David Marr vs. The LLM’s
David Marr died young (35) and died a long time ago (1980). Even so, in the years after his death he had a massive influence on AI. Marr had a pivotal role in establishing what research in AI should look like and what AI researchers should do. He was really important in defining what AI became. He wrote down his ideas on this in a paper you can find here called “Artificial Intelligence — a personal view”. His book (Marr 1982) “On Vision” was and is considered a core text and a classic of science writing. It was the starting point for AI and especially machine vision researchers for many years and is still a recommended text in many university courses (for example MIT : https://www.youtube.com/watch?v=Di_3pGAveGs&ab_channel=MITCBMM).* As another measure of his influence you can find a video of Noam Chomsky talking about his work in an interview filmed in 2012.
So — what did Marr think about AI? In “Artificial Intelligence a Personal View” Marr sets out his criteria for research in AI, this paper is the jumping off point for this blog. Marr conceived the lowest level of description to be computational. This was important because by describing processes that achieve specific effects using computational formalism allowed tangible solid results to be obtained (for example see page 331 of Marr 1982). Marr explains “once a computational theory for a process has been formulated, algorithms to implement it can be designed”. Algorithms are therefore seen to be the second level of Marr’s hierarchy and at the top (or bottom) is an implementation level — the mechanism that executes the algorithm.
Although this hierarchy is quoted as religious law in the literature of computational neuroscience it seems to me that it’s an overly literal interpretation of Marr’s thought. Reading his output beyond “On Vision”, and in particular reading the paper that’s the subject of this blog I think that Marr probably thought that some hierarchy or decomposition was required to make sense of complex systems. The classic computational->algorithmic->implementation was just one specific description that he could articulate.
Tomaso Poggio, was one of David Marr’s collaborators and survived him to carve out a glittering career over a lifetime spent teaching and researching Computer Vision at MIT. . Poggio characterizes Marr’s approach to Artificial Intelligence as requiring the description of a complex system at several different levels of detail (Poggio 2021). Specifically brains and minds should not be understood as singular systems that operate in one way. Instead they are to be conceived of as an arrangement of modules that have particular properties and functions that are achieved using different underlying mechanisms.
Marr’s framework is constructed around specific defensible features. Marr wanted to identify problems that could be abstracted from the detail of the observations of human cognition, solved, and the solutions then used to produce outcomes that are mapped back to the details of human cognitive operation.
This formulation articulates a classic idea of science, the creation of theories from evidence that predict so far unobserved facts. This idea is central to Marr’s program because it placed Artificial Intelligence firmly in the category of Natural Science. Artificial Intelligence was to be a science that studied the way that the human brain worked, in particular it was to be the study of the way that the human brain created intelligence. This study would be conducted by abstracting the details of the implementation (all the messy biology and electronics) away from the processes and algorithms that Marr saw as fundamentally important. AI would be respectable.
In the “Personal View” paper Marr introduces his idea of type 1 and type 2 theories. Type 1 theories result from the isolation of a specific information processing problem (Marr’s examples are linguistic competence & creating a 3d model of a scene from bifocal vision data) and the creation of a method of solving it. Protein structure recognition is called out as an example of a type 2 theory because it is likely to be solved or approached without providing clarity about how the information processing problem is solved. A large complex program without much theory might succeed. This will be impressive if and only the program is effective (and you can show that convincingly).
Marr is clear that a type 2 theory is a good thing if it leads to an effective solution, but a type 1 theory is interesting even if it doesn’t provide the complete mechanics of a solution. By Marr’s lights it’s ok to construct a hierarchy of processes that solve a large problem… even if the precise way to solve the components of the hierarchy is unclear.
This formulation of respectable research explanatory of human cognition or useful in a specific application is what has underpinned respectability in AI since Marr died. Interesting has been defined as biologically or psychologically plausible, useful as solving a specific application. This formulation allowed Patrick Winston to both develop his research developing a computational theory of how people think and at the same time happily acknowledge the value of engineering AI solvers and ML models into systems that solved airline schedules. At the same time the same definition included work from neuroscientists like Geoff Hinton. Now the focus of AI as the science that understands human cognition is inadequate.
