Jump to content

Session 5 summary and highlights: Panel Discussion for Future Directions of Machine Learning in Neuroscience


Recommended Posts

Shraddha Shah

The panelists for the session on discussing the future directions of Machine Learning in Neuroscience were: Andrew Saxe (A), Kim Stachenfeld (K), and Leila Wehbe (L). The session was moderated by Kristin Branson (KB). Here are some of the important takeaways from the discussion session:

  • The current state of machine learning (ML) and artificial intelligence (AI) algorithms has more predictive power but lower interpretability. The interactions between ML and neuroscience can help to improve our understanding of how to interpret ML algorithms in addition to being using for neuroscience data analysis and modeling.
  • As a neuroscience PhD student starting into exploring ML for their data, start with the more robust supervised methods first, especially in cases where the ground truth data is available. In relying on ML methods to find structure in the data where the ground truth is less obvious, it is important to have a good understanding of the fundamentals of the algorithms used, and what assumptions and biases are baked into them. This can aid robustness of the results and interpretability of the outcomes.
  • One of the key distinctions between how ML/AI agents learn and biological systems is the ability to perform “continual learning”, some avenues to ameliorate this in ML/AI agents were discussed, and this is one area where ML systems can learn from biological systems that are able to perform continual learning in the real world.
  • ML systems face a challenge wherein they lack access to rich, complex data which can be a bit expensive to obtain. Neuroscience provides a rich, large-scale dataset that can help develop better ML approaches to perform exploration and active inference which is currently hard to do.

Overall, the consensus of the panel was optimistic in that there are several avenues where neuroscience and ML communities could work together in, and mutually benefit. It was illuminating to understand that the crosstalk between the fields would end up being quite productive for both fields, and that the rich datasets from neuroscience combined with the more complex (even if less interpretable) models in ML can interact in highly informative ways.

An edited and condensed version of the questions and discussions is provided here (this is not transcribed verbatim):

  1. What are your thoughts on the use of AI to increase replicability in neuroscience across labs?

L: One of the most important ways to increase replicability is at the community level by sharing data, code, and giving a perfect description of what’s been done. Some statistical and ML approaches can help to ensure you don’t make too many conclusions from the algorithms you are using. It also helps to look at data from other labs, across subjects and conditions to verify the robustness of the results. It is also important to focus why some algorithms don’t work: if out of the 10 algorithms you tried only 2 gave a significant result, think about why the rest didn’t work.

  1. Can deep learning be used for medical applications such as seizure detection, developing biomarkers, and clinical trials?

L: Deep learning can be very valuable to understand changes in brain structure across different cognitive impairments and subjects, so mainly in clinical image analysis. It’s important to remember that while deep learning is a powerful tool that is good at making predictions of the outcome, we still need to be careful in interpreting why an outcome was predicted. Since for deep neural nets, interpretability is hard, it is important not to jump to conclusions on why an outcome was predicted.

  1. Where can neuroscience make biggest impact on ML?

A: At a high level, I think, neuroscience is good at making ML people think about the sets of problems they might not be considering at the moment, so it raises the bar for the work in ML. Two examples of these are:
(i) Theory of Mind problems: how agents interpret and interact with each other by making causal inferences about those interactions
(ii) Continual Learning problems: Biological systems are good at being able to learn online from their environment for continuous periods of time, where AI systems tend to forget much faster and thus are not adept at continual learning.

  1. What are the possible pitfalls on how we use ML in neuroscience?

L: There are many possible pitfalls, but there are some we can prevent. An important pitfall we can prevent is that even though ML algorithms are good at making predictions, they are not highly interpretable. So, if an algorithm is good at predicting an outcome you are interested in, and seems to be using a particular feature to make this prediction, it is not necessary that this feature is necessary for the behaviour or phenomenon of interest. It is important to remember that the link between these complex ML models and complex behaviours leading to prediction of specific brain areas is not very straightforward. Another pitfall that may be hard to prevent that perhaps the representations we are trying to fit in some cases are so complex, that the ML algorithms might not have that kind of resolution, especially for humans where the data available are at the resolution of fMRI, MEG etc.

A: I would like to reiterate the issue of interpretability. It is important to remember that we don’t understand the deep neural nets that seem to behave like the brain, but that makes it all the more important that we need more insight into how this complex system (deep nets) work. Predicting neural activity is a great first step, but getting to this kind of insight will be crucial.

