1. How a neural network actually learns
Training is the heart of the matter. A neural network starts with initial parameters, often random or semi-random. It is then presented with data, and its performance is evaluated.
In supervised learning, an error signal is used. For example, in image classification:
input: photo of a cat
expected output: "cat"
model output: "dog" with high confidenceThe model computes a loss function, which measures how wrong its prediction is. Then, through gradient backpropagation, it determines how to adjust each parameter to reduce that error [1].
In large language models, the most common form of training is self-supervised. There is no need to label every sentence by hand. The text itself serves as supervision: the model tries to predict the next word or token.
For example:
partial input: "The capital of Spain is"
goal: predict "Madrid"The model adjusts its weights so that, on average, it predicts the training data better.
Additional stages can then be applied:
- supervised fine-tuning with examples of instructions and responses,
- alignment through human preferences [2],
- reinforcement learning,
- safety filters,
- style and format adjustments.
None of this turns the model into a conscious being. It only modifies its parameters so that it behaves in a more useful, coherent and safe way according to human criteria.
2. Beyond text: vision, images, audio and multimodality
Although large language models have grabbed the spotlight, artificial neural networks are used in many domains.
In computer vision, convolutional neural networks (CNNs) were the dominant architecture for years [3]. They capture local patterns in images: edges, textures, shapes, objects. Vision Transformers are also used today; they apply attention mechanisms to images split into patches [4].
In image generation, many models use diffusion architectures. They do not generate pixel by pixel the way text is generated; instead, they learn to transform random noise into a coherent image through successive denoising steps [5]. The text acts as a guiding condition.
In audio, neural networks can transcribe speech, synthesize voice, separate sound sources or generate music. In multimodal systems, different encoders convert images, text, audio or video into compatible numerical representations, allowing a single model to reason across several modalities at once.
This shows an important idea: neural networks are not just “text generators”. They are universal pattern-approximation machines, as long as there is enough data and a suitable representation of the problem.
3. Biological neurons: what they are and how they work
Now let us compare all of this with the brain.
A biological neuron is a living cell specialized in transmitting electrochemical signals. It has dendrites, which receive signals from other neurons; a soma, which integrates those signals; and an axon, which transmits impulses to other cells. Communication takes place at junctions called synapses, where neurotransmitters are usually released.
Unlike an artificial neuron, a biological neuron does not compute a simple weighted sum with a fixed activation function. Its behavior is far more complex:
- it integrates signals over time,
- it has ionic dynamics,
- it can show synaptic plasticity [6],
- it receives global chemical modulation,
- it interacts with glial cells,
- it is influenced by hormones, sleep, metabolism, emotions and bodily state.
Biological learning does not work like gradient descent as it is implemented on GPUs. Synapses change their strength according to activity patterns, local rules, neuromodulators and molecular processes. The brain does not train on a centralized dataset, nor does it compute exact gradients of a global loss function, although whether it approximates something similar by other means is an open question in neuroscience [7].
Moreover, the brain uses very little energy compared with a data center. A human brain at rest is estimated to use about 20 watts [8]. Training a large model can consume more than one gigawatt-hour: the training of GPT-3 was estimated at about 1,287 MWh [9], an amount of energy that a brain would take more than 7,000 years to use up.
4. Main differences between an artificial neural network and the brain
We can summarize the differences at several levels.
Physical substrate
An artificial neural network runs on transistors, memory and processing units. A biological network runs on living cells, ion channels, synaptic vesicles and organic chemistry.
Type of signal
Artificial neurons handle continuous or discretized numbers in digital representations. Biological neurons use electrochemical pulses, action potentials and graded signals in different parts of the cell.
Way of learning
Artificial networks usually learn by adjusting weights through gradients and backpropagation. The brain learns through synaptic plasticity, structural reorganization, consolidation during sleep, neurochemical modulation and embodied experience.
Memory
In an artificial neural network, learned information ends up distributed across parameters. There is no native “episodic memory” like the human one, unless external retrieval systems are added. The brain has multiple memory systems: episodic, semantic, procedural, working, emotional, and so on.
Body and environment
The human brain is integrated into a body. Its internal states (hunger, pain, fatigue, emotion) influence cognition. Current models have no biological body and no needs of their own. They operate as mathematical functions triggered by external inputs.
Consciousness
Here we enter open territory. There is no evidence that current models are conscious [10]. They can simulate language about emotions, intentions or mental states, but that does not mean they experience them. The human brain, by contrast, seems to be associated with subjective experience, although the exact nature of consciousness remains a great scientific mystery.
5. So why do AIs seem intelligent?
Because intelligence is not binary. There are many ways to perform well on cognitive tasks.
A language model does not need to be conscious to write a poem, summarize an article, translate a text or generate working code. It only needs to capture statistical regularities rich enough to create based on what it has learned.
Scale matters enormously. When a model has enough parameters, enough data and enough compute, capabilities appear that were not explicitly programmed. This is sometimes called emergence [11], although the term can be misleading [12]. It is not magic; it is the result of optimizing predictions over complex distributions.
For example, a small model may be grammatically clumsy. A very large one can stay coherent over long texts, follow instructions, adapt its tone, detect basic irony or solve problems step by step. Not because it “understands” the way a person does, but because its internal representations have captured extremely deep patterns of language and of human reasoning as expressed in text.
