. People adjust from errors the same way AI will by adjusting their processes to better fit their past experiences and future expectations. If there is no error feedback, then nothing needs to be corrected. Testing is a very effective method of error feedback because a person can see right away what they do and don’t know. Even when a person guesses an answer, and sees that they guessed it correctly, they learn from that guess because the feedback will reaffirm their initial suspicion. The opposite is also true if they guess wrong. Testing is more effective than just studying or reading material and is more effective over time when intervals between tests gradually get wider instead of being uniform.

The last pillar is consolidation which occurs when people sleep. Information is transferred from episodic memory to semantic and procedural memory as the brain replays the previous day while sleeping. Ideas are recoded in the head and even new thoughts can occur in sleep.

Another important characteristic of learning that ties into the first three pillars is curiosity. Curiosity is the gap between what we know and what we want to know. It can be inhibited by teaching that is only catered to the receptive mode over the active mode. It is not enough for babies to see people interacting with objects, they need to form their own hypothesis and test them themselves. The same is true for students learning at a higher level. Failure, stagnation, and punishment also kill curiosity over time and convince those capable of learning that they are not.

How We Learn was a very interesting and informative read. I liked getting more details on some of the more neuroscientific aspects of Uncommon Sense Teaching. I was not a big fan of the color insert because while the information was valuable I did not like needing to flip back and forth to look at the pictures every time they were referenced. Stanislas Dehaene takes a very scientific approach to learning and how current educational shortcomings can be corrected by following the data. Unlike more emotional literature I’ve read about education, I feel like the information in this book is extremely practical and can be employed quickly by anyone no matter their resources. This should be required reading for anyone whose job involves educating other people whether it be as a teacher, personal trainer, or a professional coach.

Some information I couldn’t find a place for in my report was the fact that people map out the physical world in their brain. Certain parts of the brain activate in locations and when the events of the previous day repeat in sleep those areas will activate when going through the parts of the day spent in those areas.

Also, AI models can fail to improve if they get stuck on a local maximum where when they adjust parameters either way performance decreases and this causes them to be unable to find a global maximum of performance because they don’t want to become worse. Implementing some randomness can help solve this problem by getting a model to try a more than gradual change to determine if there is a better set of parameters out there.