The development of a machine that could do protein structure prediction, identified as a respectable type 1 problem by Marr in 1982, finally happened in about 2020. Deepmind revealed Alphafold and then Alphafold2 (Jumper 2021). This development was part of the revolution in AI and ML that has also created GPT3 and ChatGPT and Bard and a host of other huge and amazing models that understand text and chess and molecular biology using the same processes and mechanisms.
It’s evident that these are machines with “new mechanisms of understanding” (Mitchell & Krakauer, Shanahan) and LLM’s that in some way understand text are in the same category. Mitchell and Krakauer further argue that these machines are only the first of a new wave of different understanding mechanisms that will emerge in the future.
Are these machines minds or minded? This is an open debate currently underpinned by belief, just as ideas about but they aren’t “alive” in the sense that they only have short term memories, and they aren’t grounded in that they don’t know what a kitten or a glass of beer are really like. To use biblical language, they aren’t “quick” or “quickened” — the spirit isn’t in them in the way that it is in a mouse or a butterfly.
Does this differentiate their understanding from our understanding? I would say clearly yes it does, their perception is radically different from our perception. A thinker that was never living can’t fear death. It may be that what LLM’s are is literally that — the language element of cognition. The other part — the bit that honks and bites and runs round the plains, or hides in the seaweed, may be more elusive.
Does it mean that a llm’s understanding is inferior or invalid? That’s much less certain, I think so — but it’s arguable that it isn’t or won’t be when the current generation of models is perfected in the coming years or months. Regardless of qualitative judgements it does make them different.
Recent work on the intelligence of cephalopods and corvids has also shone light on cognitive systems which have an evolutionary story that separates them from ours by hundreds of millions of years. Our minds are different from the minds of octopus and crows, how could they not be? After 40 years there is lots of evidence that understanding mechanisms in human cognition is too narrow a focus for modern AI.
Just as it is clear that Crows and Octopus have a distinctive cognitive machinery it is indisputable that the computational mechanism that is used to create understanding in LLM’s is completely disconnected from the computational mechanism used by humans. We don’t operate because of massive arrays of floating point numbers updating in nanoseconds in our heads, LLM’s do. Despite this difference LLM’s produce language (although Crows and Octopus don’t).
So, a third type of work is now respectable. Despite what David Marr used to think, creating a new “structure of understanding”, and then studying it is now science. AI is climbing down the tree of self knowledge, the top branches of logic and mathematics were the first rungs we descended, then we swung down to games and puzzles, now we come to language and symbols. The inner mysteries of motivation, freedom and consciousness may be a long way down, or we may fall past them next week.
I’m reminded of the story of space travel. When I was small I wanted to be an astronaut because it seemed sure that these hero’s would voyage around the solar system in my life. But humans were stopped at the moon, and only sent little robots out to the planets. Despite all that progress no human has traveled beyond luna orbit, no human has traveled to the moon for fifty years and more.
mRNA was discovered about 70 years ago, but mRNA vaccines are new.
My point is that the pace of progress is unpredictable, especially when you don’t know where you are going. AGI may be heralded by LLM’s, but the trumpets may sound for 7, 70 or 700 years before HAL puts in an appearance.
I wonder what David Marr would have made of it all?
Of course I only know him from reading his books, just like Bard and ChatGPT only know kittens & beer from reading about them. Just as it doesn’t stop them from pronouncing on those topics it won’t stop me. I think he’d have just laughed and said “of course”.
References
Jumper, J., Evans, R., Pritzel, A. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).
Marr 1982 https://archive.org/details/vision00davi/page/330/mode/2up
Mitchell, Melanie, and David C. Krakauer. “The debate over understanding in AI’s large language models.” Proceedings of the National Academy of Sciences 120, no. 13 (2023): e2215907120.
Poggio, Tomaso. From Marr’s Vision to the Problem of Human Intelligence. Center for Brains, Minds and Machines (CBMM), 2021. http://34.201.211.163/bitstream/handle/1721.1/131234/CBMM-Memo-118.pdf?sequence=1&isAllowed=y
M. Shanahan. Talking about large language models, 2022. arXiv:2212.03551.
*It’s also obvious that Marr made lasting and significant contributions to Neuroscience, but Neuroscience is something I know very little about, and that part of Marr’s work is beyond the scope of this post.
