Do Deep learning models really ‘get it,’ or do they make statistical predictions? Is it possible for artificial intelligence to reach the level of conceptual understanding? If it is possible, how do you think it will be?
A typical argument I have encountered on why people that believe that deep learning algorithms 'get it' is as follows:
a) Neural networks simulate a biological neural network, therefore they must be doing something right
b) The output is coherent, then they must be doing something right
c) if A and be are conclusions that they are doing something right, then it must be that they 'get it'
Let us take this step by step:
Analysis of a): Just because something is simulating does not grant it equivalence. Also, at what level of simulation is the neural network doing its computation? is it simulating the travel along the axon to the dendrites and doing the simulation of the chemical reactions in the same way as a biological network does? are these details important to truly simulate a neural network? all these questions have to be answered before a) can be sustained as true.
Analysis of b) : the problem with this conclusion (which is itself used as a premise for c) ) lies on an underlying premise that coherent output means understanding such that the AI 'gets it'. The problem here is to define what really is the criteria for understanding. For some, coherent output is evidence enough to show that there is intelligence. Is this the best metric? I could definitely find evidence to the contrary.
I would have to answer your second question as positive. I think that AI can reach conceptual understanding, but not with the current trend in deep learning as the main means of implementing AI. I think that right now the focus is on building subsymbolic systems, but it would be a mixture of symbolic and subsymbolic systems that is required. I think that the current trend in Deep learning is too biased towards Hinton's subsymbolic approach using distributed memories. My best guess is that distributed and local subsymbolic systems are needed to completely integrate them with symbolic systems.
AI today is described in breathless terms as computer algorithms that use silicon incarnations of our organic brains to learn and reason about the world, intelligent superhumans rapidly making their creators obsolete. The reality could not be further from the truth. As deep learning moves from the lab into production use in mission critical fields from medicine to driverless cars, we must recognize its very real limitations as nothing more than a pile of software code and statistics, rather than the learning and thinking intelligences we describe them as.
In my opinion, while current deep learning models do not understand it in a human sense, future advances in AI could potentially lead to systems with some kind of conceptual understanding. However, this would likely require fundamentally new approaches that go beyond today’s statistical methods.
Thank you, Arturo, Shafagat, and Hamidreza, for your valuable insights. The perspectives each of you offers make essential points about the limits and potential of deep learning.
Arturo Geigel , your analytical approach highlights critical points to consider when evaluating the "understanding" capacity of deep learning. In particular, your thinking on integrating subsymbolic and symbolic systems could be the key to reaching a more profound understanding in the future.
Shafagat Mahmudova , your comparative perspective on what AI is and how it is perceived is very important. Your emphasis on the need to position existing systems correctly without ignoring the limitations of deep learning provides a valuable framework for understanding the current state of the technology.
Hamidreza Zahedi , your point that current deep learning models do not have human-like understanding but that this may be possible in the future is a promising perspective on the potential of AI. As you mentioned, new methods beyond today's statistical approaches are needed to get there.
In conclusion, the comments from each of you show that whether AI really "gets it" is a technical issue and involves a philosophical and methodological debate. Thank you for your contributions to this fruitful discussion. I conclude the discussion here, but I hope we will have the opportunity to discuss this topic in more depth with new participants.
Kubilay Ayturan Unless we're going too deep into philosophy or metaphisics, I'd say that for practical purposes it already kind of "get's it".
1) It works. Seriously, for me that is basic reality check concerning whether we face rudimentary understanding: is it able to solve problems where neither memorizing correct answer nor expressing fluffy bs would help, but one would need to understand the problem? Yes, it is able to do so and gets quite good at programming.
2) It already passed good old fashion Turing test which was intended as proxy for thinking. If we need to move goalpost, then we kind of admit that is doing fine.
3) In "chain of thought prompting" instead giving outright answer, it first analyses the problem. It presents its line of reasoning (which generally shows that it understands what's the problem) and uses this reasoning to provide much better answer than the same model asked to give answer straight away. (so demonstrates that this analysis of problem indeed helped thus showing some understanding)
Marcin Piotr Walkowiak I appreciate your thoughtful words. The fact that artificial intelligence is effective in practical applications and demonstrates problem-solving skills may indicate that it comprehends certain concepts, but whether this is the same as human cognition will remain a topic of debate.
Although a passing score on the Turing Test is an important indicator of good performance, its applicability has long been debated. Although 'chain of thought prompting' and other techniques demonstrate that the model can understand it, this process doesn't require self-awareness or consciousness.
As a result, the question of whether AI really ‘gets it’ will continue to be debated as both a philosophical and technical issue.
The question of whether deep learning models "truly understand" is a complex philosophical and technical one, and there's no simple answer. Here's a breakdown of the key considerations:
What Deep Learning Models Do:
Pattern Recognition:Deep learning models excel at recognizing complex patterns in vast datasets. They can identify statistical relationships and correlations that humans might miss.
Prediction and Generation:Based on the patterns they learn, they can make predictions and generate new content, such as text, images, or audio. They can also perform complex tasks, like translating languages or playing games.
What Deep Learning Models May Lack:
Consciousness and Subjectivity:There's no evidence that deep learning models possess consciousness, subjective experience, or self-awareness. They don't have feelings, emotions, or a sense of "self."
True Understanding of Meaning:While LLMs can manipulate language effectively, they don't necessarily understand the underlying meaning of the words they use in the same way humans do. They are really good at predicting what words statistically belong together, and in what context.
Intentionality and Purpose:Deep learning models don't have intentions or purposes of their own. Their behavior is determined by the data they've been trained on and the algorithms they use.
Causal Reasoning:Although advances are being made, deep learning models often struggle with causal reasoning. They can identify correlations, but they may not understand the underlying causes of those correlations.
The "Understanding" Debate:
Statistical vs. Semantic:Some argue that deep learning models are simply sophisticated statistical machines that manipulate symbols without understanding their meaning. Others argue that their ability to generate coherent and contextually relevant outputs suggests a form of understanding, even if it's different from human understanding.
The Black Box Problem:The "black box" nature of deep learning models makes it difficult to determine what they're truly doing internally. This lack of transparency contributes to the ongoing debate about their capabilities.
In summary:
Deep learning models are incredibly powerful tools that can perform complex tasks.
However, there's no evidence that they possess consciousness, subjective experience, or true understanding in the human sense.
The question of whether they "understand" is a subject of ongoing debate and research.
Deep learning models do not truly "understand" in the way humans do—they approximate understanding by detecting patterns and statistical relationships in data. When we say a deep learning model "understands" language, vision, or context, we often mean it performs well on tasks that require understanding (like translation or image classification). However, this success stems from pattern matching and correlation, not conceptual reasoning, intentionality, or awareness.
These models lack a grounded sense of meaning, conscious goals, or an internal model of the world. For example, a model might correctly answer questions about a story without grasping the plot or characters—it’s simply leveraging patterns seen in its training data. Critics like Noam Chomsky argue that this is "imitation without comprehension," while others believe such statistical models may eventually approach something functionally equivalent to understanding.
So, while deep learning can simulate understanding, it is fundamentally syntactic, not semantic—effective but shallow. True understanding, as in grasping meaning, context, and consequences, remains uniquely human—for now.