  1. As a neuroscientist, how to learn ML? What are the parts of ML that are worth getting into?

A: It really depends on what you are interested in. Broadly, there are two paths one can take: use of ML for data analysis, and use of ML models to test new theoretical ideas about how brains work i.e. using ML as models of the brain. For using ML for data analysis, it is important to develop a good understanding of the fundamentals of the algorithms. For using ML as models of the brain, deep learning is an obvious first place to start, and a good one too. Start with the Parallel Distributed Processing books by Rumelhart et al. There are good Coursera courses on the topic (by Andrew Ng), a book on Deep Learning by Ian Goodfellow that is pretty good. Try to focus on time-tested ideas and not necessarily the latest and what might seem the greatest.

KB: It is a good idea to focus on supervised ML approaches that can help automate and improve some analyses processes, as most of these methods are well-developed and are pretty robust. These applications also allow to compare with the ground truth data more easily. For using ML as models of the brain, a deeper understanding of ML and more expertise is required.

  1. What ML algorithms have concretely contributed to neuroscience and vice versa?

K: One of the most important contributions – the idea of interpretation of dopamine signal as a reward prediction error, a model that has led to a lot of explanatory power in neuroscience. In general, Reinforcement Learning has had a lot of explanatory power in neuroscience. Another interesting contribution is in the domain of big data coming out in neuroscience where ML algorithms provides several exploratory methods to analyze the data.

A: The two fields of ML and neuroscience have been intertwined from the start, since the inception of perceptrons and neural nets. It is incredible that deep neural nets are doing well and at least they are vaguely like parts of the brain, that didn’t have to be the case. One could imagine that intelligent systems and computers look nothing like the brain, and yet there has been a strong overlap. The correspondence between ML and brains is not one-to-one, but in the future hopefully the interplay of neural features and ML systems will be stronger.

  1. What is the role of simulated and generated data for neuroscience?

A: The main role of simulated or generated data is to help improve interpretability. By simulating simple toy models and using clean, simple datasets, one can gain a better understanding how these complex systems do a task, and we can then compare it to the full nonlinear system and see if it behaves similarly. As a simple example, data generated from hierarchically structured generative model fed into a deep net shows that the deep net learns the structure of the data by starting at the broadest distinction and then branching into finer distinctions.
L: One can think of simulated data as coming from a theoretical model or as another type of data – use this as to build models of the brain and use the model to think of new cases the brain might encounter. One can further test the model on previously published results to see how well the model generalizes or what features of the model generalize.

  1. What ML algorithms and approaches are best for high-dimensional but low sample data?

K: This is a fundamentally hard problem, especially for using deep learning methods. So best thing to do is to not fit too many parameters when the number of parameters far exceeds number of data samples. Another way to combat is to augment lots of datasets. Otherwise, stick to traditional, linear methods instead like Support Vector Machines, Principal Components Analysis etc.

KB: One can also simulate the kind of noise or predicted transformations on the training data and thus artificially augment the training data.

K: Other ways to constrain data – using structured hypotheses about neural data, adding inductive biases, regularizing etc.

A: It is important to remember that even Deep Nets are operating in this regime, in that the number of parameters of Deep Nets can be much higher than training examples. Thus, this is a fundamental problem even for deep nets. Performance in this regime can thus depend quite strongly on the choice of the prior.

KB: That’s the surprising thing that even then deep nets seems to generalize well, something we want to understand.

  1. Kim mentioned explorations is one direction for ML in neuroscience? What are your thoughts? This seems different from how animals behave, where we see animals balancing the exploration-exploitation dichotomy. Whereas, current ML only have exploration built into them. What would enable to use the learnings from animal behaviours literature to inspire ML algorithms in this domain?

K: A lot of data I would be interested in is how animals and humans behave in new environments even without the neural underpinnings. Mice and rats do clever things, if you move the reward from one end of the maze, they go to the other end directly, they aren’t bumping into walls etc. or randomly running around. Thus, better characterization of behaviour is important. This is a bit tricky as it is hard to get a lot of data for behaviour in “new” environments, as its hard to get too many trials of this. The neural data that would help is to find neural correlates of how the brain uses the structure we’ve seen before to decide where we think we might find more information.

