What has neuroscience done for machine intelligence? (Updated)

Today on the twitters, Michael Hendricks asked, “Why do AI people bother with how animal brains work? Most good inventions work by doing things totally unlike how an animal would.”

The short answer is that animal brains can already solve the problems that AI researchers want to solve; so why not look into how they are accomplishing it?

The long answer is that in the end, the algorithms that we ultimately use may end up being dramatically different – but we need a starting point somewhere. By looking at some of the algorithms that have a neural inspiration, it is clear that by thinking about ideas of how the nervous system works machine learning/AI researchers can come up with clear solutions to their problems:

  1. Neural networks. In the 1940s and 50s, McCulloch, Pitts, and Hebb all contributed to modeling how a nervous system might work. In some sense, neural nets are trapped in this 1940s view of the nervous system; but why not? At an abstract level, it’s close…ish.
  2. Deep learning. Currently the Hot Shit in machine learning, these are like “neural networks 2.0”. Some quick history: traditionally, neural networks were done one layer at a time, with strict feedforward connectivity. One form of recurrent neural network proposed by Hopfield can be used to memorize patterns, or create ‘memories’. A variant on this, proposed by (computational neuroscientist) Terry Sejnowski and Geoff Hinton is the Boltzmann machine. If you combine multiple layers of Boltzmann machines with ideas from biological development, you get Deep Learning (and you publish it in the journal Neural Computation!).
  3. Independent Component Analysis. Although this story is possibly apocryphal, one of the earliest algorithms for computing ICA was developed – by Tony Bell and Terry Sejnowski (again) – by thinking about how neurons maximize their information about the physical world.
  4. Temporal difference learning. To quote from the Scholarpedia page: “This line of research work began with the exploration of Klopf’s 1972 idea of generalized reinforcement which emphasized the importance of sequentiality in a neuronal model of learning”

Additionally, companies like Qualcomm and the Brain Corporation are attempting to use ideas from spiking neural networks to make much more energy efficient devices.

In the other direction, neuroscientists can find that the brain appears to be implementing already-known ML algorithms (see this post on Nicole Rust). Many ideas and many biological specifics will be useless – but research is the hope of finding the tiny fraction of an idea that is useful to a new problem.


Over on reddit, downtownslim offers two more examples:

Neocognitron was the foundation for the ConvNet. Fukushima came up with the model, LeCun figured out how to train it.

Support Vector Machines This last one is quite interesting, not many people outside the neural computation community know that Support Vector machines were influenced by the neural network community. They were originally called Support Vector Networks.

5 thoughts on “What has neuroscience done for machine intelligence? (Updated)

  1. I have mixed feelings on this topic. On the one hand, I can get pretty excited about natural algorithms and think that for engineering problems, it is great to turn to whatever ideas we can take, be they inspired by nature or not. On the other hand, I find that stressing the connections can be counter-productive. A thing that seems to happen eerily often is to tell the biologists/neuroscientists that the work is of interest for mathematical or computational reasons, and to the mathematicians/CSers that the work is of biological or neurological interest; while in reality, the work is not interesting on either side. Sometimes this can grow to a whole field that is not competitive from an engineering perspective, but also not of interest or applicability to the scientists. I am thinking of genetic algorithms here, where they are usually outperformed by simulated annealing algorithms or SAT-solvers in practice, and not very reasonable for biological insights, either.

    Admittedly, at the neuro-ML interface things are a little bit different. A lot of NN work has been pretty well developed, so the engineering is solid, but I feel like it can be damaging to theory. From the perspective of neuro, it seems like the NN work is potential steps towards ‘simplified’ theories of the brain, but I don’t think they are actually useful as theoretical tools for neuro since they can’t offer much other than simulation, and I definitely see them abused in cognitive psych. From the perspective of theory, it seems like the complexity of NN is justified as capturing some part of neuroscience, which makes ML folks more comfortable with the fact that there are much fewer theoretical guarantees.

    Of course, the above might just be my pessimism. A more optimistic interpretation is that NN work motivates the neuroscientists to think more formally and abstractly by offering them steps towards a computational framework, and motivates the theorists to develop new tools to better characterize these beasts.

    • I suppose where I find the connection to be useful isn’t the exact placement of ion-channels in complex models etc etc, but in taking ideas about how the brain functions as inspiration for solving engineering problems. Sometimes it may lead to dead ends (eg genetic algorithms. Are those a dead end? I hadn’t realized it, or thought about it). Others may be fruitful. I certainly wouldn’t think it’s something that should be a *focus* of most ML research, but it seems useful as *one of many* ways of investigating the problem.

      ANNs are useful, sometimes, in theoretical understandings of the brain. Othertimes, not so much. But I don’t think we should get hung up on insisting on a close connection between the two fields all the time.

  2. I would agree that generally it’s not obvious that looking to neuroscience to solve engineering problems is a viable approach: different hardware may require different solutions. Though of course it can’t hurt to take inspiration from something you know works.

    When it comes to deep nets for visual object recognition (specifically the convolutional nets that have been winning ImageNet), I think the optimal approach there happens to be incredibly similar to biology, and as a computational neuroscientist who studies vision I find that incredibly lucky. Deep nets provide a framework to create a simulation of the visual system that could actually perform a task from start to finish. Having such a simulation means we could test out theories of computation and perform specific manipulations and see their effect without an animal staring at a screen.
    Granted deep nets as they are will need to be tweaked to have them merge with models of visual cortex before this is actually all viable, but still, it is exciting.

    • “Granted deep nets as they are will need to be tweaked to have them merge with models of visual cortex before this is actually all viable, but still, it is exciting.”

      Yeah this is what I find most exciting – the back and forth between the two fields that brings people together who are thinking about similarish issues but from dramatically different perspectives.

  3. Pingback: Proxem » La lettre du 6 octobre : les enjeux de la rencontre entre la culture des humanités et celle de l’ingénierie informatique

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s