Opening the Black Box: Low-Dimensional Dynamics in High-Dimensional Recurrent Neural Networks

Recurrent neural networks (RNNs) are useful tools for learning nonlinear relationships between time-varying inputs and outputs with complex temporal dependencies. Recently developed algorithms have been successful at training RNNs to perform a wide variety of tasks, but the resulting networks have been treated as black boxes: their mechanism of operation remains unknown. Here we explore the hypothesis that fixed points, both stable and unstable, and the linearized dynamics around them, can reveal crucial aspects of how RNNs implement their computations. Further, we explore the utility of linearization in areas of phase space that are not true fixed points but merely points of very slow movement. We present a simple optimization technique that is applied to trained RNNs to find the fixed and slow points of their dynamics. Linearization around these slow regions can be used to explore, or reverse-engineer, the behavior of the RNN. We describe the technique, illustrate it using simple examples, and finally showcase it on three high-dimensional RNN examples: a 3-bit flip-flop device, an input-dependent sine wave generator, and a two-point moving average. In all cases, the mechanisms of trained networks could be inferred from the sets of fixed and slow points and the linearized dynamics around them.

[1]  Ramón Huerta,et al.  Reproducible sequence generation in random neural ensembles. , 2004, Physical review letters.

[2]  E. Ott Chaos in Dynamical Systems: Contents , 2002 .

[3]  Steven H. Strogatz,et al.  Nonlinear Dynamics and Chaos , 2024 .

[4]  Xiao-Jing Wang Decision Making in Recurrent Neuronal Circuits , 2008, Neuron.

[5]  Stephen P. Boyd,et al.  Convex Optimization , 2004, Algorithms and Theory of Computation Handbook.

[6]  Henry Markram,et al.  Real-Time Computing Without Stable States: A New Framework for Neural Computation Based on Perturbations , 2002, Neural Computation.

[7]  Mw Hirsch,et al.  Chaos In Dynamical Systems , 2016 .

[8]  M. Tsodyks,et al.  Working models of working memory , 2014, Current Opinion in Neurobiology.

[9]  J J Hopfield,et al.  Neural networks and physical systems with emergent collective computational abilities. , 1982, Proceedings of the National Academy of Sciences of the United States of America.

[10]  D. Lathrop Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering , 2015 .

[11]  Xue Z. Wang,et al.  (Q)SAR Modelling of Nanomaterial Toxicity - A Critical Review , 2015 .

[12]  Ranulfo Romo,et al.  Flexible Control of Mutual Inhibition: A Neural Model of Two-Interval Discrimination , 2005, Science.

[13]  Eduardo D. Sontag,et al.  Computational Aspects of Feedback in Neural Circuits , 2006, PLoS Comput. Biol..

[14]  Dean V Buonomano,et al.  Embedding Multiple Trajectories in Simulated Recurrent Neural Networks in a Self-Organizing Manner , 2009, The Journal of Neuroscience.

[15]  Ranulfo Romo,et al.  Neural dynamics of perceptual detection under temporal uncertainty , 2014, BMC Neuroscience.

[16]  James Martens,et al.  Deep learning via Hessian-free optimization , 2010, ICML.

[17]  Geoffrey E. Hinton,et al.  Learning internal representations by error propagation , 1986 .

[18]  Ilya Sutskever,et al.  Learning Recurrent Neural Networks with Hessian-Free Optimization , 2011, ICML.

[19]  Zachary F Mainen,et al.  Preparing and selecting actions with neural populations: toward cortical circuit mechanisms , 2015, Current Opinion in Neurobiology.

[20]  J. Hopfield,et al.  Computing with neural circuits: a model. , 1986, Science.

[21]  Matthew T. Kaufman,et al.  A neural network that finds a naturalistic solution for the production of muscle activity , 2015, Nature Neuroscience.

[22]  L. F. Abbott,et al.  Generating Coherent Patterns of Activity from Chaotic Neural Networks , 2009, Neuron.

[23]  Oliver Obst,et al.  Guided Self-Organization of Input-Driven Recurrent Neural Networks , 2013, ArXiv.

[24]  Alireza Goudarzi,et al.  DNA Reservoir Computing: A Novel Molecular Computing Approach , 2013, DNA.

[25]  Yoshua Bengio,et al.  Learning long-term dependencies with gradient descent is difficult , 1994, IEEE Trans. Neural Networks.

[26]  R. Romo,et al.  Neuronal correlates of parametric working memory in the prefrontal cortex , 1999, Nature.

[27]  Adam Ponzi,et al.  Sequentially Switching Cell Assemblies in Random Inhibitory Networks of Spiking Neurons in the Striatum , 2010, The Journal of Neuroscience.

[28]  Dursun Delen,et al.  Analyzing the Predictability of Exchange Traded Funds Characteristics in the Mutual Fund Market on the Flow of Shares Using a Data Mining Approach , 2014, 2014 47th Hawaii International Conference on System Sciences.

[29]  Ramón Huerta,et al.  Transient Cognitive Dynamics, Metastability, and Decision Making , 2008, PLoS Comput. Biol..

[30]  Eli Shlizerman,et al.  Symmetries constrain the transition to heterogeneous chaos in balanced networks , 2015, BMC Neuroscience.

[31]  Harald Haas,et al.  Harnessing Nonlinearity: Predicting Chaotic Systems and Saving Energy in Wireless Communication , 2004, Science.

[32]  David Sussillo,et al.  Neural circuits as computational dynamical systems , 2014, Current Opinion in Neurobiology.

[33]  V. Jayaraman,et al.  Intensity versus Identity Coding in an Olfactory System , 2003, Neuron.

[34]  Drew N. Robson,et al.  Brain-wide neuronal dynamics during motor adaptation in zebrafish , 2012, Nature.

[35]  Srdjan Ostojic,et al.  Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons , 2014, Nature Neuroscience.

[36]  Michael J. Black,et al.  A freely-moving monkey treadmill model , 2014, Journal of neural engineering.

[37]  Paul Miller,et al.  Natural stimuli evoke dynamic sequences of states in sensory cortical ensembles , 2007, Proceedings of the National Academy of Sciences.

[38]  W. Newsome,et al.  Context-dependent computation by recurrent dynamics in prefrontal cortex , 2013, Nature.

[39]  N. Parga,et al.  Dynamic Control of Response Criterion in Premotor Cortex during Perceptual Detection under Temporal Uncertainty , 2015, Neuron.

[40]  Hideo Mabuchi,et al.  A coherent perceptron for all-optical learning , 2015, 1501.01608.