  1. What is your advice for neuroscience PhD students interested in ML approaches?

A: Be ambitious, find ways to move beyond just incremental work. PhD is a good time to try something big and ambitious, but also needs to be doable. Do keep that in mind.

K: Just get started with the methods, code up some algorithms and mess with the architecture and parameters etc., there is a lot of low-hanging empirical fruit that just isn’t tried out. Try to have 1-2 safety nets in terms of projects, but then try something more ambitious along with it. Computational projects are not as intensive to be able to get started compared to some other more involved experimental methods.

L: Don’t think that you are not a particular kind of person. Computational work can be learned, just try it and don’t wait for the perfect opportunity.

  1. According to you, what could be the biggest possible achievement of ML-neuroscience research in the next 10 years?

L: To be able to do joint-modeling of specific areas of AI like language, vision, reinforcement learning and neuroscience of those fields. This would enable to learn principles of brains and also to design better deep learning algorithms. This would also help to understand if the reason why these deep learning algorithms work has anything to do with brains.

K: One of the challenges in ML is trying to draw conclusions in domains where data is expensive whereas neuroscience has lot of rich, large-scale and higher-dimensional datasets. This can help to explore ML approaches of active inference and exploration on the neural data, as such approaches are harder in core ML as they depend on expensive and rare types of data.

A: If we can remove the M from ML, and be able to develop a better understanding of the learning and understand whether this kind of learning can also describe part of the learning processes in biological systems – that would be very interesting and a very powerful tool for understanding the brain.

KB: An important avenue is to develop a better understanding of making better use of ML to find possible structures in neural data.

  1. Sleep is important in animals for learning. Is there anything equivalent in ML where shift of states enables better learning?

A: This hasn’t been proven yet, but one idea is that in sleep, experiences or memories stored in the hippocampus is replayed into the neocortex – this is an idea with a long history. The modern version of that is the complementary learning hypothesis. Replaying episodes from hippocampus into neocortex can be seen as one solution to the continual learning problem, as this allows slowly integrating new knowledge without quickly overwriting old knowledge. Another idea is a probabilistic learning model where during the day, the model can see what exists in the world and what can happen, but to do proper probabilistic inference one needs to downweigh, at night, what can’t happen and thus forget those unlikely scenarios. These learnings could be potentially incorporated in ML approaches.

K: Replay buffers are important for Reinforcement Learning. In continual learning where agents are interacting in the real world, a lot of sample data can be very correlated (as seen in natural environments). Mixing in data from memory can be helpful to prevent catastrophic forgetting and promote continual learning. The cyclic nature of sleep has parallels with the cyclic periods of learning processes – such as the forward and backward pass in neural networks. Taking in information and using that information to update parameters is useful – Anna Schapiro has some conceptually interesting work on this topic comparing oscillatory nature of learning and oscillatory nature of human brain states.

  1. Is there any particular area of ML that is up and coming and you are excited about its impact on neuroscience?

L: Excited in the area of language – where we can take a sequence of words and this can be represented as a vector. Instead of just looking at which areas respond more to complex sentences vs. simple sentences, this allows us to see which areas are involved in complex meaning beyond just words. This opens the door for using natural language and communicative language in experiments.

A: The area that I am excited is about all the ML approaches in causal learning, this probably has lots of connection to neuroscience.

K: There are two areas in ML that I am very excited about:
(i) Development of neural networks that are capable of relational reasoning i.e. performing computations over complex and rich pairwise information. That would have the capacity to model rich human behaviour and animal processes.
(ii) The second area is called meta-learning. The power these approaches gives to neuroscience is they help constrain the models by using a distribution over tasks which is probably a more natural way to constrain models on how an animal should be able to do a set of behaviours.

---- end of discussion session ----

Here are some links to explore this topic further:

Q&A with Jonathan Pillow: Machine Learning, Big Data and Neuroscience

Current Opinion in Neurobiology special issue edited by Maneesh Sahani and Jonathan Pillow: Machine Learning, Big Data, and Neuroscience

Hassabis et al, Neuron, July 2017: Neuroscience-Inspired Artificial Intelligence

Link to comment
Share on other sites

Please sign in to comment

You will be able to leave a comment after signing in

 Share

×
×
  • Create New...