The Roles of Supervised Machine Learning in Systems Neuroscience
by Joshua I. Glaser, Ari S. Benjamin, Roozbeh Farhoodi, Konrad P. Kording
システム神経学（Systems Neuroscience）でどのようにMLが使われているかのSurvey
- Solving engineering problems
MLは医学的診断やBMI、その他研究のためのツールの機能を向上させた。
セグメンテーションの例：
segmenting and labeling parts of an image, which usually requires manual annotation, is an especially important task. However, as imaging techniques improve and the volume of data increases, it will become infeasible to rely on manual annotation. To solve this problem, many ML techniques have been developed to automatically segment or label new images based on a dataset of previously labeled images. These techniques have been used to label medical images (Litjens et al. 2017; Ronneberger, Fischer, and Brox 2015) . They have also been used to understand the connections and morphologies of neurons from electron microscopy (Helmstaedter et al. 2013; Funke et al. 2017; Turaga et al. 2010) .
- Identifying predictive variables
MLはある現象に対して、どの変数が影響を与えているかを見出すことができる。leave-one-out strategy(各変数を抜いてみて、精度がどれくらい下がるかを観察する)やbest-first strategy(各変数一つだけにしてみて精度の変化をみる)といった方法があるが、いまだ研究中である「The development of feature selection algorithms that appropriately allow for nonlinear feature interactions is an active field in ML and statistics (Tang, Alelyani, and Liu 2014) .」
ただ、画像処理など変数そのものが生の入力データに明示されていないこともある。これはこのカテゴリ(2.)の対象外。
観察できる事象から神経の活動を推測すること(encoding)にもML（XGBoostやアンサンブル学習法、深層学習）が力を発揮している。MLを使うことで、シンプルなモデルが理解に役立っていたとしても、重要な特徴の見落としがわかることがある。
脳のある領域の活動からどんな情報が読み取れるかという逆問題(decoding)の解析にも役立つ。
もちろん、観察可能な変数からの疾患の推測にも使われている。モデルを仮定せず分析できるのも利点。
クラシフィケーション技術も、疾患の分析から脳のセグメンテーション等様々に使われている。
- Benchmarking simple models
MLはあるモデルの性能を計るのに用いることができる。
モデルの予測のぶれが、入力のノイズによるものなのか、それともモデル自体の不備によるものかの解析は難しいが、DNNを用いることで、モデルがどの程度の構造を説明できるかのUpper Boundを示すことができるらしい。これは、これまでのモデル同士を比較する手法からのパラダイムシフトである。 - Serving as a model for the brain
MLは脳モデルとして用いることができる。
似ているところは当たり前なのでスキップ。
- there are a few reasons to be skeptical (Lake et al. 2017 (Check!))
- neural networks require large amounts of data to train, while the brain can often learn from few examples (Carey and Bartlett 1978; F. Xu and Tenenbaum 2007) .
- artificial networks are plainly different in both architecture and response patterns from biological brains
- One focus on creating biologically plausible neural networks is on having spiking (binary), as opposed to continuous, units. Many recent research papers have begun to create spiking neural networks that successfully solve typical machine learning (ML) problems (Zenke and Ganguli 2017; Nicola and Clopath 2017; Mozafari et al. 2018)
- backpropagation, which is used for training neural nets, yet is not considered a biologically
plausible mechanism for credit assignment. One recent paper showed that random feedback weights still allows for successful learning (Lillicrap et al. 2016) , solving one of the implausible aspects of backpropagation. - Other work has presented networks based on the apical/basal dendrites to solve the problem of credit assignment (Guerguiev, Lillicrap, and Richards 2017; Körding and König 2001/Related (more original?) article) . Plus, there have been many other recent (Scellier and Bengio 2016; Bengio et al. 2015, 2017 ,check slides!) and historic (Hinton and McClelland 1988) works creating more plausible credit assignment mechanisms
- many biologically-motivated deep learning algorithms do not scale well to large datasets (Bartunov et al. 2018)
Interpretability / There has been much recent work to develop methods to better understand what is occurring within neural networks
- methods for visualizing what features and information are represented at different scales within convolutional neural networks (Olah et al. 2018, video)
- methods for understanding the dynamics within recurrent neural networks (Sussillo and Barak 2013)
- researchers are also developing new model architectures that are more easily interpretable (Foerster et al. 2017; Linderman et al. 2017)
- researchers recently tested whether the tuning of individual units in neural networks were important for classification generalization (Morcos et al. 2018)
- fully observable neural networks may serve as test-beds for new neural data analysis methods, which are greatly needed to increase understanding (Jonas and Kording 2017/video)
focusing on which cost functions the brain is optimizing rather than the final properties of the network. Similarly, it is important to focus on determining the learning rules implemented by the brain. We have discussed some of these matters recently in a separate review, particularly focusing on how neuroscience should learn from machine learning’s quest to understand neural networks (Marblestone, Wayne, and Kording 2016 / Summary by google) .
NN以外も研究対象である。
- Decision trees, for example, offer a compelling framework to model decision-making as a prunable tree of potential decisions (Huys et al. 2012) .
- Simple neural circuits have also been modeled with threshold decision trees (Uchizawa, Douglas, and Maass 2006)
- The cerebellum and similar structures (like the mushroom body in Drosophila ) can potentially be modeled as implementing a weighted nearest-neighbors calculation (Dasgupta, Stevens, and Navlakha 2017)
- information processing in the brain relates to the ML concept of random projections (Arriaga and Vempala 2006)
- comparisons of the brain to reinforcement learning (Wang et al. 2018; Gläscher et al. 2010; Glimcher 2011; Doya 2000) and unsupervised learning (Hinton and Sejnowski 1999; Doya 2000
Note:
As is the case with any modeling, we wish to remind readers that it is important to be cautious
when interpreting ML models. High predictability does not mean causality. This is especially true
because there are so many unobserved variables within neuroscience. For instance, the
apparent importance of a variable within a model may be inaccurate due to other unobserved
variables (Stevenson 2018) . Moreover, high model accuracy does not mean that it is actually a
causally correct model of what is occurring in the brain. High accuracy is a necessary, but not a
sufficient condition for model correctness. This is because there is an enormous space of
potential models that could explain the data well. This is a difficulty of modeling the brain and
identifying predictive variables with ML, but does not impact the use of ML for engineering, or
benchmark, applications.