6. Simulated understanding versus human understanding
A useful distinction is between competent symbol manipulation and embodied understanding.
When a person says “fire burns”, they are not just associating words. They have sensory experience, emotional memory, bodily knowledge of pain and a causal understanding of the physical world.
When a model generates the sentence “fire burns”, it is producing a highly probable sequence given the context. It has never felt heat. It has never avoided getting burned. It has no fear and no desire. It does not have a model of the world anchored in direct perception, even if it may have learned remarkable correlations from text.
This does not mean the model is useless or merely a stochastic parrot [13]. It is much more than that. It has learned abstract structures that allow it to generalize. But it is not a human cognitive subject either.
The philosophical question of whether any artificial system could truly understand remains open. What we do know is that current models do not work like brains, even though they can imitate certain products of human thought remarkably well.
7. A final analogy: musician, map and orchestrator
To close, we can use a few analogies.
A language model is not a library that looks up books. It is more like a musician who has listened to millions of pieces and now improvises a new one. It does not copy an exact song; it generates something consistent with the patterns it has internalized.
Nor is it a dictionary where each word has a number-meaning. It is more like a map with an enormous number of dimensions, where words are coordinates and their spatial relationships encode regularities of use.
And it is not a brain in miniature. It is more like a mathematical orchestrator: it receives numerical inputs, transforms them through layers of attention and computation, and produces probabilistic outputs that, assembled token by token, form language.
References
[1] Rumelhart, D. E., Hinton, G. E. and Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323, 533–536. https://doi.org/10.1038/323533a0
Historical source. Popularized backpropagation in multilayer networks and showed that hidden layers learn useful representations.
[2] Ouyang, L. et al. (2022). Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35 (NeurIPS 2022). https://doi.org/10.48550/arXiv.2203.02155
Reliable. Describes supervised fine-tuning followed by reinforcement learning from human preferences (RLHF).
[3] LeCun, Y., Bottou, L., Bengio, Y. and Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278–2324. https://doi.org/10.1109/5.726791
Reliable. Landmark work on convolutional networks (LeNet-5), used commercially to read checks.
[4] Dosovitskiy, A. et al. (2021). An image is worth 16×16 words: Transformers for image recognition at scale. ICLR 2021. https://doi.org/10.48550/arXiv.2010.11929
Reliable. Introduces the Vision Transformer, which treats image patches as tokens.
[5] Ho, J., Jain, A. and Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 33 (NeurIPS 2020). https://doi.org/10.48550/arXiv.2006.11239
Reliable. The basis of today’s diffusion image generators.
[6] Bliss, T. V. P. and Lømo, T. (1973). Long-lasting potentiation of synaptic transmission in the dentate area of the anaesthetized rabbit following stimulation of the perforant path. Journal of Physiology, 232(2), 331–356. https://doi.org/10.1113/jphysiol.1973.sp010273
Reliable. First description of long-term potentiation (LTP), the reference synaptic model of learning. Study in rabbits.
[7] Lillicrap, T. P., Santoro, A., Marris, L., Akerman, C. J. and Hinton, G. (2020). Backpropagation and the brain. Nature Reviews Neuroscience, 21, 335–346. https://doi.org/10.1038/s41583-020-0277-3
Theoretical framework. Perspective article: proposes that the cortex might approximate the principles of backpropagation through its feedback connections. A hypothesis without direct experimental confirmation.
[8] Raichle, M. E. and Gusnard, D. A. (2002). Appraising the brain’s energy budget. PNAS, 99(16), 10237–10239. https://doi.org/10.1073/pnas.172399499
Reliable. The brain accounts for about 2% of body mass and 20% of resting energy consumption. The figure of about 20 W comes from applying that percentage to an adult’s basal metabolic rate.
[9] Patterson, D. et al. (2021). Carbon emissions and large neural network training. arXiv:2104.10350. https://doi.org/10.48550/arXiv.2104.10350
With caveats. Preprint by Google and the University of California, Berkeley. The GPT-3 figure (1,287 MWh) is an estimate based on hardware and performance data provided by OpenAI, not a direct measurement.
[10] Butlin, P., Long, R. et al. (2023). Consciousness in artificial intelligence: Insights from the science of consciousness. arXiv:2308.08708. https://doi.org/10.48550/arXiv.2308.08708
Theoretical framework. Report by 19 researchers in neuroscience, AI and philosophy that assesses AI systems against the main scientific theories of consciousness. It assumes computational functionalism, which not everyone accepts.
[11] Wei, J. et al. (2022). Emergent abilities of large language models. Transactions on Machine Learning Research. https://doi.org/10.48550/arXiv.2206.07682
With caveats. Defines emergent abilities as those that appear in large models but not in small ones. Its interpretation depends on the metric used (see [12]).
[12] Schaeffer, R., Miranda, B. and Koyejo, S. (2023). Are emergent abilities of large language models a mirage? Advances in Neural Information Processing Systems 36 (NeurIPS 2023). https://doi.org/10.48550/arXiv.2304.15004
Reliable. Shows that many apparent jumps disappear when continuous metrics are used instead of all-or-nothing metrics.
[13] Bender, E. M., Gebru, T., McMillan-Major, A. and Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of FAccT ’21, 610–623. https://doi.org/10.1145/3442188.3445922
Theoretical framework. Peer-reviewed position paper that coined the term “stochastic parrot”. It argues about risks and limits, without experimental data of its own.